Butterfly Research
How much can we learnabout your business?
Type your web address. Butterfly reads what your company makes public, breaks it down in front of you and shows the work AI could take on first.
- Your website, waiting
- Your pictures, waiting
- The public register, waiting
- What you do, waiting
- Who you are 0
- What you say 0
- What you show 0
- Where you are 0
A live read takes a few seconds. You watch it find each part.
Your posts appear here
Company-level, public sources: your website, the posts it links to and the UK company register. Nothing about your people: no names, no reviews.
02 · The breakdown
Your words,broken down.
Every line we found, read again here in your browser and sorted into four kinds.
Read a company above and its breakdown appears here.
Read my company03 · Likely tasks
Aviva’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
Example10 offerings → 7 kinds of work in insurance
A first guessStarted by customers on aviva.co.uk, and the work behind each
On your site
Kinds of work, by AI potential
From your site: Car Insurance, Home Insurance, Electric car insurance, Van insurance, Travel Insurance, Health insurance
Test first · 01Claims handling
In your filmSettling a home insurance claim after the excess and limits
As our research words it Work out the settlement after excess, limits and depreciation.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £27.33 each timeIllustrative
- Doing it 33 minutes of a claims handler, at £35 an hour with on‑costs£19.25
- Checking it 7 minutes of checks and corrections, at the same rate£4.08
- Software Claims system, Policy admin, Payments£4
- Checked today by
- Claims handlers
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Car Insurance, Home Insurance, Electric car insurance, Van insurance, Travel Insurance, Health insurance
Test first · 01Claims handling
In your filmSettling a home insurance claim after the excess and limits
Work out the settlement after excess, limits and depreciation.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £27.33 each timeIllustrative
- Doing it 33 minutes of a claims handler, at £35 an hour with on‑costs£19.25
- Checking it 7 minutes of checks and corrections, at the same rate£4.08
- Software Claims system, Policy admin, Payments£4
- Checked today by
- Claims handlers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
The film: one task, its cost and who could do it, narrated.
Make it real
Your proposal: the Diagnostic, priced.
03 · Likely tasks
Allianz’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
7 offerings → 5 kinds of work in insurance
A first guessOn your site
Kinds of work, by AI potential
From your site: Kfz-Versicherung, E-Auto-Versicherung, Motorrad-Versicherung, Versicherung
Test first · 01Claims handling
In your filmDeciding a motor claim against the policy wording
As our research words it Decide a claim against the policy wording: pay, part-pay or decline, with reasons.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £30.25 each timeIllustrative
- Doing it 36 minutes of a claims handler, at £35 an hour with on‑costs£21
- Checking it 9 minutes of checks and corrections, at the same rate£5.25
- Software Claims system, Policy admin, Document store£4
- Checked today by
- Claims handlers
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Kfz-Versicherung, E-Auto-Versicherung, Motorrad-Versicherung, Versicherung
Test first · 01Claims handling
In your filmDeciding a motor claim against the policy wording
Decide a claim against the policy wording: pay, part-pay or decline, with reasons.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £30.25 each timeIllustrative
- Doing it 36 minutes of a claims handler, at £35 an hour with on‑costs£21
- Checking it 9 minutes of checks and corrections, at the same rate£5.25
- Software Claims system, Policy admin, Document store£4
- Checked today by
- Claims handlers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Allianz’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Allianz: the Diagnostic, priced.
03 · Likely tasks
AstraZeneca’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
10 offerings → 4 kinds of work in pharma and life sciences
A first guessOn your site
- Cell therapies
- Oligonucleotide therapeutics
- AI and data
- Genomics
- New Approach Methodologies
- Oncology
- BioPharmaceuticals
- Cardiovascular, Renal and Metabolism
- Respiratory & Immunology
- Infectious Disease
Not tied to one kind of work yet: your Diagnostic maps them.
Kinds of work, by AI potential
Every company in pharma and life sciences does this
Test first · 01Regulatory submissions
In your filmChecking a regulatory document for claims the data don’t support
As our research words it Check a submission document for claims the data do not support and for conflicts with other documents.
- How often
- Dozens a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £480 each timeIllustrative
- Doing it 5 hours of a regulatory reviewer, at £75 an hour with on‑costs£375
- Checking it an hour of checks and corrections, at the same rate£75
- Software Document system, Submission tool, Reference library£30
- Checked today by
- Regulatory affairs staff
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
Every company in pharma and life sciences does this
Test first · 01Regulatory submissions
In your filmChecking a regulatory document for claims the data don’t support
Check a submission document for claims the data do not support and for conflicts with other documents.
- How often
- Dozens a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £480 each timeIllustrative
- Doing it 5 hours of a regulatory reviewer, at £75 an hour with on‑costs£375
- Checking it an hour of checks and corrections, at the same rate£75
- Software Document system, Submission tool, Reference library£30
- Checked today by
- Regulatory affairs staff
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
AstraZeneca’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for AstraZeneca: the Diagnostic, priced.
