Key Takeaways

• Verification is a system, not a mood. Teams that trust their AI employee's numbers are not more trusting people; they set up sources-with-every-number, one-minute spot checks, and corrections that persist.
• Make the sources part of the deliverable. Every number should arrive with where it came from: the tool, the date range, the filter. A report you can trace is a report you can check in minutes.
• Spot-check one number, not every number. Pick the figure that would be most embarrassing if wrong and verify it against the source system. One deep check beats ten shrugs.
• The brief decides the accuracy. Ambiguous definitions ("revenue", "active users") produce confidently wrong output. Define terms once and make the definition permanent.
• Corrections must outlive the conversation. If you fix the same mistake twice, the problem is no longer the AI employee. Persistent memory turns one correction into a permanent rule.

The number that made it to the board deck

A finance lead asks her AI employee for last month's revenue summary on a Tuesday. The summary looks right, the formatting is clean, and the topline lands in Thursday's board deck. On Friday the CEO quotes it to an investor, and the investor's associate, working from the data room, quotes a different figure back. The gap turns out to be a definition: one number included a one-time services invoice, the other did not. Nobody lied. Two reasonable definitions of "revenue" walked into the same meeting.

That story is the real fear behind "can I trust the output", and it is worth taking seriously, because delegated work you have to silently redo is worse than no delegation at all. The answer is not to re-verify everything, which deletes the time savings. The answer is a light verification system that catches definition gaps and source errors in minutes. Here is the one that works.

Rule one: every number arrives with its source

The single highest-leverage change is to make provenance part of the deliverable. Not as a favor, as a standing requirement: every figure comes with the tool it came from, the exact date range, and any filters applied.

@Viktor reconcile last month's paid spend. Pull total spend per campaign
from Google Ads and Meta Ads for June 1-30, compare against the budget
column in the "Paid Media 2026" sheet, and flag anything more than 10%
over. For every number, note the account, the date range, and the
currency, so we can trace anything that looks off.

The output this produces is checkable by construction. When a campaign shows as 14% over budget, the row tells you it is the June 1-30 window from a specific ad account compared against a named column in a named sheet. If something looks wrong, you know exactly where to look, and the look takes two minutes instead of an afternoon of archaeology.

This also changes the AI employee's own behavior. Work that must show its sources cannot paper over a gap. If a connector returned partial data or a sheet tab was missing, that surfaces in the provenance instead of hiding inside a plausible total.

Rule two: the one-number spot check

You do not need to verify every figure. You need a habit: for each recurring deliverable, pick the one number that would hurt most if it were wrong, and check it against the source system directly.

For a pipeline report, that is probably the quarter's weighted pipeline: open HubSpot, apply the same stage filter, compare. For an ads report, the total monthly spend against the Google Ads dashboard. For a support digest, the ticket count against the Gmail label or Linear filter. One number, one tool, one minute.

The spot check does two jobs. It catches systematic errors, because a wrong extraction method is usually wrong in the headline number too. And it builds calibrated trust: after six consecutive weeks of the spot check matching, you have evidence, not vibes. Teams that skip this step end up in one of two bad places: blind trust that eventually gets burned, or permanent low-grade suspicion that makes the delegation pointless.

Rule three: fix the brief before blaming the output

Most "the AI got it wrong" incidents trace back to a definition that was never made. Here is what that looks like in practice:

The pattern is always the same: a term that has two defensible meanings, and a human who did not say which one they meant. A colleague would ask, and a good AI employee often does ask, but the reliable fix is upstream. Write the definition into the request the first time, and you will never have the board-deck moment from the intro. The craft of doing this well is its own topic, covered in how to write a brief your AI employee can run with.

Rule four: ask for the confidence report, not just the answer

For anything that feeds a decision, add one line to the request: "note anything you could not verify and anything you had to assume." The answer changes character immediately. Instead of a smooth wall of certainty, you get the honest edges: a connector that only returned 90 days of history, a spreadsheet tab that had not been updated since May, an assumption about currency conversion.

This is the same information a careful human analyst would volunteer, and it separates output you audit from output you review. Auditing means hunting for hidden problems. Reviewing means reading the stated caveats and deciding whether they matter. The second is ten times faster, and it is only possible when the caveats are stated.

Research deliverables deserve the same treatment: claims with links, and a clear line between what a source says and what is inference. The delegation pattern for that is covered in how to delegate research to an AI employee.

Rule five: corrections must stick

The most expensive verification failure is the repeated one. You catch the mistake, explain the fix in the thread, and three weeks later the same mistake is back, because the explanation lived and died in that conversation.

This is where an AI employee has to behave like an employee. When you correct Viktor, "we report net revenue, always exclude the services invoices", the correction becomes a persistent skill, shared across the team, applied to every future report. One person's fix becomes everyone's default. If your current tool cannot do this, you are not verifying work, you are grading homework forever. How that memory works, and how to seed it deliberately, is covered in how to give your AI employee memory.

A useful monthly habit: reread the recurring deliverables and ask what you silently fixed by hand this month. Anything you fixed twice should become a stated rule. The goal is a system where every mistake is only possible once.

What this looks like after a month

Week one, you spot-check everything and write definitions. Week two, the provenance is standard and the caveats arrive unprompted. Week four, the Monday report lands, the one-number check matches for the fourth time, and you read the caveat section in ninety seconds before forwarding it. Nobody redid anything.

Verification never disappears entirely, and it should not: the approval gates on outbound actions, the ones that keep an AI agent from acting without asking, stay in place permanently. What disappears is the anxiety, replaced by a short, boring routine, which is exactly what trust in a colleague looks like.

Frequently Asked Questions

How can I check whether my AI employee's numbers are correct?

Require every figure to arrive with its source: the tool, the date range, and the filters applied. Then verify one high-stakes number per deliverable against the source system directly. A report with traceable provenance can be checked in minutes, and a weekly spot check that keeps matching builds evidence-based trust.

Why does AI produce wrong numbers even when the data connection works?

The most common cause is definitional ambiguity, not broken data. "Revenue", "active users", and "open deals" each have multiple defensible definitions, and an unstated definition produces a confidently reasonable-looking wrong answer. Writing the definition into the request removes the largest error source before any data is pulled.

Do I need to check every report my AI employee produces?

No, and trying to deletes the benefit of delegating. Check every deliverable's single most consequential number, read the stated caveats, and keep human approval on outbound actions. Full re-verification is only warranted for a brand-new deliverable type or after a correction, until the fix is proven.

What should I do when I catch a mistake?

Correct it in the thread and make the correction permanent. With Viktor, a correction becomes a persistent skill shared across the team, so the same mistake cannot quietly return next month. If you find yourself fixing the same issue twice, turn the fix into an explicit standing rule rather than a conversational note.

Can an AI employee tell me what it is unsure about?

Yes, if you ask for it. Add "note anything you could not verify and anything you assumed" to requests that feed decisions. You will get the honest edges of the work: partial data windows, stale source tabs, assumptions made. Reviewing stated caveats is far faster than hunting for unstated ones.

Is AI-produced analysis reliable enough for board or investor material?

It can be, under the same standard you would apply to a junior analyst: sources attached, definitions stated, one independent spot check by the owner, and a human review before anything leaves the company. The failure mode to avoid is not AI involvement; it is unverified numbers of any origin reaching external audiences.