Where should a business start with AI?

Many businesses start with AI by buying accounts, running a training session or asking the team to “find ways to use it”. After a few weeks, people may produce content faster, but the way the business operates barely changes.

The reason isn't that the tools aren't powerful enough. It's usually that the business hasn't chosen a problem clear enough to work on, measure and own together.

1. Start from the process, not the tool

Pick a task that happens regularly: compiling weekly reports, researching customers before meetings, processing documents, looking up internal policies or preparing content drafts. Map the current process with five questions:

• Who does the work?
• Where do the inputs come from?
• What are the main processing steps?
• Which output counts as acceptable?
• Which exceptions usually stop the process?

A process that can't be described in words shouldn't be automated with AI yet.

2. Measure the current state before testing

You don't need a complex dashboard. For the first pilot, a business can record four simple things: processing time, number of reworks, number of significant errors and internal user satisfaction.

A baseline stops the team from arguing on gut feeling. It also prevents attaching an attractive ROI figure to the project before there is enough data.

3. Choose a pilot small enough but valuable

A good pilot usually repeats often enough, has fairly stable inputs, outputs that can be checked and risks that can be contained. If the process involves legal, financial or HR decisions or sensitive data, there must be a stopping point for human review.

An example close to SMEs

Instead of building an “executive assistant for the whole company”, start by consolidating the weekly report from three defined sources. AI can help standardise the data, flag missing parts and produce a summary; the manager still reconciles it before sending it to the CEO.

4. Write the pass criteria before building

A pilot needs clear pass/fail criteria. For example: less manual handling time under the same conditions; no required data fields missed; every conclusion traced back to its source; and the owner can handle it when the system fails.

Criteria should cover business value, output quality, user acceptance and governance requirements.

5. Hand over before scaling

If only one person knows how to run it, the business doesn't have a system yet — just a demo that depends on one individual. Before scaling you need a user guide, access rights, a list of exceptions, a way to report errors, a change approver and a review schedule.

1. Pilot with a small group of users.
2. Log errors and exceptions.
3. Update the process and documentation.
4. Train the main operators.
5. Decide to stop, fix or scale.

The best starting point

The best starting point isn't the most impressive-sounding use case. It's a process important enough to create value, clear enough to measure and small enough for the team to learn how to implement AI responsibly.