03 · Likely tasks
Aviva’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
10 offerings → 7 kinds of work in insurance
A first guessStarted by customers on aviva.co.uk, and the work behind each
On your site
Kinds of work, by AI potential
From your site: Car Insurance, Home Insurance, Electric car insurance, Van insurance, Travel Insurance, Health insurance
Test first · 01Claims handling
In your filmSettling a home insurance claim after the excess and limits
As our research words it Work out the settlement after excess, limits and depreciation.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £27.33 each timeIllustrative
- Doing it 33 minutes of a claims handler, at £35 an hour with on‑costs£19.25
- Checking it 7 minutes of checks and corrections, at the same rate£4.08
- Software Claims system, Policy admin, Payments£4
- Checked today by
- Claims handlers
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Car Insurance, Home Insurance, Electric car insurance, Van insurance, Travel Insurance, Health insurance
Test first · 01Claims handling
In your filmSettling a home insurance claim after the excess and limits
Work out the settlement after excess, limits and depreciation.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £27.33 each timeIllustrative
- Doing it 33 minutes of a claims handler, at £35 an hour with on‑costs£19.25
- Checking it 7 minutes of checks and corrections, at the same rate£4.08
- Software Claims system, Policy admin, Payments£4
- Checked today by
- Claims handlers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Aviva’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Aviva: the Diagnostic, priced.
03 · Likely tasks
Barclays’ likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
10 offerings → 5 kinds of work in banking
A first guessStarted by customers on barclays.co.uk, and the work behind each
On your site
Kinds of work, by AI potential
From your site: Credit cards
Test first · 01Complaints and card disputes
In your filmDeciding a card dispute under the card scheme’s rules
As our research words it Decide a card dispute under the card scheme’s rules: refund, re-present or reject.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £16.33 each timeIllustrative
- Doing it 21 minutes of a disputes analyst, at £32 an hour with on‑costs£11.20
- Checking it 4 minutes of checks and corrections, at the same rate£2.13
- Software Disputes system, Card scheme portal, Core banking£3
- Checked today by
- Operations staff
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Credit cards
Test first · 01Complaints and card disputes
In your filmDeciding a card dispute under the card scheme’s rules
Decide a card dispute under the card scheme’s rules: refund, re-present or reject.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £16.33 each timeIllustrative
- Doing it 21 minutes of a disputes analyst, at £32 an hour with on‑costs£11.20
- Checking it 4 minutes of checks and corrections, at the same rate£2.13
- Software Disputes system, Card scheme portal, Core banking£3
- Checked today by
- Operations staff
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Barclays’ film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Barclays: the Diagnostic, priced.
03 · Likely tasks
Birmingham City Council’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
7 offerings → 4 kinds of work in public sector and government
A first guessOn your site
Kinds of work, by AI potential
From your site: Roads, travel and parking
Test first · 01Permits and licensing
In your filmChecking a planning application for anything missing
As our research words it Check a planning or licence application for completeness and list what is missing.
- How often
- Hundreds a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £36.33 each timeIllustrative
- Doing it 41 minutes of a planning officer, at £40 an hour with on‑costs£27.33
- Checking it 9 minutes of checks and corrections, at the same rate£6
- Software Planning system, Case system, Document store£3
- Checked today by
- Planning or licensing officers
AI potential · goodLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Roads, travel and parking
Test first · 01Permits and licensing
In your filmChecking a planning application for anything missing
Check a planning or licence application for completeness and list what is missing.
- How often
- Hundreds a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £36.33 each timeIllustrative
- Doing it 41 minutes of a planning officer, at £40 an hour with on‑costs£27.33
- Checking it 9 minutes of checks and corrections, at the same rate£6
- Software Planning system, Case system, Document store£3
- Checked today by
- Planning or licensing officers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Birmingham City Council’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Birmingham City Council: the Diagnostic, priced.
03 · Likely tasks
BT’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
9 offerings → 5 kinds of work in telecoms
A first guessStarted by customers on bt.com, and the work behind each
On your site
Kinds of work, by AI potential
The work your film follows
Test first · 01Field service
In your filmDeciding whether a broadband fault needs an engineer’s visit
As our research words it Decide from the line test results whether a fault needs a home visit or a network fix.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £7 each timeIllustrative
- Doing it 10 minutes of a fault-desk adviser, at £30 an hour with on‑costs£5
- Checking it 2 minutes of checks and corrections, at the same rate£1
- Software Line-test tools, CRM, Field scheduling£1
- Checked today by
- Field supervisors
AI potential · goodLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
The work your film follows
Test first · 01Field service
In your filmDeciding whether a broadband fault needs an engineer’s visit
Decide from the line test results whether a fault needs a home visit or a network fix.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £7 each timeIllustrative
- Doing it 10 minutes of a fault-desk adviser, at £30 an hour with on‑costs£5
- Checking it 2 minutes of checks and corrections, at the same rate£1
- Software Line-test tools, CRM, Field scheduling£1
- Checked today by
- Field supervisors
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
BT’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for BT: the Diagnostic, priced.
03 · Likely tasks
DHL’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
6 offerings → 5 kinds of work in logistics, freight and shipping
A first guessOn your site
- Global Logistics
- International Shipping
Not tied to one kind of work yet: your Diagnostic maps them.
