Over the last eighteen months we have shipped AI features into production for hospitals, manufacturers, lenders and retailers. The pattern that works looks nothing like the demos.

Start with a task, not a technology

The projects that succeed begin with a named task that costs measurable time: reading vendor invoices, answering the same forty support questions, or estimating next month's SKU demand. The projects that stall begin with “we should do something with AI”.

Write the task down with its current cost — hours per week, error rate, or delay in days. That number becomes the acceptance criterion, and it also tells you how much accuracy you actually need.

Retrieval before fine-tuning

For nine out of ten business use cases, retrieval-augmented generation over your own documents beats fine-tuning on cost, speed and auditability. Keep your knowledge in a versioned store, retrieve the relevant passages, and require the model to cite them.

“If the answer cannot be traced to a source document, it is not an answer — it is a liability.”

The four controls we always build

  • Citations: every generated answer links to the source passages used.
  • Confidence thresholds: low-confidence cases route to a human queue automatically.
  • Full logging: prompt, retrieved context, response and reviewer verdict are stored for audit.
  • Cost ceilings: per-tenant token budgets with alerts, so a bad loop cannot produce a bad invoice.

Data readiness is the real project

In most engagements, 60% of effort goes into data: extracting from legacy systems, normalising vendor formats, resolving duplicate master records and defining ownership. Budget for it honestly and the model work becomes almost boring — which is the goal.

Measure after go-live, not before

We instrument every AI feature with the same rigour as a payment flow: accuracy sampling weekly, human-override rate, latency, cost per resolved task. If those numbers do not improve on the pre-AI baseline within eight weeks, we turn the feature off and say so. That discipline is why clients let us try the next idea.

A realistic first project

Pick one document type, one department and one metric. Aim for a six-week production pilot with fifty real users, not a six-month lab exercise. Ship it, measure it, then expand.

If you want a second opinion on an AI idea before you spend, our consultation call is free and we will tell you plainly if software rules would do the job better than a model.