The most common advice for evaluating an AI tool is also the most dangerous: "try it on your real work."
It sounds rigorous. It isn't. When you hand a tool a live email, a fresh invoice, a new support ticket, you usually can't verify its output — not quickly, not with certainty. The reply sounds right. The summary looks complete. The categorization seems sensible. You're grading on vibes, and vibes are exactly what AI is good at producing.
If you can't verify the answer, the test tells you nothing. There's a discipline that fixes this, and it comes from an unexpected place: a biotech lab that pre-measured an entire experimental landscape just to grade AI agents fairly.
The lab's answer to the vibes problem
Last month, Leash Bio released Minos: pre-measured experimental worlds for AI agents. The team measured millions of small molecules against many proteins — the whole landscape, upfront — then let agents explore it at compute speed. Because every possible answer was already measured, each agent's choices could be graded against ground truth: what it said it would do, what it actually did, what the consequences were, and what would have happened had it chosen differently.
When the world is pre-measured, evaluation is repeatable and counterfactuals are free. Critically, a model can't bluff a measurement.
That's the principle worth stealing, and it has nothing to do with chemistry: grade AI against answers you already hold, not against your impression of its output.
A recent agent review made the same move at desk scale. A reviewer at the How I AI podcast, testing Meta's Muse agent, built a billing ticket with planted traps: the ticket was linked to the wrong customer account, the refund was a duplicate charge that needed the correct amount calculated rather than the requested amount blindly issued, and the ticket's text contained a planted instruction designed to sidetrack the agent. Every trap had a known right answer before the agent touched anything. The test wasn't "does the output look good" — it was "does the agent catch what a careful human would catch."
The 20-minute known-answer test
Here's the small-business version of both:
Step 1: Pick one task family. One job you're considering handing to a tool: invoice categorization, customer-email replies, ticket triage, quote drafting, refund handling.
Step 2: Pull five inputs with known ground truth. From your history — last month's invoices where you know the correct categories, your trickiest customer email where you know what the right reply looked like, an old support ticket where you know the resolution, a quote request where you know which jobs were profitable. The rule: for each input, you must be able to state the correct answer in one sentence before you run the test.
Step 3: Run all five through the tool. No hints. Give it the same instructions you'd actually give it on a Thursday afternoon — no prompt-engineering heroics.
Step 4: Score pass/fail per input. Not 1-to-5. Not "pretty good." Binary: did it produce the answer you already know is right, with no invented facts and no missed traps? A near-miss on money is a fail. A confident wrong answer is a fail — a worse fail than "I'm not sure," because confident-wrong is what actually damages a business.
Step 5: One failed input kills the tool for that job. Not for every job — for that job family. A tool that gets 4/5 on invoice categorization isn't "80% there"; it's wrong on real money 20% of the time, which means you're reviewing every output anyway and the tool saves you nothing. Move on, try the next tool, or keep that task manual.
Where this works — and where it doesn't
The known-answer test works wherever you have history: records you've already corrected, replies you've already sent, numbers you've already checked. That covers more of an SMB's AI workload than people expect — invoicing, triage, categorization, reply drafting, quoting, and refund logic are all history-rich.
It doesn't work for genuinely net-new work: a market analysis you've never done, a first-draft strategy with no prior version. There, no answer key exists, and you fall back to self-scored trial runs — with the honest caveat that subjective scoring is exactly where AI tools tend to look better than they are. Treat a tool that only passes the subjective test as probationary until it accumulates some history.
Three boundaries worth writing down:
- Ground truth must be unambiguous. If two careful people in your business would answer differently, the input isn't a test — it's a debate. Pick inputs with one defensible right answer.
- Don't run money-moving tests on live systems. Known answers first, inside a test setup; a tool that fails the known-answer stage never earns real access.
- Re-run the five inputs on major vendor model updates. Agent behavior changes overnight with a model swap — in both directions. The known-answer file is what makes that recheck cheap.
Tonight
- Pick the one task you've been telling yourself to try an AI tool on.
- Pull five past instances where you can state the right answer in one sentence each.
- Run them through the tool, score pass/fail, and write the verdict down with the date.
- One miss: the tool doesn't get that job. Five passes: it earns the next, bigger test.
"Try it on your real work" is how tools with great demos end up in your stack. "Try it where you know the answer" is how you find out which ones deserve to stay.
Want the boring half pre-built? The Boring Automation Pack from SMB Scale Up includes plug-and-play automation blueprints plus the standing-instructions worksheet — the checklists and tracking files that keep delegated work inside lines you actually wrote down.
Top comments (0)