Kinds of work, by AI potential
Every company in logistics, freight and shipping does this
Test first · 01Warehouse operations
In your filmTracing a stock discrepancy across receipts, picks and counts
As our research words it Investigate a stock discrepancy across receipts, put-aways, picks and counts, and post the right adjustment.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £25.33 each timeIllustrative
- Doing it 41 minutes of a stock controller, at £28 an hour with on‑costs£19.13
- Checking it 9 minutes of checks and corrections, at the same rate£4.20
- Software Warehouse system, Scanners, Inventory reports£2
- Checked today by
- Warehouse supervisors
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
Every company in logistics, freight and shipping does this
Test first · 01Warehouse operations
In your filmTracing a stock discrepancy across receipts, picks and counts
Investigate a stock discrepancy across receipts, put-aways, picks and counts, and post the right adjustment.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £25.33 each timeIllustrative
- Doing it 41 minutes of a stock controller, at £28 an hour with on‑costs£19.13
- Checking it 9 minutes of checks and corrections, at the same rate£4.20
- Software Warehouse system, Scanners, Inventory reports£2
- Checked today by
- Warehouse supervisors
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
DHL’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for DHL: the Diagnostic, priced.
03 · Likely tasks
GSK’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
5 offerings → 6 kinds of work in pharma and life sciences
A first guessOn your site
Kinds of work, by AI potential
The work your film follows
Test first · 01Manufacturing quality
In your filmReviewing a batch record before the batch goes to release
As our research words it Review a batch record for errors and decide whether the batch can go forward to release.
- How often
- Hundreds a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £200 each timeIllustrative
- Doing it 2½ hours of a quality specialist, at £60 an hour with on‑costs£150
- Checking it 30 minutes of checks and corrections, at the same rate£30
- Software Quality system, Batch records, Lab results£20
- Checked today by
- Quality assurance staff
AI potential · goodLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
The work your film follows
Test first · 01Manufacturing quality
In your filmReviewing a batch record before the batch goes to release
Review a batch record for errors and decide whether the batch can go forward to release.
- How often
- Hundreds a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £200 each timeIllustrative
- Doing it 2½ hours of a quality specialist, at £60 an hour with on‑costs£150
- Checking it 30 minutes of checks and corrections, at the same rate£30
- Software Quality system, Batch records, Lab results£20
- Checked today by
- Quality assurance staff
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
GSK’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for GSK: the Diagnostic, priced.
03 · Likely tasks
Guy’s and St Thomas’ likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
5 offerings → 4 kinds of work in healthcare providers
A first guessOn your site
Kinds of work, by AI potential
Every company in healthcare providers does this
Test first · 01Clinical coding
In your filmCoding a hospital stay from its clinical notes
As our research words it Assign the diagnosis and procedure codes for a hospital stay from its notes, and catch the errors before the record is submitted.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £8.50 each timeIllustrative
- Doing it 12 minutes of a clinical coder, at £30 an hour with on‑costs£6
- Checking it 3 minutes of checks and corrections, at the same rate£1.50
- Software Coding software, Patient record, Reporting£1
- Checked today by
- Clinical coders
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
Every company in healthcare providers does this
Test first · 01Clinical coding
In your filmCoding a hospital stay from its clinical notes
Assign the diagnosis and procedure codes for a hospital stay from its notes, and catch the errors before the record is submitted.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £8.50 each timeIllustrative
- Doing it 12 minutes of a clinical coder, at £30 an hour with on‑costs£6
- Checking it 3 minutes of checks and corrections, at the same rate£1.50
- Software Coding software, Patient record, Reporting£1
- Checked today by
- Clinical coders
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Guy’s and St Thomas’ film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Guy’s and St Thomas’: the Diagnostic, priced.
03 · Likely tasks
Linklaters’ likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
10 offerings → 5 kinds of work in legal services
A first guessOn your site
- Aerospace, Defence and Security
- Automotive
- Banks
- Energy and Utilities
- Healthcare and Life Sciences
- Infrastructure
Not tied to one kind of work yet: your Diagnostic maps them.
Kinds of work, by AI potential
Every company in legal services does this
Test first · 01NDA and contract triage
In your filmTriaging an incoming NDA: sign it, amend it, or escalate it
As our research words it Triage an incoming NDA: sign as is, amend or escalate, with reasons.
- How often
- Hundreds a weekA first guess for a company of 2,000 to 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £110 each timeIllustrative
- Doing it 32 minutes of an associate solicitor, at £150 an hour with on‑costs£80
- Checking it 8 minutes of checks and corrections, at the same rate£20
- Software Contract tools, Document management, Matter system£10
- Checked today by
- In-house lawyers
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
Every company in legal services does this
Test first · 01NDA and contract triage
In your filmTriaging an incoming NDA: sign it, amend it, or escalate it
Triage an incoming NDA: sign as is, amend or escalate, with reasons.
- How often
- Hundreds a weekA first guess for a company of 2,000 to 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £110 each timeIllustrative
- Doing it 32 minutes of an associate solicitor, at £150 an hour with on‑costs£80
- Checking it 8 minutes of checks and corrections, at the same rate£20
- Software Contract tools, Document management, Matter system£10
- Checked today by
- In-house lawyers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Linklaters’ film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Linklaters: the Diagnostic, priced.
03 · Likely tasks
The University of Manchester’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
8 offerings → 4 kinds of work in education and training
A first guessStarted by customers on manchester.ac.uk, and the work behind each
On your site
- Accommodation
- Study abroad and exchange
Not tied to one kind of work yet: your Diagnostic maps them.
