Where AI pays off: how we run an AI audit

Most teams we meet already have a list of AI ideas. It usually comes from a workshop, a few vendor demos and a handful of requests from people who tried a chatbot at home. What the list lacks is an order. Every idea looks promising in a slide, and the team has budget for one or two.

An AI audit turns that list into a ranked roadmap. It is a short, fixed-scope engagement, and its output is a decision: which use case to build first, what it needs, and how we will know it worked.

Start from the work

We begin with the work itself. For each team in scope we ask three things: what tasks take the most hours, which of those tasks follow a pattern, and where mistakes are expensive.

The answers rarely match the idea list. A support team might ask for a customer-facing chatbot, when the bigger cost sits in agents copying order details between three systems before they can answer anything. An operations team might ask for forecasting, when the real delay is a weekly report someone assembles by hand from spreadsheets.

So we sit with the people doing the work. We watch a support shift, follow a purchase order from request to approval, and read the documents a new hire is given in week one. That is where the useful use cases show up.

Score each use case

Every candidate gets scored on the same five questions:

• Value: how many hours, errors or lost sales does this task cost today, measured from the company's own data where it exists?
• Data: does the information the agent needs exist, is it reachable through an API or database, and is it current?
• Integration: which systems does the agent need to read from and write to, and how much access do those systems allow?
• Risk: what happens when the agent gets it wrong, and can a person catch the mistake before it reaches a customer?
• Adoption: will the people who own this task use the tool, and who on their side owns it after launch?

A use case with high value and missing data drops down the list, however exciting it sounds. A modest use case with clean data, simple integration and an eager owner often goes first, because it ships quickly and builds trust for the harder ones.

Check the data early

Data is where most AI projects stall, so we check it during the audit instead of finding out in month two. For each top candidate we pull samples from the real systems: a week of support tickets, the product catalogue, the policy documents a knowledge assistant would answer from.

We look for gaps that would break an agent. Product data split across two systems that disagree. Policies that exist only as scanned PDFs. Order histories that cannot be queried by customer. Each gap gets a size and an owner, because fixing it is part of the project plan.

Decide how success is measured

Before anything is built, we agree on how the first use case will be measured. For a support agent that might be the share of tickets resolved without a hand-off, the time to first answer and a weekly sample of answers reviewed by the support lead. For an internal assistant it might be how often staff use it and how often its answers are marked wrong.

These measures go into the evaluation set we build for the agent, so the same questions that justify the project also test it.

What the audit delivers

At the end the team receives:

• A ranked list of use cases with the scores and the reasoning behind them
• A data and integration assessment for the top candidates
• A recommended first project with scope, timeline and fixed price
• The success measures and the first draft of the evaluation set
• The risks we found, and the controls each one needs

The audit is also where we say plainly when AI is the wrong tool. Some tasks are better served by a clean integration, a simple rule or a fixed form. Finding that early saves a project that would have looked impressive in a demo and failed in daily use.

Where to start

If you have more AI ideas than budget, an audit is the cheapest way to choose. Bring the list, the people who do the work and access to a sample of the data. We will come back with the order to build in, and the first project scoped and priced.