Hiring is the use case where an unexamined model does the most damage, and where the damage is hardest to see from the inside. The engineering controls matter more here than anywhere else.
A lawyer does not want a summary of the indemnity clause. They want to be taken to it, with the words intact, so they can form their own view. That distinction changes the whole build.
Logistics paperwork is the canonical extraction problem, and the difficulty is never the model. It is that the input is a photograph taken in a lorry cab by someone in a hurry.
It is unglamorous, it is everywhere in finance operations, and it is a better fit for a language model than most of what gets demonstrated on stage — provided the model never touches the arithmetic.
In text, a two-second pause is imperceptible. In conversation it is the difference between a system that feels present and one people talk over. Almost every voice architecture decision follows from that.
A new model version is a dependency upgrade to the least deterministic component in the system, and it usually arrives without a changelog that describes what will change about your behaviour.
Conventional monitoring tells you the request succeeded. For a model-backed system that is nearly useless — the interesting failure returns 200 and says something wrong.
A reference architecture for getting data out of systems of record and into an application that can rely on it — sync, identity, permissions, and the tests that catch an upstream change before your users do.
Teams estimate token cost and are surprised by the bill. The gap is almost never the price per token — it is retries, context growth, and the runs nobody counted.
Per-seat pricing assumes a human is the unit of value. When the software does the work instead, the meter is pointing at the wrong thing — and the cost side is no longer fixed either.
A working method for deciding whether a change made an agent better or worse — how to build the set, what to measure, how to gate a release, and how to keep it honest as production moves.
A required human approval is often treated as a compliance formality bolted on at the end. Designed properly it is the most useful architectural constraint you will get.
AI in production
Eleven questions that separate a system from a demo. None of them are about model quality, and most projects that stall have not answered more than three.
LLM workflows
A prompt change alters system behaviour in production. If it can reach users without review, a test, or a way to roll back, then the most volatile part of your system is also the least governed.
LLM workflows
A model that returns prose is a conversation. A model that returns a validated object is a component you can build on. Getting from one to the other is mostly unglamorous contract work.
Data and integration
The model is rarely the hard part. Getting clean, current, permissioned data out of the systems of record is the hard part, and it is where most projects quietly stop.
RAG
When a retrieval-augmented system gives a bad answer, the model is almost never the reason. It answered correctly from the wrong documents, because retrieval handed it the wrong documents.
Agent architecture
Most problems described as needing a team of collaborating agents need one competent agent, a work queue, and a way to hand off to a person. The orchestration is the easy part to add and the hard part to debug.
Agent architecture
"You must never issue a refund over £500" in a system prompt is not a control. It is a request. The boundary belongs in the tool layer, where it can be enforced.
LLM workflows
The useful question is not whether a language model can do a step. It is whether the step tolerates being wrong occasionally, and what happens when it is.
Evaluation
An eval set is a sample from a distribution you chose. Production is a sample from one you did not. Almost every surprising failure lives in that gap.