Pick a process that happens often, follows rules you can write down, costs little if a draft is wrong, and has someone who owns it. That is the whole test. A first AI project should be boring, frequent and easy to check, because the first goal is a working habit, not a dramatic result.
Why the first project matters more than the tool
A first project teaches the team how to review AI output, who approves what, and where the information lives. A well-chosen one earns trust. A badly chosen one (vague, high stakes, owned by nobody) teaches the team that AI is unreliable. The tool is rarely the cause of either outcome.
This is also where most small-business projects stall: the idea was too big, the rules were never written down, or nobody was responsible for checking the output. Writing the process down first fixes the second and third problems before any build.
The five tests
Score each candidate process against these five questions. Use plain words (yes, partly, no). Numbers would suggest a precision this exercise does not have.
| Test | The question | A good first project says |
|---|---|---|
| Frequency | How often does this happen? | Several times a week, so you learn quickly and the time saved is real. |
| Clear rules | Could you write the steps and exceptions on one page? | Yes. If two experienced people would do it differently, write the rule first. |
| Cost of an error | What happens if one output is wrong? | A person catches it before it reaches a customer, or the harm is small. |
| Available information | Is everything the task needs already in a system someone can connect? | Yes. Information that lives in one person's head is a documentation job first. |
| Named owner | Who is responsible for the result and for changing the process? | One named person with time to review the output every week. |
A decision table you can copy
The table below applies the five tests to five candidate processes in an imagined small service business. It is a hypothetical illustration of the method, not a QS client and not a recommendation for any real business.
| Candidate process | Frequency | Clear rules | Error cost | Information ready | Verdict |
|---|---|---|---|---|---|
| Sort and label new inquiries | High | Yes | Low (a person reads the label) | Yes (form and inbox) | Strong first project |
| Draft replies to common questions | High | Mostly | Medium (reviewed before sending) | Partly (needs approved answers) | Good second project |
| Weekly summary of open jobs | Weekly | Yes | Low | Yes | Good, but small time saving |
| Quote pricing | Medium | No (judgement) | High | Partly | Leave with a person |
| Handling a complaint | Low | No | High | Partly | Leave with a person |
Notice what the table does not say. It does not say pricing and complaints can never use AI. It says they are poor first projects, because the error cost is high and the rules are not written. An assistant can still help a person prepare for them later.
Decide what you will measure before you start
Choose one or two measures that come from records you already have, and note today's values so you can compare later. Examples: the time between an inquiry arriving and a first reply, the number of inquiries unanswered after a day, or the minutes someone spends on a weekly report. If you cannot measure the current state, that is a finding in itself, and recording it is a fair first step.
Do not set a target from someone else's case study. Your volume, tools and team are different. Set a target after you have two or three weeks of real baseline data.
Warning signs of a bad first project
- The goal is "use AI" rather than a named task with a before and after.
- The output goes straight to customers with no review step.
- It involves money, dates or promises the business must keep.
- The process changes every week, or nobody agrees how it should work.
- The information it needs is scattered across people's personal accounts.
- No one has an hour a week to review results.
Do you need an agent at all?
Often the best first project is a plain workflow: a fixed sequence of steps, with an AI model used for one step such as summarizing or drafting. Anthropic's engineering team draws the same line in its guidance on building effective agents: workflows follow predefined paths, agents decide their own steps, and it recommends finding the simplest solution that works before adding complexity (see sources). For a small business, "simplest" usually means a workflow with a person reviewing the result.
Who decides, and what happens next
The owner or the process owner picks the project. A consultant can help rank candidates, but should show the reasoning in a table like the one above so you can disagree with it. At QS Digital, this ranking is the output of the Diagnose phase; the next step is writing the process down (Map) before anything is built. You can also start with a single paid hour to rank your own candidates with Pat.
Sources and evidence
- Anthropic, Building effective agents. Primary source, read 2026-10-10. Used for the workflow-versus-agent distinction and the advice to start with the simplest solution.
- QS Digital selection method. QS's own judgement, not an industry standard. The example table is hypothetical.