AI earns its cost where a task is language-shaped, high-volume, and tolerant of a checkable error. It disappoints where the task requires a single correct answer that nobody will verify.
Where it genuinely works
- Reading things. Extraction, classification and summarisation over documents and messages, where a human previously read everything to find the few that mattered.
- Finding things. Search over private knowledge where the searcher does not know the vocabulary the document uses.
- Drafting things. First drafts of routine written work that a person will review anyway — replies, summaries, descriptions, structured records.
- Triaging things. Sorting variable input into the right queue with the right context attached.
The pattern: high volume, language-shaped, and a human either reviews the output or the error is cheap and visible.
Where it disappoints
- Single-answer arithmetic. If a number must be exactly right, compute it in code and let the model explain the result, not produce it.
- Stable rule-following. If the rules are written down and do not change, a rules engine is cheaper, faster, auditable and does not need an evaluation set.
- Decisions nobody will check. The dangerous combination is plausible output plus no reviewer plus consequential action.
- Tasks without a definition of correct. If your own team cannot agree what a good output looks like, no system can be evaluated, and the project has no way to end.
The tell
Ask one question about any proposed use: how will we know it is wrong?
If the answer is a person reads it, a downstream system rejects it, or a check compares it to a source, the case is sound. If the answer is a shrug, the risk is not that the system will fail visibly — it is that it will fail invisibly for months.
The prerequisite nobody wants to hear
Most AI programmes that stall do so on data, not on models. The content is scattered, stale, contradictory or locked in formats nobody can parse. That work is unavoidable and it is the majority of the timeline.
The compensation is that it is not wasted. An inventory of your own knowledge, cleaned and permissioned, has value independent of any model you put in front of it.
AI & Intelligent Automation
Most useful AI work is not a model. It is retrieval, tooling, evaluation and the boring parts of integration done properly. That is the part we build.
AI & Intelligent Automation