Technology · AI at work
AI Adoption in SMEs: From Pilot Project to Everyday Practice
Adopting AI in SMEs: the three-step approach of pilot area, tool plus workflow, and time tracking – with a worked example and the most common mistakes.
By Boaz Lichtenstein
For SMEs, whether to adopt AI is no longer the question – but a lot of money gets burned between the two most common answers. Some wait until “the technology is mature” (for many tasks, it already is), others licence tools for everyone and call it a strategy. The profitable path is narrower and far less spectacular – it doesn’t start with a tool subscription, but with a single, well-chosen question.
Key takeaways
- The three-step approach – pilot area → tool plus workflow → measure and scale – beats any company-wide rollout hands down.
- A single process with demonstrable time-sink potential is a better starting point than a blanket tool licence.
- Data rules, sceptics taken seriously, and a mandated owner determine whether adoption sticks – not the choice of software.
- After eight to twelve weeks, simple time tracking shows honestly what pays off; only then does scaling follow.
- The costliest single mistake: buying the tool before the process, and mistaking broad licensing for strategy.
Where to start: finding the right pilot question
The right starting point is a single area with a measurable time-sink problem – not the whole organisation at once. Typical candidates are quote preparation, customer service pre-qualification, reporting, or processing incoming documents. What matters isn’t the department, but that today’s effort can be quantified in hours.
Three criteria filter out good candidates: volume is high enough that time savings add up over weeks; the task is repetitive enough for patterns to emerge, but not so sensitive that a single mistake becomes costly; and there’s someone on the ground who actively backs the pilot rather than merely tolerating it. A sales team that manually puts together several quotes a day is a better pilot than the entire executive communications function – not because the latter matters less, but because mistakes there weigh more heavily and the learning effect is smaller.
From experience: If you’re unsure where to start, don’t ask senior management first – ask the teams directly. “Where do you spend the most time on dull repetition?” delivers the right answer more reliably than any top-down analysis – and builds buy-in within the pilot team along the way.
Step by step: from pilot to standard practice
AI adoption succeeds through seven traceable steps – from choosing the area to consolidation. Skip a stage – usually setting up standard cases or time tracking – and you lose exactly the trust you’d need for the next stage of expansion.
- Choose a pilot area and agree it with the affected team, rather than imposing it top-down.
- Name two or three concrete tasks that demonstrably cost hours today.
- Choose a tool that fits the task – not the one with the biggest marketing budget.
- Set up standard cases: templates, connected data, documented example solutions – half a day of joint setup beats any generic training session.
- Trial it in production for eight to twelve weeks and document time per case, before and after.
- Spot-check quality and sharpen the rules, rather than just tallying time saved at the end.
- Declare what works the new standard, drop the rest, and appoint an owner with a clear mandate.
It’s exactly this consolidation – deliberately drawing a line between what stays and what’s phased out – that distinguishes genuine adoption from a never-ending experiment.
The conditions that get forgotten
Three unglamorous factors determine whether AI adoption sticks more than any choice of tool: clear data rules, resistance taken seriously, and a named owner with a mandate. Miss any one of them, and even a successful pilot fragments back into isolated initiatives within a few months.
Data rules before the first prompt: Which data may go into which tools? A one-page, plain-language policy (allowed / allowed with anonymisation / off-limits) prevents both recklessness and paralysis – and, as a side effect, covers the training obligation under the EU AI Act.
Take the sceptics seriously: Resistance usually stems from real concerns – job anxiety, pride in quality, bad first experiences. What works is anything that restores a sense of control: time to learn, a say in which cases get picked, and the visible rule that AI takes on tasks, not responsibility. Our article on AI adoption in teams covers how to build acceptance in detail.
An owner with a mandate: Without a named person who gathers experience, maintains standards and evaluates new options, every rollout fragments into isolated solutions – each department trials its own tool, and none of it ever gets consolidated.
Worked example: what a pilot project actually delivers
A rough worked example shows the order of magnitude: a five-person sales team that prepares quotes daily can save several hours of work per day through AI-assisted first drafts – with tool costs that typically pay for themselves within the first month.
Assume five employees each prepare three quotes a day – 15 quotes daily. Manually, a quote takes 30 minutes on average – 7.5 hours of total effort per day. With an AI-assisted first draft (the model pulls product data and the template, a human checks and adjusts), effort drops to around 12 minutes per quote – 3 hours a day. Saving: 4.5 hours daily, which works out to nearly 1,000 hours a year across roughly 220 working days. At an internally calculated hourly rate of €45, that’s worth roughly €44,000 a year – against tool costs of perhaps €2,000 a year for the five licences. Exact figures vary widely by industry and starting point; the order of magnitude – several times the tool cost – is typical for well-chosen pilots.
The most common mistakes in AI adoption
Most failed AI adoptions don’t fail because of the technology – they fail because of a handful of recurring organisational mistakes, all avoidable once you know them.
Buying the tool before the process: Rolling out licences to all staff before a concrete use case exists. Fix: identify the process with the greatest time-sink potential first, then procure tools specifically for it.
No success measurement: “It feels like it saves time” isn’t enough to justify scaling. Fix: document time per case, before and after, from day one.
Dismissing sceptics: Resistance gets written off as laziness instead of taken seriously as a signal. Fix: give people a say in which cases get picked, and address concerns concretely.
No owner appointed: Every department trials its own tool, nothing gets consolidated. Fix: one person with a clear mandate for standards and evaluation.
Scaling too fast, too broadly: A single successful case gets rolled out company-wide immediately, without consolidation. Fix: cement two or three proven cases first, then expand.
Choosing a pilot area: which area, when?
Not every business area is equally suited as a first pilot – the table shows how typical time sinks differ by risk and suitability.
| Area | Typical time sink | Risk if it goes wrong | Suitability as first pilot |
|---|---|---|---|
| Quote preparation | High | Low–medium | Very good |
| Reporting | Medium | Low | Very good |
| Customer service pre-qualification | Medium–high | Medium | Good |
| Incoming documents/invoices | High | Medium–high | Good, with oversight |
| Contract review | Low–medium | High | Only after experience |
Rule of thumb: the higher the volume and the lower the risk of a mistake, the better suited an area is for the first pilot. Highly sensitive areas such as contract review belong in the second wave, once the company already has experience with AI-assisted quality control.
The bottom line
SMEs rarely win with AI through anything spectacular – they win through the sum of unspectacular hours: faster quotes, lighter caseloads, better first drafts. Anyone who, after a year, can point to two or three firmly established use cases with documented time savings has handled adoption better than most. The next step is then no longer a leap but a staircase: once you’ve run the three-step approach cleanly, you can apply it to further areas – all the way to agents handling entire process chains. But the first step is always the same: one question, one area, one round of time tracking.