Kinds of work, by AI potential
From your site: Online and blended learning
Test first · 01Marking and feedback
In your filmMarking an essay against its rubric, as a second marker would
As our research words it Mark a script against the rubric and give feedback a second marker would agree with.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £23.50 each timeIllustrative
- Doing it 24 minutes of a second marker, at £45 an hour with on‑costs£18
- Checking it 6 minutes of checks and corrections, at the same rate£4.50
- Software Marking platform, Rubrics, Marks records£1
- Checked today by
- Academic markers
AI potential · goodLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Online and blended learning
Test first · 01Marking and feedback
In your filmMarking an essay against its rubric, as a second marker would
Mark a script against the rubric and give feedback a second marker would agree with.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £23.50 each timeIllustrative
- Doing it 24 minutes of a second marker, at £45 an hour with on‑costs£18
- Checking it 6 minutes of checks and corrections, at the same rate£4.50
- Software Marking platform, Rubrics, Marks records£1
- Checked today by
- Academic markers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
The University of Manchester’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for The University of Manchester: the Diagnostic, priced.
03 · Likely tasks
Octopus Energy’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
10 offerings → 4 kinds of work in energy and utilities
A first guessStarted by customers on octopus.energy, and the work behind each
On your site
- Electric vehicles
- Electric vehicles & chargers
- Intelligent Octopus Go
- Octopus Tracker
- Agile Octopus
- Octopus Fan Club
- Intelligent Octopus Flux
Not tied to one kind of work yet: your Diagnostic maps them.
Kinds of work, by AI potential
From your site: Energy
Test first · 01Metering and billing
In your filmChecking a suspect meter reading before the bill goes out
As our research words it Validate a suspect meter reading and decide whether to estimate, re-read or bill it.
- How often
- Tens of thousands a weekA first guess for a company of 2,000 to 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £3.80 each timeIllustrative
- Doing it 5 minutes of a billing analyst, at £28 an hour with on‑costs£2.33
- Checking it 1 minute of checks and corrections, at the same rate£0.47
- Software Billing system, Meter data, CRM£1
- Checked today by
- Billing specialists
AI potential · goodLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Energy
Test first · 01Metering and billing
In your filmChecking a suspect meter reading before the bill goes out
Validate a suspect meter reading and decide whether to estimate, re-read or bill it.
- How often
- Tens of thousands a weekA first guess for a company of 2,000 to 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £3.80 each timeIllustrative
- Doing it 5 minutes of a billing analyst, at £28 an hour with on‑costs£2.33
- Checking it 1 minute of checks and corrections, at the same rate£0.47
- Software Billing system, Meter data, CRM£1
- Checked today by
- Billing specialists
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Octopus Energy’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Octopus Energy: the Diagnostic, priced.
03 · Likely tasks
PwC’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
5 offerings → 4 kinds of work in accounting, audit and tax firms
A first guessOn your site
- Consulting
- Economics
Not tied to one kind of work yet: your Diagnostic maps them.
Kinds of work, by AI potential
Every company in accounting, audit and tax firms does this
Test first · 01Bookkeeping and outsourced accounting
In your filmReconciling a client’s bank and ledger accounts
As our research words it Reconcile a client’s bank and ledger accounts and post the adjusting entries.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £44.25 each timeIllustrative
- Doing it 36 minutes of an accountant, at £55 an hour with on‑costs£33
- Checking it 9 minutes of checks and corrections, at the same rate£8.25
- Software Accounting software, Bank feeds, Workpapers£3
- Checked today by
- Accountants
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
Every company in accounting, audit and tax firms does this
Test first · 01Bookkeeping and outsourced accounting
In your filmReconciling a client’s bank and ledger accounts
Reconcile a client’s bank and ledger accounts and post the adjusting entries.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £44.25 each timeIllustrative
- Doing it 36 minutes of an accountant, at £55 an hour with on‑costs£33
- Checking it 9 minutes of checks and corrections, at the same rate£8.25
- Software Accounting software, Bank feeds, Workpapers£3
- Checked today by
- Accountants
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
PwC’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for PwC: the Diagnostic, priced.
03 · Likely tasks
Rolls-Royce’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
4 offerings → 4 kinds of work in manufacturing
A first guessOn your site
Kinds of work, by AI potential
From your site: Testing capabilities
Test first · 01Quality and non-conformance
In your filmDeciding what happens to a part outside its drawing’s limits
As our research words it Decide the disposition of a non-conforming batch (use as is, rework, scrap or return to supplier) against the drawing and specification.
- How often
- Hundreds a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £140 each timeIllustrative
- Doing it 1½ hours of a quality engineer, at £65 an hour with on‑costs£97.50
- Checking it 30 minutes of checks and corrections, at the same rate£32.50
- Software Quality system, Drawings, Inspection data£10
- Checked today by
- Quality engineers
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Testing capabilities
Test first · 01Quality and non-conformance
In your filmDeciding what happens to a part outside its drawing’s limits
Decide the disposition of a non-conforming batch (use as is, rework, scrap or return to supplier) against the drawing and specification.
- How often
- Hundreds a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £140 each timeIllustrative
- Doing it 1½ hours of a quality engineer, at £65 an hour with on‑costs£97.50
- Checking it 30 minutes of checks and corrections, at the same rate£32.50
- Software Quality system, Drawings, Inspection data£10
- Checked today by
- Quality engineers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Rolls-Royce’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Rolls-Royce: the Diagnostic, priced.
