Enterprise Search
Acceptance Testing AI Agents for Business: Measure Errors, Clarifying Questions and Rework Separately
In a pilot, an AI agent often looks good. Whether it saves more work in day-to-day operation than review and maintenance cost is something a success rate alone won’t tell you. For that, you have to record errors, clarifying questions and rework separately. This post describes an acceptance test for the office that you can run with your own cases.
by Sebastian @searchit · · 10 min read
AI-generatedAnswering emails, entering appointments, preparing quotes – these are typical examples of AI agents in the office, and each of them uses several tools in a row. In a pilot, usually only one thing counts: did the right result come out at the end? For deciding whether to put an agent into operation, that isn’t enough for me. An agent that gets nine out of ten jobs right can still create more work than it saves – because it keeps asking questions, because someone has to check every result, or because the one error was a faulty email to a customer.
A note up front: we make searchit. The acceptance test below is a draft method from our own work, based on OpenAI’s Evaluation best practices. It is not a standard. My advice for introducing AI in your business: adapt the test to your own workflows and check every result against your own data – with every vendor.
In short
- A single rate hides too much. Count errors, clarifying questions, rework and review effort separately.
- Errors come in classes. A wrong sentence in a draft is something different from an email to the wrong recipient. An error with external impact never goes into the average.
- The expected result is fixed before the test – by someone who knows the workflow, not by the agent and not after the run.
- Testing happens without external impact: as a draft, in a dry run, or with a test account that has realistic permissions.
- The balance sets the time saved against the effort for review, rework and maintenance – and is drawn up again after every change.
Why a success rate is not enough
“The agent handles 90 percent of cases correctly” sounds like a clear answer. For operation, it leaves the decisive questions open:
- How did it finish? On its own, after two clarifying questions, or only after someone reworked the result?
- What does review cost? Does every result have to be checked, or are spot checks enough?
- What happened in the remaining cases? Did the agent stop and report it – or deliver a wrong result that nobody noticed?
- What does upkeep cost? Instructions, permissions and tools change; a new language model behaves differently from the old one.
In an AI workflow with several tools, errors often arise at the hand-offs. The agent picks the wrong tool, passes a correct value on incorrectly, or keeps going even though a tool reported an error. You see little of this in the final result – but you do in the run log.
Four counts instead of one grade
Record four areas separately for each test case. Only together do they give a meaningful picture.
1. Errors – by class
An error is a result that deviates from the expected result defined beforehand. Assign every error to a class:
| Class | Example | How it counts |
|---|---|---|
| critical | email to the wrong recipient, appointment with the wrong customer, record overwritten, confidential information passed on | individually, never in the average – one case is enough for “not accepted” |
| factual | wrong deadline, wrong amount, wrong contact person in the draft | as a rate, with a threshold |
| formal | salutation, formatting, tone | as a rate, with its own threshold |
Also note whether the error was noticeable. An agent that reports “I can’t find a contract for this customer” has not done its job, but it has disclosed the gap. An agent that quotes an invented contract instead is more dangerous – even though both cases count as “not done”. Distinguish these cases in your records.
2. Clarifying questions – justified or avoidable
Clarifying questions are not bad in principle. An agent that asks before sending an email with external impact is doing exactly the right thing. So distinguish:
- justified questions: the information really was missing, or the action needs approval.
- avoidable questions: the information was in the job, in the document or in the mailbox, and the agent didn’t find it.
Avoidable questions point to missing knowledge or unclear instructions – often the cheapest problem to fix.
3. Rework – in minutes
Rework is everything a person changes in the result before it moves on. Measure it in minutes rather than as yes or no. A draft in which only two words need correcting saves almost all of the time; one you have to rewrite saves nothing.
4. Review and maintenance
Record the review time even when the result was correct. On top of that comes ongoing upkeep: sharpening instructions, adjusting permissions, retesting after a model change. Leave that out and you are flattering the agent – and half a year later you’ll wonder why the savings never actually showed up.
The acceptance test in eight steps
The procedure is a draft for the office. It requires no programming, but it does need a few hours from someone who knows the workflow. Before the first run, define both the expected results and the thresholds from step 6.
- Define one task narrowly. Not “the agent helps in sales”, but “the agent answers enquiries about delivery dates from the mailbox and creates a follow-up reminder”.
- Put together a test set of real cases. Better a few real cases than many invented ones. Deliberately include the difficult ones: incomplete enquiries, contradictory information, and cases in which the agent should do nothing.
- Define the expected result beforehand. For each case, someone who knows the workflow writes down what would be correct – before the agent runs. Otherwise the expectation unconsciously shifts towards the result.
- Run without external impact. Emails as drafts, workflows in a dry run, a test account with the permissions of a real role. A test account with full access measures something different from operation.
- Fill in the four counts for each case – in the scoring sheet below. To do so, read the run log, not just the result.
- Define thresholds beforehand. For example: no critical error, factual errors below a limit you set yourself. Evaluate the run against these predefined thresholds. A critical error ends the acceptance, no matter how good the rest looks.
