Who Needs to Be Involved in an AI Build

The role you cannot outsource

Somebody in your business has to be able to say, case by case, what the right answer is. Without that, there is no evaluation set, no accuracy measurement and no way to know whether the system works.

We can build everything else. We cannot supply that, and projects where it is unavailable stall at exactly the point where they should be proving themselves.

What we bring

• Engineering — the pipeline, integration, retrieval, deployment
• Evaluation design — how to measure this specific thing meaningfully
• Interface design for review and correction, which is a design problem more than a coding one
• Operational design — monitoring, cost control, failure behaviour

What you bring

1. A domain expert for two to four days across the project, mostly on the evaluation set
2. An owner who decides, with about half a day a week
3. Access to real data, early, including the awkward cases
4. The people who will use it, at demonstrations, with honest reactions

Book the domain expert's time before the project starts. Their availability is the most common cause of AI projects running late, and it is entirely predictable.

You do not need a data scientist

For the overwhelming majority of business AI, there is no model training and no statistical modelling. The work is engineering, evaluation and interface design.

Businesses that believe they need to hire a data scientist before starting usually delay by a year for no benefit.

After launch

Someone owns quality: watches the correction rate, reviews the evaluation results, decides when thresholds move. It is a few hours a month and it is what stops a good system quietly degrading.

That person should be in the business rather than at a supplier, because the judgement is about your work rather than about the software.

Frequently asked questions