03 · Likely tasks
Sage’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
6 offerings → 4 kinds of work in software and SaaS
A first guessOn your site
Kinds of work, by AI potential
From your site: Sage Intacct, Sage X3, Brightpearl by Sage, Sage 100, Sage 300
Test first · 01Technical support
In your filmDiagnosing a customer’s problem from the ticket and the logs
As our research words it Diagnose a customer’s technical problem from the ticket, logs and documentation, and give the fix.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £22.42 each timeIllustrative
- Doing it 29 minutes of a support engineer, at £35 an hour with on‑costs£16.92
- Checking it 6 minutes of checks and corrections, at the same rate£3.50
- Software Support desk, Logs, Knowledge base£2
- Checked today by
- Senior support engineers
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Sage Intacct, Sage X3, Brightpearl by Sage, Sage 100, Sage 300
Test first · 01Technical support
In your filmDiagnosing a customer’s problem from the ticket and the logs
Diagnose a customer’s technical problem from the ticket, logs and documentation, and give the fix.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £22.42 each timeIllustrative
- Doing it 29 minutes of a support engineer, at £35 an hour with on‑costs£16.92
- Checking it 6 minutes of checks and corrections, at the same rate£3.50
- Software Support desk, Logs, Knowledge base£2
- Checked today by
- Senior support engineers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Sage’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Sage: the Diagnostic, priced.
03 · Likely tasks
Severn Trent’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
4 offerings → 4 kinds of work in energy and utilities
A first guessOn your site
- Bioresources
- Hafren Dyfrdwy
Not tied to one kind of work yet: your Diagnostic maps them.
Kinds of work, by AI potential
From your site: Severn Trent Green Power, Severn Trent Water
Test first · 01Asset inspection and maintenance
In your filmGrading a pump’s condition from its inspection report and photos
As our research words it Grade an asset’s condition from an inspection report and photos, and decide whether to repair, monitor or replace it.
- How often
- Hundreds a weekA first guess for a company of 2,000 to 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £60 each timeIllustrative
- Doing it 48 minutes of an asset engineer, at £55 an hour with on‑costs£44
- Checking it 12 minutes of checks and corrections, at the same rate£11
- Software Asset system, Inspection app, Work orders£5
- Checked today by
- Asset engineers
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Severn Trent Green Power, Severn Trent Water
Test first · 01Asset inspection and maintenance
In your filmGrading a pump’s condition from its inspection report and photos
Grade an asset’s condition from an inspection report and photos, and decide whether to repair, monitor or replace it.
- How often
- Hundreds a weekA first guess for a company of 2,000 to 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £60 each timeIllustrative
- Doing it 48 minutes of an asset engineer, at £55 an hour with on‑costs£44
- Checking it 12 minutes of checks and corrections, at the same rate£11
- Software Asset system, Inspection app, Work orders£5
- Checked today by
- Asset engineers
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Severn Trent’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Severn Trent: the Diagnostic, priced.
03 · Likely tasks
Tesco’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
7 offerings → 5 kinds of work in e-commerce and retail
A first guessOn your site
Kinds of work, by AI potential
From your site: Online Groceries
Test first · 01Stock and replenishment
In your filmSetting a product’s reorder quantity from sales, stock and lead times
As our research words it Decide the reorder quantity for a product from sales, stock and supplier lead times.
- How often
- Tens of thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £2.75 each timeIllustrative
- Doing it 2 minutes of a replenishment analyst, at £35 an hour with on‑costs£1.17
- Checking it 1 minute of checks and corrections, at the same rate£0.58
- Software Replenishment system, Stock records, Supplier portal£1
- Checked today by
- Replenishment planners
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Online Groceries
Test first · 01Stock and replenishment
In your filmSetting a product’s reorder quantity from sales, stock and lead times
Decide the reorder quantity for a product from sales, stock and supplier lead times.
- How often
- Tens of thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £2.75 each timeIllustrative
- Doing it 2 minutes of a replenishment analyst, at £35 an hour with on‑costs£1.17
- Checking it 1 minute of checks and corrections, at the same rate£0.58
- Software Replenishment system, Stock records, Supplier portal£1
- Checked today by
- Replenishment planners
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Tesco’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Tesco: the Diagnostic, priced.
03 · Likely tasks
Whitbread’s likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks follow the moment the read finishes.
5 offerings → 4 kinds of work in travel, hospitality and airlines
A first guessOn your site
Kinds of work, by AI potential
From your site: Book a room
Test first · 01Reservations and changes
In your filmWorking out the refund when a guest cancels under the rate’s rules
As our research words it Cancel a booking and work out the refund or credit under the fare and policy rules.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £4.47 each timeIllustrative
- Doing it 6 minutes of a reservations adviser, at £26 an hour with on‑costs£2.60
- Checking it 2 minutes of checks and corrections, at the same rate£0.87
- Software Booking system, Rate rules, Payments£1
- Checked today by
- Servicing agents
AI potential · highLab demand, value per task and recorded outcomes, less the effort to test it and the regulation around it.- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
From your site: Book a room
Test first · 01Reservations and changes
In your filmWorking out the refund when a guest cancels under the rate’s rules
Cancel a booking and work out the refund or credit under the fare and policy rules.
- How often
- Thousands a weekA first guess for a company of more than 10,000 people. Your Diagnostic counts it from your logs.
- What it costs to run
- £4.47 each timeIllustrative
- Doing it 6 minutes of a reservations adviser, at £26 an hour with on‑costs£2.60
- Checking it 2 minutes of checks and corrections, at the same rate£0.87
- Software Booking system, Rate rules, Payments£1
- Checked today by
- Servicing agents
- Lab demand
- Value per task
- Recorded outcomes
- Effort to test
- Regulation
Clears the bar or not: every agent that could take it on would be tested on about a hundred held-back cases in your cloud before anyone relies on one.