- Draw up the balance. Compare the previous manual processing time with review time, rework and estimated upkeep. If too little is left, that is a result – not a failure.
- Repeat after every change. New language model, new instructions, new tool, changed permissions: the same test set, the same expected results, comparison with the last run.
The scoring sheet
| Case | Expected result reached? | Error class | noticeable? | Clarifying questions (justified / avoidable) | Rework (min.) | Review time (min.) |
|---|---|---|---|---|---|---|
| 1 | ||||||
| 2 | ||||||
| … |
At the end, the results of the four areas stand side by side instead of disappearing into a single grade. In our own measurement runs on document enrichment by language models, we also work with a reference version defined in advance. We record separately what the model found correctly, missed and invented.
What an agent should offer for acceptance testing
How well an agent can be acceptance-tested also depends on the software it runs in. Ask every vendor:
- Is there a dry run? Can a workflow be played through without anything going out?
- Who approves? Is drafting a workflow separate from approving it?
- What does the confirmation show? Does the user see before sending what actually goes out – recipients, attachments, text?
- What are the permissions tied to? To the agent, or to the user it works for?
- Can you trace which model produced what? Only then can a model change be compared cleanly against the old run.
- Are corrections recorded? Rework that isn’t written down anywhere shows up in no balance.
How searchit supports acceptance testing
searchit is the memory beneath your AI agents and comes with its own agent, the searchit AI Agent. For the acceptance test, this means:
- Play workflows through in fast-forward. Since version 8.0, a user describes a workflow in their own words, and the agent turns it into a draft. searchit checks it against the permitted building blocks and plays it through in fast-forward: which message would go to whom and when – without anything being sent. The workflow only becomes active after approval. Drafting and approving are distinct roles.
- Versions instead of surprises. Workflows are versioned. If you change a workflow after the test, a running case stays on its version – the next test run compares old against new cleanly.
- Confirmation before sending. Before the searchit AI Agent sends an email or text message (SMS) or deletes emails, a dialog at the workstation shows recipients, subject, attachments and text. “Don’t send” is preselected.
- Permissions tied to the user. What an agent may do via the interface depends on the roles of the signed-in user. If a role is missing, the server refuses and names it. A test account with a real role therefore checks the same roles that apply in operation.
- Provenance of every AI field. Every field that a language model generates during indexing carries the model, model version, timestamp and prompt identifier. After a model change, you can see which results come from the old run and which from the new one.
- Corrections to recognised cases stay visible. If searchit has recognised a case incorrectly, the user corrects it with one click and sees their corrections together with their effect in a list. For service agents, the agent management shows the latest jobs and errors.
How workflows, agents and memory mesh together is shown on the slide “One gear drives the next” on searchit-agent.ai.
Frequently asked questions
How many test cases does an acceptance test need?
There is no fixed number. The mix matters more than the quantity: real cases, including the difficult ones and those in which the agent should do nothing. If new types of errors appear in operation, those cases belong in the test set.
Who defines the expected result?
Someone who does the workflow by hand today – before the run and without knowing the agent’s result. For disputed cases, a second person helps.
What is a critical error?
Any error that has an effect outside the agent or is hard to undo: a faulty message sent out, information changed or deleted by mistake, confidential information passed on without authorisation. Which actions count in your case, you define before the test.
When do I have to retest?
After every change that affects behaviour: a different language model, changed instructions or skill, a new tool, different permissions. With the same test set, the comparison is done in a short time.
Is it enough if the agent asks before every action?
No. A confirmation question protects against external impact, but it costs time. Anyone who confirms dozens of questions a day eventually stops reading carefully. That is why the test also records clarifying questions.
Conclusion
For me, AI automation for business includes a balance that also captures the review effort. Under this draft method, an AI agent in your business is accepted when it makes no critical error, stays within the defined thresholds and, on balance, saves more time than review, rework and upkeep cost. A success rate alone won’t show you that. For that, you need the scoring sheet with its four areas and a test set of your own cases.
We’ll be glad to show you how a workflow is created in searchit, run in fast-forward and approved – on your own data, in a demo.
Related posts: Best Enterprise Search Software 2026 · Hardware for Local LLMs
Source for the method: OpenAI, “Evaluation best practices”, retrieved in November 2026. The acceptance test in this post is our own draft and not a recommendation by OpenAI. Product and company names mentioned are trademarks of their respective owners.
You might also like
AI-generatedEnterprise Search
AI Document Search: Checking PDF Tables in Your RAG System Without the Overall Score Hiding Errors
A RAG system with a very good overall score can still assign the decisive amount wrongly. Especially with multi-page PDF tables, headers, columns and page references get lost …
AI-generatedEnterprise Search
Permission to Read Is Not Permission to Act: How to Safely Test AI Agents with Write Access
An AI agent with read permissions finds information. With write permissions, it sends emails, changes records and creates appointments. Before you allow that, you need three …
AI-generatedEnterprise Search
AI Agent Memory for Business: Company Knowledge, Facts and Open Cases Compared
An AI agent without memory starts from scratch every day. But “memory” means something different depending on the vendor: the knowledge in your organisation’s data, facts about …
