The AI Strategy That Wins in SMEs and Startups

The enterprise AI playbook — a platform team, a data lake, a centre of excellence, a two-year roadmap — is not a smaller version of what an SME should do. It is a different strategy, built for different constraints. Copying it at 60 employees produces the costs without any of the benefits.

Your actual advantages

Strategy starts from what you have that larger competitors do not:

• You can change a process in an afternoon. No change board, no committee, no six-week impact assessment. This is the single largest structural advantage available to you, and most firms never use it.
• The decision-maker is in the room. The person who owns the workflow, the budget and the outcome is often the same person. Enterprise programmes spend most of their time reconciling those three.
• You know your customers individually. You can evaluate whether an output is good by asking someone, today.
• You have no legacy platform to protect. No prior investment demands defending, no internal team's existence depends on the old approach.

Your disadvantages are equally clear: little historical data, no specialist staff, no tolerance for a failed six-figure programme. A winning strategy leans entirely on the first list and refuses to fight on the second.

The five principles

1. Buy the capability, own the workflow

Do not build models. Do not fine-tune. Use commercial models through an API and spend your effort on the part that is genuinely yours: the specific sequence of steps your business runs, the data it produces, and the judgement encoded in your staff's heads. Models are becoming a commodity; your workflow is not.

2. Automate the spine before the flourishes

The spine is the flow that turns effort into cash — enquiry to quote to delivery to invoice to receipt. Improvements here compound because every transaction crosses them. Marketing-content generation is attractive and visible, but it does not touch the spine, and it is where most small firms spend their first AI budget.

3. One workflow at a time, all the way through

Half-finished automation across five workflows produces zero benefit and five new maintenance burdens. One workflow taken from manual to measured to automated to improved produces a result you can point at — and, more importantly, an internal template for the next.

4. Keep the human where the consequences are

Automate drafting, extraction, classification, routing, monitoring and assembly. Keep humans on pricing decisions, customer relationships, quality sign-off and anything with legal or safety consequences. This boundary is not timidity; it is what makes the automated portion trustworthy enough to be left alone.

5. Measure in cash and hours, never in adoption

"Eighty percent of staff have used the tool" is not a result. Days off the cash cycle, hours returned to the owner, quotes issued per week, error rate on invoices — these are results. If you cannot express the benefit in cash or hours, you have not found the benefit yet.

The 90-day plan

Days 1–14: Find the constraint

• List every recurring task consuming more than four hours a week across the business.
• For each: frequency, who does it, minutes per run, and what breaks downstream when it is late.
• Identify the one workflow where a week of delay costs the most. That is your constraint — and it is frequently not the workflow that feels most annoying.

Days 15–30: Digitise its inputs

• Every event in that workflow gets captured digitally, same day, by the person who caused it.
• One identifier per customer and per order, carried through every artefact.
• Record the baseline: cycle time, volume, error rate, hours consumed. Write it down where others can see it.

Days 31–60: Automate the deterministic steps

• Notifications, reminders, document generation, scheduled reports, threshold alerts.
• Rules, not models. Cheap, reliable, auditable, and they establish that the programme delivers.
• Re-measure. You should already see movement; if not, stop and find out why before adding intelligence to it.

Days 61–90: Apply AI to the residue

• Whatever still needs reading, drafting, classifying or predicting.
• Shadow-run for two weeks against the human before anyone depends on it.
• Set a confidence threshold from your own shadow-run data. Above it, straight through; below it, human queue.
• Publish the 90-day delta against the day-30 baseline — including what did not work.

Budget shape for a 20–150 person firm

• Model and tool spend: typically the smallest line. Usage-based API costs for a single automated workflow usually land in the low hundreds of dollars monthly.
• Integration and build: the largest line, and it is mostly plumbing — forms, exports, scheduling, notification delivery.
• Change and training: consistently underestimated. Budget at least a quarter of build cost. This is where programmes actually fail.
• Run and maintenance: assume 15–20% of build cost annually. Automations are not furniture; they need an owner and occasional repair.

If a proposal inverts this shape — heavy on models, light on change management — it was written for a different kind of company.

Five mistakes that end programmes

1. Starting with the most exciting use case instead of the most constrained one.
2. Buying a platform before running a workflow. The platform will encode assumptions you have not yet earned the right to make.
3. Hiring a data scientist at 40 employees. You need an operations person who can automate, not a modeller. The work is 80% plumbing.
4. No baseline. Without a before number, every result becomes a matter of opinion, and opinion loses at budget time.
5. Treating it as a project. Projects end. Operating capabilities need an owner, a budget line and a review cadence.

What winning looks like at 12 months

• One core workflow measurably faster, with the numbers published monthly.
• Two or three deterministic automations running unattended, owned by name.
• One AI-assisted step live with a known straight-through rate and a functioning exception queue.
• A dataset that did not exist a year ago, now good enough to answer questions that used to require a guess.
• Staff proposing the next candidates without being asked.

That is not a dramatic transformation, and it will not make a conference keynote. It is, however, a compounding advantage — and it is achievable by a firm of thirty people with no specialist staff and a budget that would not fund a single enterprise discovery phase.