How often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
Whitbread’s film: one task, its cost and who could do it, narrated.
Make it real
A proposal for Whitbread: the Diagnostic, priced.
03 · Likely tasks
likely tasks,from its own words.
Your products and services point to work your teams do every day. Butterfly’s research scores 220 kinds of work across 35 sectors for AI potential. Your Diagnostic then finds the thousands of tasks inside them.
The likely tasks appear after the read.
Start the readHow often it happens×What it costs to run: hours × salary + software=What it is worth to you, in £
Of the thousands of tasks in your business, the first 25 are the ones worth more to you than the rest, ranked by what each costs a year and how often it happens, so the return you can track far outweighs the cost.
See it as a story
The film: one task, its cost and who could do it, narrated.
Make it real
Your proposal: the Diagnostic, priced.
04 · Your industry
A research enginefor your industry.
Every week the module reads your market, company by company, and turns what changes into tasks to test, a quarterly report and predictions we keep score on.
Ten industries, one engine each. Add your company and it opens on yours.
What it watches, every week
- Rivals’ pagesRate, fee and product wording at the banks you name, week by week
- Job postsRoles that name the work: KYC review, credit analysis, payment operations
- Your own wordingWhat your pages promise about decisions and turnaround
- Customer feedbackPublic complaint themes, product by product
One signal, followed to a test
- 01 · The signalA rival starts promising same-day decisions on business lending
- 02 · The task it points toLoan spreading and covenant extraction
- 03 · The testCan an agent spread a statement to your standard, run after run, with a person signing the credit memo?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
Nearly half of UK financial firms report a partial understanding of the AI they use, and a third of their use cases are third-party.
Source: Bank of England / FCA, 21 Nov 2024 7
EU AI Act: credit scoring is high-risk, with duties from 2 Dec 2027. PRA SS1/23 asks banks for independent model validation.51
What it watches, every week
- Rivals’ pagesCover wording, excesses and claims promises at the insurers you name
- Job postsClaims handlers, underwriting assistants, bordereaux analysts
- Your own wordingWhat your pages promise about claims: speed, channels, cover
- Customer feedbackPublic complaint themes about claims and renewals
One signal, followed to a test
- 01 · The signalYour homepage promises help with a claim at any hour
- 02 · The task it points toFirst-notice-of-loss triage and coverage decisions
- 03 · The testDoes an agent route every claim the way your handlers would, and hand over when it should?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
AIG says underwriting reviews ran faster and data accuracy rose in early rollouts.
EU AI Act: life and health insurance pricing is high-risk, with duties from 2 Dec 2027.51
What it watches, every week
- Other providers’ pagesHow other providers describe referrals, waiting times and appointments
- Job postsRoles that name the work: clinical coders, booking teams, appeals staff
- Your own wordingWhat your pages promise patients about appointments, referrals and letters
- Patient feedbackPublic complaint themes about appointments, letters and bills
One signal, followed to a test
- 01 · The signalA run of clinical coder vacancies in your job posts
- 02 · The task it points toClinical coding
- 03 · The testDoes an agent code a hospital stay the way your coders would, every time, and query the clinician when the notes fall short?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
US physicians handle about 40 prior authorisations a week, which take them and their staff 13 hours.
Source: American Medical Association, 2026 28
EU AI Act: AI that decides eligibility for public benefits and services is high-risk, with duties from 2 Dec 2027.54
What it watches, every week
- Peers’ pagesHow peers describe their billing, payment terms and refunds
- Job postsRoles that name the work: reconciliations, accounts payable, month-end close
- Your own wordingPayment terms and invoicing promises on your own pages
- Customer feedbackPublic complaints about billing and refunds
One signal, followed to a test
- 01 · The signalA run of new month-end close roles in your job posts
- 02 · The task it points toBank and intercompany reconciliations
- 03 · The testTested on real cases held back from the start, at least 20 a task and typically about a hundred, against the bar the task owner sets.
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
On APEX-Accounting, models score well on average, yet almost none pass a task on all 8 tries.
Source: APEX-Accounting, arXiv, 29 Jul 2026 11
Every company closes its books, so this is the opening batch for every client, whatever the industry.
What it watches, every week
- Other firms’ pagesHow other firms describe their practice areas, fixed fees and turnaround
- Job postsRoles that name the work: contract managers, paralegals, knowledge lawyers
- Your own wordingWhat your pages promise clients about speed, fees and advice
- Client feedbackPublic complaint themes about service and cost
One signal, followed to a test
- 01 · The signalA rival firm starts promising fixed-fee NDA review within a day
- 02 · The task it points toPlaybook contract review
- 03 · The testDoes an agent mark a contract up to your playbook, clause by clause, run after run, and escalate what falls outside it?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
Well-known legal research AI tools hallucinated between a sixth and a third of the time.
