One right answer proves little.
Right 3 in 4 sounds good. Across about a hundred held-back cases, that’s25 wrong answersYour task owner sets the bar
Private agent benchmarks
Know which AI you can trust with which job.
Built to test AI on about a hundred of your own held-back cases, in your own cloud, against the bar you set.
01 · What a task costs
Every task has a cost.We add it up.
A person’s time, the software and the checks, times how often it comes up. Then we would find which AI could do it, and how reliably.
An example business · Finance operationsIllustrative
Matching a supplier invoice to its purchase order
Five blocks = £1
- Doing it
- 10 minutes of an accounts-payable clerk, at £28 an hour with on‑costs
- about £5
- Checking it
- 2 minutes of checks and corrections, at the same rate
- about £1
- Software
- Finance system, AP automation, Supplier portal
- about £1
- A task
- about £7
How often, a week?
A week, from£660
A year, from£34k
And a mistake costs more than money: a customer, your name, a regulator’s question. So an agent has to clear the bar you set.Why held-back cases
02 · Why it matters now
AI takes on harder work every quarter.Trust hasn’t kept up.
The harder the job, the higher the stakes. Agents may soon handle a company’s harder work, but how many businesses would hand an AI their sensitive processes untested? Butterfly’s readiness tests are built to give you that confidence, task by task, and a return you can predict.
What the top agent can do
Real freelance projects completed, start to finish13
More than eight times in a year.
The top agent completed 2.5% of real freelance projects in October 2025, and about 21% in October 2026.
Source: Scale AI, Remote Labor Index, fetched 25 Oct 2026 13
- 25%of AI initiatives delivered the return expected of them.1
Source: IBM Newsroom, 6 May 2025 1
- 56%of people have made mistakes at work because of AI, and fewer than half are willing to trust it.8
Source: KPMG / University of Melbourne, 2025 8
- 46%of UK financial firms say they understand the AI they use in part, not in full.7
Source: Bank of England / FCA, 21 Nov 2024 7
- 76%of enterprise AI use cases are now bought in from vendors, up from 53% a year before.4
Source: Menlo Ventures, 9 Dec 2025 4
03 · How it works
Butterfly Tasks,step by step.
Butterfly Tasks is the software Butterfly is building to run in your own AWS account. Three of its eight steps:
- 01Set up
- 02Sources
- 03Tasks
- 04Examples
- 05Walkthroughs
- 06Readiness
- 07Blueprint
- 08Agents
- Step 03 of 08
Tasks: the work your people repeat, found in your own records and costed.
- Step 06 of 08
Readiness: agents tested on your own held-back cases, against the bar your task owner sets.
- Step 07 of 08
Blueprint: for each task that clears the bar, how an agent would do it and where a person stays in.
Why a verdict holds
Held-back cases, any agent, your own cloud: three things every verdict will stand on, each with its own page.
Right 3 in 4 sounds good, but across about a hundred held-back cases that’s 25 wrong answers. Why held-back cases
How agents are tested
Check any agent before you commit to it.
Each one is tested on your own task and leaves with a readiness score.
£20k per agent Draft for review · subject to contract
Check an agentBuilt to run in your own cloud.
Your cloud accountYour systemsButterfly TasksResultsZero access for Butterfly staff
Your records are read inside your own account. Personal data about your customers never leaves it.
Trust & security04 · Price
Start with a Diagnostic.Then keep it tested.
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.
Your first yearDraft for review · subject to contract
First · 8–12 weeks from data accessReadiness Diagnostic£60kfixedYour tasks found and costed, the first ones tested, with blueprints.
Year one · Readiness Diagnostic £60k + Function £96k£156k
05 · The data program
Get paid for your work.Licensed, not handed over.
The tasks we find, and the synthetic data made from your records, could also be licensed to AI labs, if you choose: each made inside your cloud and approved by you. A licensed task is worth $1,000 to $2,000.
Your taskThe closed caseThe records stay in your cloud
A rewritten copyPersonal details rewrittenEvery example checked six ways, in your accountApproved
Lab 1Trains a model
Lab 2Tests an agent
Licensed ×2Licence income, back to youTasks you licenseIllustrative
$100k–200kof value for 100 tasks, at $1,000 to $2,000 each
One block, ten tasks. A large organisation may have around 10,000, and new ones every day from real work. Your own readiness tests are never licensed.
06 · Industries
Ten industries to start in.And room for more.
Where we expect strong sets of tasks to test. Each starts with work every company has, like closing the books, then its own. Open one to see what we would test first there.
- Banking and creditFirst Clearing a KYC or AML alert
- InsuranceFirst Triaging a first notice of loss
- Healthcare administrationFirst Preparing a prior-authorisation packet
- Every first batchAccounting and finance operationsFirst Reconciling a bank account
- LegalFirst Reviewing a contract against the playbook
- Pharma and life-sciences regulatoryFirst Processing an adverse-event case
- Asset and wealth managementFirst Reconciling a fund’s NAV
- Telecom and customer operationsFirst Diagnosing a fault with the customer
- Public-sector casework and permittingFirst Validating a planning application
- Management consultingFirst Sizing a market
Your proposal
Get your proposal.Then register your interest.
A readiness report made from your company’s public website: what your tasks cost, and the first 25 to test. If it reads right, register your interest: a person from Butterfly will call you to talk it through.
Or read on:How it worksPricing