Butterfly Tasks
Example company · preview
Butterfly Tasks is the software Butterfly is building to run in your own AWS account. It will find the work your people repeat, test AI agents on it against the bar your task owners set, and give you a blueprint for each task. Here is how it will work, inside Lakeholm Insurance, an example company.
Example company · preview
One case, eight steps
A Lakeholm Insurance policyholder asks for a hire car while their car is repaired. Follow that claim, and the task it belongs to, through the eight steps.
01 of 08
Installed in your company’s own AWS account, on your own Amazon Bedrock. Your admin installs it and invites people. Butterfly has no access to the account.
The aim: an hour or two for someone with basic technical knowledge.
The picture: Lakeholm Insurance’s AWS account (Ireland, eu-west-1), with Butterfly Tasks installed inside it beside Amazon Bedrock, the key store and Lakeholm’s own systems. Butterfly stands outside the account, with no access.
02 of 08
Connect email, CRM, claims or finance systems read-only, or upload files. Every record is read inside your account, and its personal details are masked before anyone sees it, with the keys in your own key store.
Each record lands in one bucket: Approved, Excluded with its reason, or Investigating.
The picture: Lakeholm’s claims mailboxes connected read-only. A policyholder’s email arrives, asking for a hire car while their car is repaired. Its 5 personal details are masked before anyone sees it, and the record is approved and linked to the task Arrange a hire car.
03 of 08
From the records, Butterfly Tasks will find the work people repeat. For each task: its team, how often it happens, how long it takes, what it costs a year, how much judgement it takes, and how often people get it right today, worked out per task, never per person.
You mark each one worth doing, not worth it or not possible yet, then take a recommended task set or build your own.
The picture: 8 of Lakeholm’s motor claims tasks ranked by what each costs a year, with Arrange a hire car lit at number 7. It is taken forward in the set Motor claims, first set.
Try itYour own company’s likely tasks, beside the same work at an example company.
Add your company’s website, and its likely areas and tasks appear here, beside the example below.
An example from insurance
Insuranceabout 1,000 peopleAWS Ireland · eu-west-1about 280 tasks found
Judgement1 block: follows fixed rules5 blocks: careful judgement, with reasons written down
Mark each one worth doing, not worth it or not possible yet, then take a recommended task set or build your own.
04 of 08
For each task in a set, a synthetic example is written from each real case, inside your account. Every personal detail is made up, every date moves by the same amount, amounts change but stay close, and the shape stays the same. Each example passes six checks before it counts.
The picture: the claim beside the synthetic example written from it. Every personal detail is made up, every date moved together, 5 days earlier, and the outcome is the same. Six checks, each passed.
Try itAnother of Lakeholm’s cases, a windscreen claim: step through it, then see a rewrite.
An example from Lakeholm Insurance
Real record4 personal details masked
Hello,
I’d like to claim for a cracked windscreen. It happened on 28 October.
I’ve attached photos and a quote. My number is Masked: phone.
Thanks,Masked: name
On the claims system€660Paid, replacement
Intake masked 4 personal details before anyone saw the record. The keys stay in your own key store, in your account.
Next, a synthetic example is written from this case.
The task’s rules
Intake masked 4 personal details before anyone saw the record. The keys stay in your own key store, in your account.
Next, a synthetic example is written from this case.
The six checks
Every name, date and amount here is made up.
Step 1 of 3: the record as intake masked it.
05 of 08
The people who do the task walk through it on a practice desktop with synthetic cases, never real ones. Their steps become the approved procedure and an eval.
Butterfly experts, your own staff, watch the replays and note each step. The task owner approves the procedure.
The picture: a claims advisor walks through the task on a practice copy of the claims system, with a made-up case. One step is changed as they explain it, their six steps become the procedure, and the task owner approves it.
06 of 08
Agents are tested on held-back real cases, kept aside and never used to build or tune them: at least 20 a task, typically 100. Each is scored three ways, the right outcome, the rules followed and the exceptions missed, and its outcome is set beside your people’s.
Every task will run on eight models, four frontier and four open-weight. Butterfly’s pick is the approved model that clears the bar for less than any other.
| To be | Right outcome | Rules followed | Exceptions missed |
|---|---|---|---|
| Clears the bar | 95%97% when the work takes judgement | 98% | None |
| With a person checking | 85%90% when the work takes judgement | Not checked | No more than one |
The four verdicts
The picture: the hire car agent tested on 100 held-back cases. 97.9% right against a bar of 95%, rules followed in 99.1% against 98%, no exception missed, beside people’s 90.9% today: it clears the bar. Eight models, unnamed, work through the same cases.
07 of 08
For each task: how an agent would do it, where a person stays in, what access it needs and how it is judged. It comes as a report and as an agent pack that drops into your own agent system.
Tasks no agent can do yet go to the Butterfly lab.
The picture: the blueprint for Arrange a hire car. The agent does 4 of its six steps alone and decides 2 with a person checking 1 in 20, and the agent pack’s files are ready for Lakeholm’s own agent system.
08 of 08
Your company runs its agents. The task owner approves going live and how many cases a person checks. Live agents will be retested every week on fresh cases, and each quarter a fuller retest shows what a newer model, extra training or a new rule would add.
The picture: the task owner approves go-live. The hire car agent goes live with a person checking 1 in 20, and it is retested every Monday on fresh cases.
Your account
Personal details are masked at intake, every example is synthetic, and nothing leaves without the approval of a person you name.
May leave, with the approval of a person you name
Never leaves
Personal data about your customers never leaves your account.
If you agree, reviewed task sets, and synthetic datasets made from your records, can be licensed to AI labs. You approve every licence and earn royalties from each one. Your held-back cases and readiness tests are never licensed.
Your proposal
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:Agent benchmarksPricing