AI Strategy in 90 Days: From Baseline to First Results

Key Takeaways: A working AI strategy does not need to take six months. In 90 days you can go from a baseline assessment to a validated proof of concept, provided you follow the right phasing and avoid common mistakes. This article describes an approach in four phases: baseline assessment (weeks 1-2), use case selection (weeks 3-4), proof of concept (weeks 5-8), and validation with scaling decision (weeks 9-12).

Why 90 days, not six months

An AI strategy that takes six months is, in 80% of cases, a strategy that never gets implemented. Lengthy strategy projects lose momentum, budget holders grow impatient, and the market moves on while you are still planning. Research by McKinsey (2025) shows that companies reaching first results within 90 days are 3.2 times more likely to scale successfully than companies with longer timelines.

The 90-day approach works because it forces focus. You cannot do everything at once, and that is precisely the point. Instead of an all-encompassing transformation plan, in 90 days you produce concrete evidence that AI adds value to your specific business, with your specific data, for your specific processes. That evidence is worth more than any strategy document.

A common objection is that 90 days is too short for "real" AI. That is true if you expect a fully automated system. But the goal of the first 90 days is not perfection; it is validation. You want to know whether it works, how much it returns, and what is needed to scale. You can get those answers in 90 days, and they are exactly the input you need for a substantiated investment decision.

Weeks 1-2: the baseline assessment

The first two weeks are dedicated to understanding your current situation without touching any technology. A solid baseline is the foundation on which all subsequent steps rest and prevents you from building a solution for a problem that does not exist.

The data audit is the starting point and takes 3-5 working days. You inventory what data your organization generates, where that data is stored, in what format, how current it is, and who has access. In practice, most SMEs discover during this audit that they have more data than they thought, but that it is scattered and inconsistent. A typical company with 50 employees has 15-25 data sources, of which only 4-6 are structured and reliable enough for direct AI application.

Parallel to the data audit, you conduct a process mapping exercise. You map the 5-10 most time-intensive or error-prone processes with concrete figures: how many hours per week does it cost, how many errors occur, what is the direct financial impact? This quantification is essential because you will later need to prioritize use cases based on expected ROI.

Finally, during this phase you speak with 5-8 key figures in the organization: the people who work with the processes daily. They know where the pain is, what workarounds they have devised, and where information gets lost. These interviews take 45-60 minutes each and yield insights that cannot be found in any dashboard.

The result of weeks 1-2 is a baseline report with three elements: a data landscape inventory, a process analysis with quantitative pain points, and an initial list of 8-12 potential AI applications. This report does not need to be 50 pages; 5-10 pages with concrete figures is more effective.

Weeks 3-4: selecting the top 3 use cases

With the baseline in hand, it is time to choose. The temptation is to select the most ambitious application, but the best first use case is rarely the most spectacular. You are looking for the application that combines the highest probability of success with impact in euros and acceptable complexity.

The ROI scoring matrix is the instrument that makes this choice objective. You score each potential application on five criteria, each on a scale of 1-5. First criterion: expected annual savings or revenue increase in euros. Second criterion: data quality and availability, because without good data every AI project is doomed. Third criterion: technical complexity, where lower is better for a first project. Fourth criterion: organizational impact, meaning how many people need to change their way of working. Fifth criterion: strategic fit with your business goals over the medium term.

The weighting of these criteria depends on your situation. A company that needs quick proof for the management team opts for high weighting on criteria 1 and 3: quick wins with low complexity. A company that wants to build broad support weighs criterion 4 higher: an application that generates little resistance.

In practice, document automation and demand forecasting most often score highest for a first AI project. Document automation scores high on data quality (documents already exist) and low organizational impact (few people need to change their workflow). Demand forecasting scores high on financial impact (direct savings on inventory costs) but requires better data hygiene.

Select your top 3 and develop a one-pager for each: the problem, the expected solution, the required data, a rough cost estimate, and the expected return. Present these three options to your decision-makers. In 70% of cases, the choice is unanimous; in the remaining cases, a brief discussion helps determine the priority.

Weeks 5-8: building the proof of concept

The four weeks for the proof of concept are the core of the 90-day journey. In this phase, you go from paper to a working prototype that demonstrates value with real business data. The discipline to do this in four weeks is essential; any proof of concept that takes longer than six weeks loses its function as rapid validation.

