The First AI Project in an SME Should Be a Business Bottleneck, Not a Technology Demo.

A practical selection test for owners and operations managers deciding what to automate first.

A small company can run a convincing AI demonstration in an afternoon. That does not tell the owner whether the business should adopt it.

The test I would trust is more ordinary: can the team name a recurring process that takes too long, creates avoidable corrections, or repeatedly sends decisions back to the owner?

For a service company, that process might begin when sales confirms a customer request and end when delivery has the right scope, owner and schedule. For a software agency, it could be what happens after a requirement changes: the tasks, estimates, project plan and customer commitment all need another look.

When I managed software delivery without an AI-assisted process, a significant amount of time went into preparing and maintaining project documentation. Research and analysis were repeated when new requirements or improvements emerged, followed by adjustments to plans, tasks and estimates. These were legitimate management activities, but manual reconciliation added substantial effort.

Later, we introduced AI assistance across the product-development workflow. Preparation and revision work became less time-consuming, and estimates were more dependable in day-to-day use. That observation about estimates is qualitative; I do not have a separately calculated estimation-error reduction to report. In that operating context, the measured research-to-release cycle became at least 40% shorter, based on before-and-after sprint-planning timelines. AI was part of a changed workflow, so the result should not be attributed exclusively to the tool.

For another SME, I would not start by promising a similar number. I would use four tests to choose the first process.

1. Is there a real operating cost?

Ask where staff repeatedly enter the same information, wait for approval, reconcile conflicting records, or correct work downstream. A workflow with a visible queue or handoff and a meaningful cost is easier to assess than an abstract ambition to "be more AI-driven."

The cost need not be expressed in money immediately. A consistent delay, senior-review burden or avoidable rework can be a usable starting baseline.

2. Who owns the process and its exceptions?

Trace one recent real example from request to completion. What is the authoritative record? Who accepts incomplete input? Who approves a changed scope, quote or delivery promise? Where does the task go when the usual path fails?

If the answer is "someone normally sorts it out," the company probably needs to make that responsibility explicit before automating the step.

3. Does the step need AI at all?

Sending a complete approved form from one system to another is often a rules-and-integration task. A language-heavy activity, such as preparing a first impact assessment from several changing documents, may be a candidate for AI assistance. A client commitment, exception approval or quality acceptance still has a responsible decision owner.

The aim is to choose the cheapest reliable intervention for the actual job, including the work of checking its outputs.

4. Can we test one improvement safely?

Choose a limited pilot with known inputs and a defined review point. Record the current processing time, correction rate, quality issues and amount of specialist attention. Then compare those same measures after the change.

Count work moved to another person as part of the cost. For example, if an AI-assisted first draft saves 20 minutes but creates 25 minutes of senior correction, total effort has increased. Those figures are illustrative, not results from my project.

At the end, make an explicit decision: keep the intervention, change it, or stop. If a simpler rule or clearer ownership performs better, use that.

Many SMEs are adopting off-the-shelf AI while still finding targeted integration difficult. The OECD's 2026 D4SME Survey reports that pattern in a non-representative sample of over 2,000 SMEs across 12 OECD countries. The observation does not establish what any individual company needs, but it explains why the first operating decision deserves attention.

If your team has a specific recurring workflow but isn't sure where AI, ordinary automation and accountable human decisions belong, I offer an independent AI Workflow Diagnostic ($299) for one current process. It covers the process map, opportunities, controls, risks, and a practical first sequence. It does not include implementing an automation or a company-wide technology rollout.

Explore the AI Workflow Diagnostic

M. Mohammad · The Tech Manager’s Playbook