How I diagnose and check the work
What should AI do, and who checks the result?
A build can look right and still miss how your team works. I start with the problem, check the records and use the people doing the job to test the result.
Learn how the work really happens.
We agree which records and sources I can use. I ask the team what the terms and rules mean, then check that against the setup. Missing information stays visible.
The result still needs checking.
AI can produce work that looks right and isn’t. I use it throughout my work, but human domain expertise decides whether the output fits your business. The agreed rules, decisions and checks belong in the documentation.
If an AI tool becomes part of your process, we agree what it can read, what it can do and who checks its output. Prices, access and customer commitments need clear rules and approvals.
Find the cause before choosing the fix.
I trace how records and tools depend on each other and test possible causes against ordinary and difficult cases. A report can calculate correctly and still answer the wrong question.
Question whether the proposed build is needed.
We compare a simpler process, an existing setting, a connection and a custom build. I show the work early enough for your team to challenge it. We also agree who can change it and how it will be maintained.
Check what happened to the actual record.
We inspect the result with the person responsible. If it is wrong, we use what we learned to correct the work and its documentation.
An illustrative example
An invoice is missing. Where would we look?
1. Agree what makes the job ready to bill
Is it job completion, an approval or a contract date? Compare the team’s answer with what triggers the invoice today.
2. Follow the job record
Check the status, the customer link and what the connection reported. Did invoicing never start, or did it start and fail?
3. Agree the fix
A missing approval and a broken integration need different fixes. Choose the correction that addresses the observed cause.
4. Check the invoice
Use the agreed example. Check the invoice and linked job, then test that sending it again does not create a duplicate. Show the owner what to do if a transfer fails.
A hypothetical walkthrough of the method; it is not a client result or a recording of a live system.
Leave the explanation with the setup.
Your team gets the documentation, decisions and live training for the agreed work. We make clear who handles maintenance, so the setup does not depend on remembering what was discussed.
An ordinary test and an exception
Ask AI to summarize the signed agreement. Then give it a deal with terms from an older version. Check which record it used, whether it flags the conflict, and what it sends to a person.
What would you like to work through?
Bring a problem where the proposed answer still doesn’t feel right. We’ll discuss what needs checking.
Free, 30 minutes. Bring one example. No technical brief needed.