Method · Technical support engineering · by Ayman Sbeiti — I support high-trust software platforms · Hiring?
Method
How I work with AI.
AI generates the scaffolding. Deciding what correct looks like, and checking that the output meets it, stays human. An AI assistant is only as trustworthy as the content and guardrails underneath it.
The division of labour
AI generates
- Scaffolding and first drafts
- Scripts and transformations
- The repetitive front of the queue
Stays human
- What correct looks like
- Verification against the real data
- Where the assistant must hand off
- Decisions that need accountability
The principles
- Keep the correctness criteria humanAI drafts the scaffolding: scripts, transformations, first passes. Deciding what the output needs to look like, and verifying that it does, remains my work.
- Test against real contentReview AI failures deliberately. Inaccurate or hallucinated answers are findings to fix at the source, not noise to tolerate.
- Fix the content, not just the behaviourWhen an assistant answers wrongly, the durable repair is usually in the knowledge it draws from.
- Bound what AI may handleDecide explicitly where it must hand off to a human. The boundaries are more consequential than the coverage.
- Adopt where it multiplies judgmentData preparation, first drafts, the repetitive front of a queue. Keep it out of decisions that need accountability.
Why this matters
The pattern across every place I've used AI in production work is the same: the leverage is real, and it is never unsupervised. Building a data-migration script, the AI generated the scaffolding. Deciding what the data needed to look like on the other side, and checking that it did, remained human work.
Configuring an AI support assistant taught the most durable lesson. The enablement took hours, but the real work was the review loop: reading its inaccurate and hallucinated answers and improving the underlying content until the answers could be trusted. An assistant's usefulness depends on the content underneath it, not the configuration on top.
The boundaries matter more than the capabilities. Deciding what an AI system must not answer, and where it hands off to a human, did more for user trust than anything it could handle. I evaluate AI tooling the way I'd evaluate any operational system: by its failure modes first.
The rule
Content debt becomes AI debt the moment the assistant goes live.