Deals & diligence
Is the target’s AIas ready as the deck says?
The Agent Target Benchmark tests a company’s AI agents on its own workflows, inside the deal timetable: £50k fixed, 2–4 weeks.
“Our agent resolves 8 in 10 tickets.”
68% resolved, between 61% and 74% at 95% confidence, over 1,600 runs
Why now
AI is in the deal now.
of 2025 tech deals had an AI component, against about one in four in 2024.
Source: Bain & Company, 27 Jan 2026 34
On APEX-Accounting, models score well on average, yet almost none pass a task every time. The two figures come from different models.
On APEX-Accounting, the top average score is 56.4%, yet the top rate of passing all 8 tries is 2.6%; the two figures come from different models.
Source: APEX-Accounting, arXiv, 29 Jul 2026 11
01 · The Agent Target Benchmark
One price.Inside the deal window.
- 2–4 weeks, fixed in scope
- No install in the target’s cloud
- Model costs included
01 · Scope
The workflows the target says its agents run, and the claims in its deck, agreed in one call.
- 01ScopeOne call
The workflows the target says its agents run, and the claims in its deck, agreed in one call.
- 02Scenarios50–100
Simulated scenarios built from those workflows, each with a known right answer.
- 03Runs1,000–2,000
About 20 runs of each scenario, across the settings the agent runs on.
- 04ReadoutA screen
Every figure with its confidence interval, and each scenario where the agent failed, and how.
02 · For each side
One benchmark,four sides of the table.
Is the target’s AI as ready as the deck says?
Its agents run simulated scenarios built from its own stated workflows, and every figure comes back with its confidence interval, before you sign.
When Between the letter of intent and signing
- A screen of each agent, with its interval
- The scenarios it failed, and how
- Questions for the management session
0302 · Questions deal teams ask
Before you scope one.
No. Nothing is installed in the target’s cloud. We need access to the agent itself, for example a test account, and the scenarios are simulated from the workflows the target describes.
Two to four weeks from the scoping call, fixed in scope, so it fits between the letter of intent and signing.
£50k, fixed, for one benchmark. The model costs of the runs are included.
Neither. It is a screen: every figure comes with its confidence interval and the scenarios where the agent failed, so you can weigh it like any other diligence finding.
Whoever commissions it. The readout is yours to use in the deal, and we share it with no one else.
No. Nothing is installed in the target’s cloud. We need access to the agent itself, for example a test account, and the scenarios are simulated from the workflows the target describes.
Neither. It is a screen: every figure comes with its confidence interval and the scenarios where the agent failed, so you can weigh it like any other diligence finding.
Whoever commissions it. The readout is yours to use in the deal, and we share it with no one else.
0403 · Scope a benchmark
Tell us about the deal.
Three short steps. A person at Butterfly reads it and replies to arrange a scoping call.
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:Agent benchmarksHow it works