Where to start with AI in an SME: the first process that pays for itself

One process first, not the whole company

In a small company, the most common mistake is trying to "roll out AI" everywhere at once. It works far better to pick one process that pays for itself, get it running, measure the effect, and only then move on. This approach doesn't require an in-house AI team — all you need is one person on the business side who knows the process inside out, plus a partner for the technical side.

Start with a question that sounds simple: where do we lose the most time on repetitive work whose output can be checked? The answer usually points to your first strong candidate.

Three selection criteria

Score processes along three dimensions, each on a scale of 1–5.

• Volume — how many times a month the process repeats. The more it repeats, the larger the base for savings.
• Repeatability — whether the steps are similar every time. A repeatable process is easier to describe and to check.
• Cost of an error — what happens when the output is wrong. A low cost of error means a safer start.

The best first process has high volume, high repeatability and a moderate cost of error. A high cost of error doesn't rule a process out, but it does require a human in the approval loop.

Process	Volume	Repeatability	Cost of error	Priority
Draft proposals / RFPs	4	4	3	High
First response in support	5	4	2	High
Organizing documents	4	5	2	High
Pricing decisions	2	2	5	Low

Concrete processes for a first attempt

Three areas tend to work best in SMEs.

Proposals and RFPs

An AI assistant drafts a proposal from the inquiry and your earlier documents. A salesperson reviews it, edits it and sends it. The volume is often high, and the draft shortens the time from inquiry to reply.

First response in support

Instead of writing every reply from scratch, an agentic workflow classifies the ticket and proposes the content. A human approves it or edits it. You keep the cost of error low by giving the assistant nothing more than the "first draft" role.

Document work

Pulling data out of invoices, contracts or forms is a classic case: high repeatability, a measurable result. Here you quickly see whether the process pays for itself.

Assistant or agent to start

For the start, choose the simpler option. An AI assistant responds to a specific instruction and finishes the task. An AI agent works across multiple steps and decides on its own what to do next — it delivers more, but it demands more oversight and stronger safeguards.

Operator's rule: start with an assistant and a human in the loop. You move to an agent only once the process is described, measured and stable.

Output quality depends on the instructions

Whatever you choose, the result depends on how you frame the task. Prompt engineering is simply writing unambiguous instructions: what format the output should take, which data may be used, what must not be done. In practice this is usually a bigger lever than the choice of model itself.

Save those instructions as templates the whole team uses. That way the result is repeatable rather than dependent on whoever happens to be writing it.

What's next

Once you've chosen the process, set a single success metric (for example, time per task or the share of drafts approved without edits), run a small pilot, and measure for a few weeks. If the numbers hold up, you expand to the next process on the list. If they don't, you have cheap proof that this particular process wasn't first in line.