Week 5 focuses on data preparation. You extract, transform, and load the required data into a working environment. This is the phase where projects typically cost 30-50% more time than planned, so consciously build in buffer. A pragmatic rule of thumb: if after five working days you still do not have a workable dataset, data quality is probably insufficient for this specific use case and you should consider taking your second choice from the top 3.

Weeks 6-7 are the actual model development. Depending on the chosen application, this may mean fine-tuning an existing language model, training a forecasting model on historical data, or building a classification model for document recognition. The choice between building yourself and using a platform has already been made in weeks 3-4. Budget for this phase: 5,000-15,000 euros in external development costs, or 80-120 internal hours if you have the expertise in-house.

Week 8 is dedicated to building a minimal user interface and testing with 3-5 end users. The proof of concept does not need to look good; it needs to work. A spreadsheet with model output that an employee manually compares to reality is sufficient. The question you want to answer at the end of week 8 is: does this model deliver a better outcome than what we currently do?

The costs for the proof of concept typically range between 8,000 and 20,000 euros, depending on complexity and whether you hire externally. This is deliberately a fraction of a full implementation. The goal is to validate, not to make production-ready.

Weeks 9-12: validation and scaling decision

The final four weeks are as important as building the proof of concept but are most often skipped in practice. Validation is the process of systematically proving that the proof of concept delivers the promised value under realistic conditions.

Weeks 9-10 focus on quantitative validation. You run the model in parallel with the existing process for two weeks and measure the results. What percentage of predictions are correct? How much time does it save per employee per day? What is the error rate compared to the manual process? These figures are the hard evidence you need for the investment decision.

For demand forecasting, you measure the Mean Absolute Percentage Error (MAPE) of the model versus the current method. An improvement of 15-30% compared to the current estimation is a realistic goal for a first iteration. For document automation, you measure the percentage of correctly processed documents; 85-95% is a realistic first score, depending on document complexity.

Week 11 is the financial substantiation. You translate the validation results into a business case: if the model yields X% improvement on an annual basis and a full implementation costs Y euros, what is the expected ROI and payback period? Include the hidden costs that many projects forget: data maintenance, model retraining, change management, and compliance.

Week 12 is the scaling decision. You present three scenarios: stop (the results do not justify further investment), develop to production (the results are convincing), or pivot (the approach works but for a different application). In practice, 60-70% of well-executed proofs of concept lead to a positive scaling decision.

The role of subsidies in de-risking your journey

Subsidies are not just a financial benefit; they function as a risk reducer for the entire journey. By deploying subsidies strategically, you can finance the 90-day project without the full amount coming from your own resources.

The WBSO (R&D Tax Credit) is directly applicable to the 90-day journey. The hours your employees spend developing and testing the proof of concept qualify as research and development work. With a team of two employees each spending 50% of their time on the project, WBSO yields approximately 8,000-15,000 euros over the quarter. Important: the WBSO application must be submitted before the work begins, so this needs to be arranged in week 1.

The MIT scheme (SME Innovation Stimulation) offers a feasibility study subsidy of 35% up to a maximum of 20,000 euros. The baseline assessment and use case selection (weeks 1-4) qualify as feasibility research. An external advisor helping with the data audit and ROI scoring matrix typically costs 5,000-10,000 euros; with MIT subsidy, you pay 3,250-6,500 euros of that yourself.

For the longer term, the SLIM scheme (Learning and Development Stimulation for SMEs) is relevant. If you decide to scale up after the 90-day project, you need employees who can work with AI tools. SLIM reimburses up to 60% of costs for a training plan, up to a maximum of 24,999 euros for individual SMEs. An AI training plan for 5-10 employees typically costs 8,000-15,000 euros; with SLIM you pay 3,200-6,000 euros.

In summary: the 90-day project costs 15,000-35,000 euros all-in. With a combination of WBSO and MIT, you can realize 10,000-25,000 euros in subsidies, bringing the net investment down to 5,000-15,000 euros. That is the equivalent of one month's salary for a piece of evidence that can save months of discussion about "whether AI is something for us."

Pitfall 1: boiling the ocean

The most common mistake in AI strategy projects is trying to do too much at once. An SME that simultaneously wants to implement a chatbot, demand forecasting, and process automation ends up with three half-baked projects instead of one successful one.

The "boil the ocean" mentality stems from understandable enthusiasm. After the baseline assessment, you see opportunities everywhere, and every stakeholder has their own favorite application. The temptation to please everyone by starting multiple projects in parallel is strong, but the consequences are predictable: fragmented attention, competition for the same data engineers, and no single project that progresses enough to show real results.

