AI Case Studies for SMEs: How to Judge One and Size Your Own
An AI case study for an SME is useful only if it gives a measured starting point, a precise scope and a follow-up period. Without those three elements, a "70% gain" tells you nothing about what you will get in your own company. This article explains how to read a case study, which questions to ask before drawing inspiration from it, and how to size your own project without waiting for someone else to do it first. It supplements our complete guide to AI in business for SMEs.
One clarification: the worked scenario below is illustrative. It does not correspond to any real GrowthPerf client and only shows the calculation method. Replace every assumption with your own data.
Why AI case studies are so hard to compare
Because almost none of them publish the same kind of information. A case study is a sales document before it is a measurement. It showcases the best result, on the best process, with the best team.
Context matters too. According to Eurostat, 20.0% of EU enterprises with 10 or more employees used AI technologies in 2025, up from 13.5% in 2024. Usage is growing fast but is still a minority practice: many SMEs have no internal benchmark and treat the first figure they read as a reference. That is exactly the risk.
Three biases come up almost every time:
- selection bias: only projects that worked get published;
- scope bias: the announced gain covers one step of the process, not the whole process;
- duration bias: the result is measured a few weeks after launch, before usage drops off.
The five elements that make a case study credible
A credible case study answers five questions: which process, which starting situation, which solution, which measured result and over what period. If one is missing, ask before projecting yourself into it.
| Element | What to require | Warning sign |
|---|---|---|
| Targeted process | A named task (e.g. sorting inbound requests) | "Company-wide transformation" |
| Starting situation | Volume, time spent, error rate before the project | No "before" figure |
| Solution | Tools, level of human supervision, integrations | "AI" mentioned with no detail |
| Result | A figure, how it was measured, who measured it | A percentage with no baseline |
| Duration | At least a few months of hindsight | Result after two weeks |
Focus on the "starting situation" row. A time saving can only be calculated against an initial measurement. If the company never timed the task beforehand, the announced figure is an estimate, not a measurement.
How to spot an inflated figure
A figure is suspect when it is round, has no calculation basis and no mention of cost. A few habits help you sort quickly.
First, ask for the total cost. A project that "saves 20 hours a week" but needs two days of weekly supervision is not the same proposition. Cost includes licences, integration, team training and human review time. Our article on the cost of AI for an SME in 2026 details these items.
Second, look for the share of work that stays human. In most SME projects, AI prepares and a person validates. A case claiming 100% automation on a task involving customers or personal data deserves questions, especially about data confidentiality.
Third, check how the gain is defined. Hours saved are not euros saved. If those hours are not reassigned to productive work, cash flow does not change. This is the starting point of the AI ROI calculation for SMEs.
An illustrative scenario to understand the calculation
Here is a worked example, not a real case: a service SME receives 120 email requests a week and wants to speed up triage. All values below are assumptions to be replaced.
- Starting situation: 6 minutes of handling per request (reading, qualifying, standard reply), or 12 hours a week.
- Solution tested: a workflow that classifies the request, extracts key information and prepares a draft reply. A person reviews and sends it.
- Time after the project: 2 minutes per request, or 4 hours a week.
- Gross gain: 8 hours a week, roughly 368 hours over 46 working weeks.
- Valuation: at a loaded hourly cost of 35 EUR (assumption), that is about 12,880 EUR a year.
- Year 1 cost: 4,500 EUR for setup and 80 EUR a month for licences (assumptions), or 5,460 EUR.
- Year 1 result: about 7,420 EUR net gain, with payback reached in five to six months.
This calculation is deliberately simple. It ignores the learning period, errors to correct and the time spent improving the setup. Its purpose is to show where the variables sit: volume, unit time, hourly cost, project cost. To manage them afterwards, rely on suitable KPIs.
How to size your own case before launching
Time the task for two weeks before choosing a tool. It is the only way to get a starting situation nobody can challenge, including you.
- Pick a repetitive process with a measurable volume (requests, invoices, quotes, meeting notes).
- Record the real time per occurrence on ten to twenty cases, interruptions included.
- Identify what must stay under human control and what can be prepared automatically.
- Set a quality indicator (share of drafts accepted unchanged, errors caught) alongside time saved.
- Define a test period and a decision threshold before starting, for example "we scale up if average time drops under 3 minutes with fewer than 5% errors".
If you do not know where to start, an AI audit for SMEs ranks processes by potential. To structure the project over time, the 90-day roadmap provides a realistic frame.
Mistakes that most often distort a case study
The most common mistakes come from management, not technology. People measure too early, compare different scopes or forget supervision time.
A few concrete examples:
- launching a pilot without a baseline, then trying to reconstruct the figures afterwards;
- measuring the gain on the first users, often the most motivated, and extrapolating to the whole team;
- ignoring undeclared parallel usage, described in the article on shadow AI;
- underestimating support, whereas change management explains a large share of the gaps between projects;
- not tracking usage over time, which the article on measuring adoption covers.
For an overview, also see AI deployment mistakes in SMEs and the adoption steps.
Build or buy: what external support changes
External support mainly reduces scoping time and the risk of choosing the wrong tool, not the need to measure. Whether you build in-house or with a provider, the starting situation remains your responsibility. The choice depends on the skills you have available, as shown in the article on build or buy for AI projects.
Either way, you will need to convince leadership. The method is detailed in pitching an AI project to leadership, and it relies on exactly the type of sizing shown above. To gauge plausible gains, the AI productivity figures for SMEs serve as a reference, to be handled with the same caution as any case study.
Frequently asked questions
Is a published AI case study reliable?
It is reliable to the extent that it shares its method. Look for the starting situation, scope, duration and cost. A case that only gives a final percentage remains a sales indication.
What gains can you reasonably expect from a first project?
It depends on the process and the volume. On a repetitive triage or drafting task, noticeable time savings are possible, but they vary widely between companies. Measure in your own context before aiming at a reference figure.
How long before you see a result?
Plan a test of a few weeks to validate feasibility, then several months to confirm that usage holds. A result measured at two weeks is a signal, not proof.
Do you need a specific tool to automate a process with AI?
Not necessarily. Many SME projects rely on process automation tools combined with a language model, with human validation. The choice depends on your existing tools and data sensitivity.
How do you present an internal case to leadership?
Present the quantified starting situation, the gain hypothesis, the total cost, the test duration and the decision threshold. Also state what remains under human control. That transparency reassures more than a large percentage.
Is a fast return on investment possible in an SME?
It is possible on high-volume, low-complexity processes, but you must demonstrate it with your own figures. AI ROI is calculated on the full cost, supervision included.
How GrowthPerf supports you
GrowthPerf is a Qualiopi-certified training organisation specialising in AI, no-code and automation for SMEs and associations. We start from your real processes: we measure the starting situation with you, then choose the tools and the level of supervision that fit. Your teams can then build skills with the operational AI or no-code and automation training.
To keep an overview of your approach, see our complete guide to AI in business for SMEs.
Free AI training audit (30 minutes): book a call with GrowthPerf to identify the first process to measure in your company.