Film 1 · Agent benchmarks
Your agent benchmark
346 words1:51 voicedNarration: AI voice
00 Introduction, for everyone
Every company runs on thousands of tasks, each one costing hours, salaries and software. AI can take on more of them every month. But which ones can you trust it with? Now let me show you, for your business.
01 Read and price
This is your company, from the outside: your website, your posts, your own words. We read all of it: what you do, how you work, and the thousands of tasks behind it. Take this one: matching a supplier invoice to its purchase order.
Each one costs about £7 in hours, salary and software. It happens hundreds of times a week, all year.
02 Test it
Your vendor says its agent can do it. But one wrong answer reaches a customer, with your name on it. So we’d set it up in your own cloud: your records, your rules, no access for our staff. It would face a hundred held-back cases.
Right three times in four sounds good. But across a hundred cases, that’s twenty-five wrong answers. We’d test every option: your own agent, and eight models, frontier and open-weight. Two clear the bar, and you’d see what each costs to run.
03 Approve
For the one you choose, you’d get its blueprint: the access it needs, where a person stays in, the test to keep passing. The task owner sets the bar, and approves it going live.
We’d test it again every week, and fully each quarter: models change. As agents take on more sensitive work, this is how your company knows which to trust, and what each job is worth.
04 What comes next
That was one task. Behind it are thousands more. We start with the twenty-five worth more to you than the rest, and test each one the same way.
Your readiness report sets them out: what each costs today, what you could save, and how it would run in your own cloud. And the tasks you choose to license could earn you income from AI labs, while your records never leave.
Agents will take on more of your work every year. With Butterfly, you’ll know which are ready. That’s your agent benchmark.