Start from work, not technology

The most common reason AI projects disappoint is that they begin with a tool looking for a problem. A better starting point is a list of the work your teams do every week: the requests they handle, the documents they read, the decisions they make and the hand-offs between them.

Good first candidates share three traits. The work happens often. It follows a recognisable pattern, even when the inputs vary. And the cost of a slow or inconsistent result is visible to someone who cares about fixing it.

Score every candidate on four questions

  • Value: if this worked, what would change? Hours saved, revenue gained, errors avoided or faster response. Put a number on it, even a rough one.
  • Feasibility: does the data exist, can you get to it, and could a capable new employee do the task with the same information?
  • Risk: what happens when the AI is wrong? A mistake in an internal draft is cheap. A mistake in a refund, a contract or a medical message is not.
  • Ownership: is there a named person who wants this, will test it and will be accountable for the result?

Score each from one to five and multiply value by feasibility to rank the list. Use risk to decide how much human review the first version needs, rather than to rule ideas out entirely. Treat ownership as a gate: no owner, no project.

Apply the new-employee test

A useful shortcut: imagine handing the task to a smart new hire in their first week, with access to the same documents and systems the AI would have. If they could do a reasonable job after reading the instructions, the task is a strong AI candidate.

If the work depends on knowledge that is written down nowhere, the first step is not AI at all. It is capturing that knowledge: interviewing the people who hold it and turning their judgement into clear rules, examples and exceptions. That work pays off whether or not AI is involved.

Measure the baseline before you build

Before writing any code, measure how the work performs today: volume per week, time per item, error or rework rate and response time. Without a baseline, even a successful project ends in an argument about whether it worked.

Agree the target too. "Halve first-response time on WhatsApp enquiries" is a target. "Use AI in customer service" is not. A good target names the metric, the direction and roughly how far it should move.

First projects that tend to work

  • Answering repeat customer questions from a knowledge base, with hand-off to a person.
  • Qualifying inbound leads and booking meetings.
  • Extracting data from invoices, forms or contracts into a system of record.
  • Drafting first versions of proposals, reports or listings for expert review.
  • Summarising calls, meetings or case histories into structured notes and tasks.

Each is frequent, measurable and easy to supervise, which is exactly what a first project needs. Success here builds the data connections, governance and confidence that bigger projects depend on.

What to avoid the first time

  • Fully autonomous decisions with legal, financial or medical consequences.
  • Use cases that depend on data you have not collected yet.
  • Projects without a clear owner, or that span five departments at once.
  • Building a general "company AI" before solving one specific problem well.