The solution is disciplinarily simple but politically challenging: choose one project, make it successful, and use the success as leverage for the next project. An SME that delivers a working proof of concept in 90 days builds more support for AI than a company that after six months has three slide decks but no working system.

In practice, this means you allocate 80% of your AI budget and capacity to the first project and at most 20% to preparing the second project. The second use case from your top 3 only begins when the first is in production or when you consciously decide to pivot.

Pitfall 2: tool-first thinking

The second major pitfall is starting with the technology rather than the problem. "We need to do something with ChatGPT" or "I want an AI dashboard" are not strategy. They are solutions in search of a problem, and they almost always lead to disappointing results.

Tool-first thinking results in projects that are technically impressive but commercially irrelevant. A sophisticated deep learning model that predicts 2% more accurately than a simple regression model but costs three times as much to maintain is not a good investment. The right question is not "which AI tool should we use?" but "which business problem costs us the most money, and can AI solve it better than our current approach?"

The 90-day framework partially protects against this because the first two weeks are entirely process-oriented. You start with pain, not technology. But the temptation arises again in weeks 5-8 when the technology choice must be made. Hold to the principle that the simplest solution that works is the best one. A rule-based system that correctly handles 85% of cases is often more valuable than a neural network that scores 92% but fails unpredictably on the remaining 8%.

A practical aid is the "could an intern do this?" test. If an intelligent person with a spreadsheet and sufficient time could achieve the same result, AI is probably overkill. AI adds value when it requires pattern recognition in large datasets, when speed is essential, or when human inconsistency is a problem.

Pitfall 3: never closing the proof of concept

The third pitfall is subtler and therefore more dangerous: the proof of concept that continues indefinitely without a definitive go/no-go decision being made. This happens when results are "promising but not convincing" and the team wants more time to improve the model.

The risk is that the proof of concept imperceptibly transforms into a production system without the necessary architecture, monitoring, and support. Employees begin relying on the model, decisions are based on it, but no one has allocated budget for maintenance or defined a fallback procedure for when the model makes errors.

The solution is maintaining a hard deadline at week 12. The proof of concept is assessed at that point against pre-agreed criteria, and there are only three outcomes: production development approved with budget X, project stopped with documented learnings, or pivot to alternative use case with a new 90-day cycle. "Let's just keep going a bit longer" is not a valid option.

Your first week: concrete starting steps

This article is worthless if it does not lead to action. Therefore, we end with five concrete steps you can take this week to start the 90-day journey.

Step 1: identify your internal AI champion. This is the person who drives the project, not as a side task but as primary responsibility for the next three months. In companies up to 50 employees, this is often the owner-director or operations manager; in larger companies, a data analyst or IT manager. Budget at least 2 days per week of this person's time.

Step 2: schedule the data audit. Book 3-5 working days in the calendars of the people who manage your core systems: the ERP administrator, the CRM owner, the financial employee. Ask each the same five questions: what data does your system generate, in what format, how far back does the history go, who has access, and what are the known quality issues?

Step 3: submit a WBSO application. This must be done before work begins, so do it now. The application takes 4-8 hours to prepare and can yield 8,000-15,000 euros. The risk of rejection is limited; approval rates for WBSO are above 90%.

Step 4: define your success criteria. Write down on paper what you want to know in 90 days. Not "whether AI works" but specifically: "can AI reduce our inventory costs by at least 10%?" or "can AI correctly process 80% of our invoices without human intervention?" Measurable criteria prevent scope creep and make the go/no-go decision in week 12 objective.

Step 5: communicate the project internally. Inform your team that an AI pilot is taking place over the next three months, that their input is needed, and that the goal is to learn, not to replace jobs. That last message is crucial: fear of job loss is the biggest saboteur of AI adoption in SMEs. Emphasize that AI will help reduce tedious work, not make their roles redundant.

Conclusion: strategy is action

An AI strategy in 90 days is not half a strategy; it is a complete cycle of analysis, design, build, and validation. The difference from longer timelines is not that you do less, but that you talk less and do more. After 90 days you do not have a strategy document but a result: hard data on what AI can mean for your specific business, what it costs, and what the next step is. That information is the foundation for every subsequent decision, whether that is scaling up, pivoting, or consciously choosing not to pursue AI.