Source: Stanford RegLab / HAI, arXiv, 30 May 2024 21
Court cases involving AI-hallucinated content keep mounting: 2,217 counted so far, 76 of them in the UK.22
What it watches, every week
- Peers’ pagesHow peers describe their trials, pipelines and medical information
- Job postsRoles that name the work: safety case processors, regulatory writers, medical information
- Your own wordingWhat your pages say about trials, safety reporting and enquiries
- Public enquiriesThemes in public enquiries, product by product
One signal, followed to a test
- 01 · The signalA new indication approved for one of your products
- 02 · The task it points toAdverse-event case processing
- 03 · The testDoes an agent decide whether a case is serious, expected and reportable, and by when, the way your safety physicians would?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
What it watches, every week
- Rivals’ pagesHow other managers describe fees, performance and reporting
- Job postsRoles that name the work: fund accountants, client reporting, operations analysts
- Your own wordingWhat your pages promise about reporting, suitability and service
- Client feedbackPublic complaint themes about advice and service
One signal, followed to a test
- 01 · The signalA rival starts promising daily reporting to every client
- 02 · The task it points toClient-report commentary
- 03 · The testDoes an agent write the quarter’s commentary from the numbers, accurately, every time, with a person signing it off?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
Three in four enterprise AI use cases are now purchased from vendors rather than made in-house, and spending tripled.
Source: Menlo Ventures, 9 Dec 2025 4
Firms outside its scope use PRA SS1/23 as the benchmark for models made in-house or by vendors, AI included.56
What it watches, every week
- Rivals’ pagesTariffs, bundles and service promises at the providers you name
- Job postsRoles that name the work: complaints handlers, billing analysts, fault-desk advisers
- Your own wordingWhat your pages promise about faults, bills and switching
- Customer feedbackPublic complaint themes: billing, faults and cancellations
One signal, followed to a test
- 01 · The signalA rival starts promising same-day fault fixes
- 02 · The task it points toGuided fault diagnosis
- 03 · The testDoes an agent find the fault and choose the fix your engineers would, and hand over when a customer needs a person?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
Klarna’s AI handled two-thirds of its customer chats in its first month; a year later Klarna rehired people because quality suffered.
EU AI Act: from 2 Aug 2026, an AI voice must tell the caller it is one.51
What it watches, every week
- Other bodies’ pagesHow other councils and agencies describe services, targets and waiting times
- Job postsRoles that name the work: planning officers, benefits assessors, licensing teams
- Your own wordingWhat your pages promise residents about applications and decisions
- Residents’ feedbackPublic complaint themes about applications and response times
One signal, followed to a test
- 01 · The signalNew guidance changes what a planning application must include
- 02 · The task it points toPlanning-application validation
- 03 · The testDoes an agent spot every missing document in an application, run after run, and say exactly what is missing?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
Documented AI incidents hit a record, and agents still fail about one in three structured benchmark attempts.
Source: Stanford AI Index 2026, 2026 23
EU AI Act: AI that decides eligibility for public benefits is high-risk, with duties from 2 Dec 2027.54
What it watches, every week
- Other firms’ pagesHow other firms describe their services, sectors and case studies
- Job postsRoles that name the work: analysts, bid writers, research associates
- Your own wordingWhat your pages promise clients about speed and insight
- Client feedbackPublic complaint themes about delivery and cost
One signal, followed to a test
- 01 · The signalA rival starts promising a market sizing within two days
- 02 · The task it points toMarket sizing
- 03 · The testDoes an agent size a market from the sources your analysts use, every assumption shown, the same way run after run?
- Your market what changed, rival by rival
- Frontier changes new models, from the Agent Reliability Index
- Predictions tracked, with their hit rate
The same tests carry into diligence: the Agent Target Benchmark checks a target’s AI before a deal closes.
Your industry outside these ten? We set its engine up with you, case by case.
What the module costs05 · Rules and dates
The rules that ask for proof,and when they apply.
Regulators already expect evidence that the AI a company uses works as it should. These are the dates that bear on testing it, and what Butterfly’s runs give you for each.
- 17 May 2024Today
PRA SS1/23, model risk management
- Who
- UK banks, building societies and PRA-designated investment firms with internal-model approval; others use it as a benchmark
- Butterfly gives you
- Independent test evidence for each model and agent, run after run, ready for your validation file.
What it asks, and the source
Validate the models you use, made in-house or by a vendor, AI included. The current version took effect on 23 April 2026.
Source: Bank of England, PRA SS1/23, 23 Apr 2026 56
- 17 Jan 2025Today
DORA, digital operational resilience
- Who
- Banks, insurers, investment firms and their ICT providers in the EU
- Butterfly gives you
- Test runs of each AI service a provider supplies, on your own work, kept with your records on that provider.
What it asks, and the source
EU rules on digital resilience for 20 types of financial firm and the ICT providers they rely on, with EU oversight of critical third-party providers. Managing the risk of those providers is one of its core areas.
Source: EIOPA, Digital Operational Resilience Act, fetched 26 Oct 2026 58
- 1 Jan 2026Today
US prior-authorisation deadlines
- Who
- US Medicare Advantage, Medicaid and CHIP plans
- Butterfly gives you
- How often an agent gets a request right, and how long it takes, run after run, on your own cases.
What it asks, and the source
Prior-authorisation requests must be decided within 72 hours when urgent and seven calendar days otherwise. A prior-authorisation API follows from 1 January 2027.
Source: CMS, Interoperability and Prior Authorization final rule (CMS-0057-F), 17 Jan 2024 59
- 2 Aug 2026Today
EU AI Act, transparency duties
- Who
- Anyone using AI with people in the EU
- Butterfly gives you
- A check in every run that each agent says what it is, where your rules say it must.
What it asks, and the source
People must be told when they deal with AI: an AI voice or an AI interviewer says so.
