Is the model quality the problem?

Rarely. Most off-the-shelf models are capable enough. The gap is context and where the tool sits relative to the work.

Do we need to build our own AI?

Not from scratch. It's usually about wrapping a capable model in your context, data and workflow — not training one.

How do we start without a big project?

One task, one workflow, one loop proven end to end — then expand from what worked.

You bought AI tools that don't fit

If you find yourself a few months past an AI rollout and the honest status is “we have the licences, but people aren't really using it,” the problem usually isn't the model. It's the fit.

Key takeaways

AI fails in operations when it's bolted beside the work instead of built into it.
Generic tools assume a generic process — yours isn't generic.
The fix is shaping AI to the workflow, not retraining people around the tool.

Why the off-the-shelf tool gets abandoned

It asks the team to change how they work to suit it. It doesn't have your context — your parts, your customers, your rules — so its answers are generic. And it lives in a separate window from where the work actually happens. So people try it once, shrug, and go back to the way that works. None of that is a model problem.

Built-in beats bolted-on

AI earns its place when it shows up inside the existing workflow, with your data and your rules, doing a step the person was already doing. No new tab, no new habit, no copy-pasting between tools. The best AI feature is often one nobody calls “the AI tool” — it's just the work, faster.

Start from the task, not the technology

Pick one repetitive task that involves text or judgement, give the AI the same context a person would have, and put the output where the work already lives. Prove that single loop end to end before adding anything. A narrow tool that nails one task builds trust; a broad one that does everything adequately gets switched off.

Keep a human where it matters

Fit also means knowing where AI drafts and a person decides. Tools that respect that line get trusted and kept; tools that overreach get quietly disabled. Designing that boundary deliberately is part of making AI stick.