← Back to perspectives
/Intellicon FDE Team

Before Accepting a Tool Built by an AI Agent, Define These Four Testing Requirements

Finding the best-value supplier can yield different answers depending on whether you compare unit prices or totals including shipping. Your acceptance checklist should clearly define input data, expected outputs, exception handling, and the basis for comparison.

AIEgentWrX人機協作導入
Before Accepting a Tool Built by an AI Agent, Define These Four Testing Requirements

Suppose you ask an AI Agent to build a procurement price-comparison tool. During the demo, it reads supplier quotes and recommends suppliers. But can you still rely on those recommendations when the quotes omit shipping costs or use different currencies? When reviewing a tool for acceptance, managers need to look beyond whether the demo produces a result and confirm that the software handles exceptions according to company rules. At Intellicon Solutions, our view is that teams should first define clear specifications so they can test repeatedly and verify results, then decide how many AI Agent roles the work requires.

According to iThome, AWS released AWS Emulator, developed by autonomous AI Agent teams without engineers writing or reviewing the product code. AWS did not, however, skip verification: the team generated tests based on existing specifications, compared how the emulator and the actual cloud services behaved under the same conditions, and assigned testing, discrepancy analysis, and fixes to separate groups. Enterprises can learn from this verification approach, but should not simply adopt a development model that bypasses human review.

Before Acceptance Testing, Define What Counts as a Correct Result

AWS has existing specifications and live services to compare against. Enterprises developing internal tools should likewise establish how they will determine whether results are correct. Business teams and IT should jointly prepare an acceptance checklist: business teams confirm the operating rules, while IT determines whether those rules can be turned into executable tests. This helps avoid discovering, only after development is complete, that the two sides understood the requirements differently.

For a procurement price-comparison tool, a requirement that merely says “find the best-value supplier” leaves the developer unsure whether to compare unit prices, totals including shipping, or total costs that also meet delivery requirements. Each test case should therefore clearly specify the following, so that anyone reviewing the same data can explain why the result is correct:

  • Input data: Specify the quote version, currency, quantity, and delivery schedule, and use a fixed dataset so that the data does not change between test runs.
  • Expected outputs: State what should be compared and which fields the results must include. For example, if procurement decides to compare totals including shipping, the results should include a cost breakdown and the basis for the calculation.
  • Exception handling: If shipping costs are missing or an exchange rate cannot be confirmed, the tool should flag insufficient data and refer the issue to procurement. It must not make assumptions about missing data and rank suppliers as usual.
  • Basis for comparison: Include the rules approved by procurement and manually verified results. If results from a legacy system will be used as a reference, first confirm that they meet the current requirements.

For text outputs, check that all required information is present, citations are correct, and no unsupported conclusions have been added. There is no need to require identical wording every time. However, if business teams and IT still disagree on the expected results, clarify the specifications or narrow the scope of the pilot first. Do not leave an AI Agent to decide the company’s procurement standards.

Keep Actual Test Results After Every Change

A successful demo proves only that a particular operation produced the result shown. It does not establish whether the tool will remain correct with different data or scenarios. Keep the demo data as a test case, then add scenarios involving missing fields, duplicate data, and service connection failures. IT should run the software in the existing test environment and verify both the actual outputs and any subsequent actions the software takes. If the tool writes data, confirm that the data was actually written and that its contents are correct, rather than relying on a success message on screen.

After each change, retest not only the cases that previously failed but also existing functions that may have been affected. For example, after changing how totals including shipping are calculated, verify that supplier rankings are still correct, rather than simply checking whether the cost fields appear. Test records should include the software version, input data, expected results, and actual results, along with the steps needed to reproduce the test so that another colleague can repeat the verification.

Business managers and IT should agree in advance on which errors will block release, such as incorrect monetary calculations, access to data outside authorized permissions, or skipped mandatory approvals. Record unresolved items in a fix list, clearly stating their impact and who is responsible for resolving them. Reports should not present only an overall pass rate: managers also need to know whether failed items could affect payments or commitments to external parties.

Skills in EgentWrX can capture standard practices, such as confirming specifications before development and listing test cases and discrepancies at delivery, so that an AI Agent can reuse them for similar work. However, Skills provide working rules; IT still needs to execute the actual tests in the test environment. Acceptance requires execution records. Having the inspection steps written down does not mean the checks have passed.

Agree on Each Group’s Deliverables Before Dividing the Work

Clarify each group’s tasks and responsibilities before deciding whether to add more AI Agent roles. Those preparing test cases should deliver the input data, expected results, and the basis for judging correctness. Those analyzing discrepancies should identify which cases failed and how the actual results differ from expectations. Those making fixes should explain what changed and which functions might be affected, then hand the work back to IT to rerun the relevant tests.

EgentWrX supports delegation through Sub-agents, allowing a primary AI Agent to assign work to other AI Agent members. Frequently collaborating members can also be organized into teams, with delegation capacity limits configured. Enterprises can divide the work according to the deliverables described above. However, agreement among different members that a result is correct does not mean it has been verified. Acceptance still requires checking actual execution results against specifications confirmed by the business team.

For your next development acceptance meeting, choose one clearly scoped internal requirement. Ask the business manager to confirm the acceptance checklist and IT to demonstrate how to rerun the tests, then review the failed items together. Before the tool enters operational use, the business manager should approve which tasks it may be used for, and IT should assess the technical risks. If the team cannot even explain what the results should be compared against, complete the specifications first. Adding more AI Agent roles will not solve that problem.

FAQ

Can a tool built by an AI Agent enter operational use after a successful demo?

No. A successful demo proves only that a particular operation produced a result. Keep the demo data and add test cases involving missing fields, duplicate data, and connection failures. IT should run the tests and record differences between actual and expected results, allowing managers to decide whether to approve release based on the impact of the failed items.

Should business teams or IT define the acceptance criteria?

Business teams and IT should define them together, with the business manager confirming the operating rules. Business teams need to explain which results are acceptable and which situations require human review. IT then turns those requirements into repeatable test cases. Where agreement is still lacking, clarify the specifications before development begins.

Can an AI Agent check its own development work?

An AI Agent can help generate test cases, but its own assessment that the work has passed is not enough to complete acceptance. Business teams should first confirm the basis for the expected results, then IT should run the tests and retain actual outputs and discrepancy records. The same requirements apply even when different AI Agent members perform the checks.

References

30 minutes to map out which work to hand to AI first

Want every employee to have their own AI teammate?

A consultant will be in touch shortly.

ISO/IEC 27001:2022 Certified ISMS