Source: European Commission, in force 2026-07-27 51
- 8 Oct 2026Today
NAIC model bulletin on insurers’ use of AI
- Who
- Insurers in the US states that adopted it
- Butterfly gives you
- A record per AI system: what it was tested on, how often it passed, and who checked it.
What it asks, and the source
A model bulletin on how insurers use AI. As of 8 October 2026, 29 jurisdictions had adopted it and 4 more had rules or guidance of their own.
Source: NAIC, implementation map, 8 Oct 2026 55
- 7 Dec 2026Today
EU T+1, the first deadline
- Who
- EU market participants
- Butterfly gives you
- How often an agent matches and confirms a trade correctly inside the new window, run after run.
What it asks, and the source
Allocations and confirmations must go out faster, the first step of the EU’s move to T+1 on 11 October 2027.
Source: ESMA, Statement on T+1 preparations, 20 Jul 2026 60
- 1 Jan 2027Today
California’s automated decision-making rules
- Who
- Businesses using it for credit, housing, education, jobs, work allocation or healthcare
- Butterfly gives you
- Evidence of how each agent decides, case by case, with every run kept for review.
What it asks, and the source
A business already using automated decision-making for significant decisions (credit, housing, education, jobs and the allocation of work, healthcare) must comply by 1 January 2027.
Source: California Privacy Protection Agency, CCPA regulations, §7200, fetched 26 Oct 2026 57
- 11 Oct 2027Today
EU T+1 settlement
- Who
- EU market participants
- Butterfly gives you
- Test runs of the agents that clear breaks and send instructions, timed against the one-day window.
What it asks, and the source
Trades settle one business day after they are made. The final deadline also covers sending settlement instructions early.
Source: ESMA, Statement on T+1 preparations, 20 Jul 2026 60
- 2 Dec 2027Today
EU AI Act, high-risk duties
- Who
- Firms whose AI output is used in the EU, in those uses
- Butterfly gives you
- Evidence of how accurate and how steady each agent is, task by task, measured on your own work, for the technical file.
What it asks, and the source
Duties for high-risk uses, from testing to accuracy and human oversight: credit scoring, life and health insurance pricing, public benefits, recruitment, and AI that allocates or monitors work.
Source: EU AI Act, Annex III, fetched 26 Oct 2026 54
Source: European Commission, in force 2026-07-27 51
- 2 Aug 2028Today
EU AI Act, AI inside regulated products
- Who
- Makers of products covered by Annex I that carry AI
- Butterfly gives you
- Repeatable test runs of the AI part, kept with the product’s own records.
What it asks, and the source
The high-risk duties for AI that is part of a product already under EU safety law (Annex I) apply from this date.
Source: European Commission, in force 2026-07-27 51
A summary to plan with, never legal advice: Butterfly supplies the test evidence, and each company stays responsible for meeting each rule.
06 · The Agent Reliability Index
How reliable are agents,really?
Four findings from the public record, each with its source and date. The Index will run models × effort levels × task families, and score how often an agent passes.
Passing once is not passing every time.
Any agent right 75% of the timeArithmetic, Jan 2026
Source: Anthropic Engineering, 9 Jan 2026 9
Illustrative arithmetic, assuming independent tries.
APEX-AccountingJul 2026
Source: APEX-Accounting, arXiv, 29 Jul 2026 11
The two figures come from different models.
A customer gets one try. So the Index will report how often an agent is right across its held-back cases, not its average.
Passing once is not passing every time.
Any agent right 75% of the timeArithmetic, Jan 2026
Source: Anthropic Engineering, 9 Jan 2026 9
Illustrative arithmetic, assuming independent tries.
APEX-AccountingJul 2026
Source: APEX-Accounting, arXiv, 29 Jul 2026 11
The two figures come from different models.
A customer gets one try. So the Index will report how often an agent is right across its held-back cases, not its average.
How the Index measures it
Models × effort levels × task families, on held-back cases.
- 01Task familiesWork written for the Index: Reconciliations, claims triage, KYC alert review, billing disputes, contract review, clinical coding.
- 02Every model and effort levelEach model at low, medium and high reasoning effort, on the same tasks.
- 03Held-back casesReal cases kept aside from the start, never used to build or tune an agent: at least 20 a task, typically about a hundred.
- 04Scored on every caseHow often it is right across the cases, the rules it followed and the cost of a success.
A perfect streak proves less than it looks: 8 clean runs in a row still allow up to 31% failure.61Tested again every quarter, and after each major model release. No client’s tasks or results appear in the Index.
07 · Read more
Short reads,built on the record.
08 · Pricing
Free to read.A module to keep reading.
The first read is free. The module keeps reading every week; the managed tier adds a named analyst.
Your read
Free
One read of your company, its breakdown, its likely tasks and your film.
- Company-level public sources
- The breakdown and the likely tasks
- Your narrated film
Butterfly Research
£25–60ka year
Company-level monitoring of your market, feeding your task map, your quarterly report and our predictions.
- Rivals’ pages, job posts, your own wording, customer feedback
- An engine for your industry
- Signals that become tasks to test
Managed
£100ka year
For large accounts: named analyst hours, predictions with tracked hit rates, and escalations.
- A named analyst
- Predictions, scored when they land
- Escalations when something moves
Draft for review · subject to contractAn add-on to the readiness subscription (from £96k a year). Company-level data: never your people, never named individuals.