How we deliver Artificial Intelligence projects
Every engagement follows a clear sequence of stages. We designed this process after running more than 40 AI projects since 2018, and we refine it regularly based on what we learn.
Six stages from first call to production
Below is the typical path a project takes. Timelines vary by scope, but most projects complete within eight to sixteen weeks.
Discovery call
We start with a 30-minute video call. You describe the business problem; we ask about your data, systems and team. There is no charge for this call, and there is no obligation. By the end we both know whether the problem is a good fit for AI.
If it is not, we say so. We have recommended off-the-shelf software, simple rule-based automation or even a spreadsheet formula to about 20% of the people who contact us. Honesty at this stage saves everyone time.
Data diagnostic
This is a paid two-week engagement. We sign an NDA, connect to your data sources (or receive sample exports) and assess volume, quality, coverage and bias. We interview the people who generate and use the data day to day, because they know things the data dictionary does not.
The output is a diagnostic report: a 10-to-15-page document that describes the data landscape, identifies feasible use cases, estimates accuracy ranges and lays out a project plan with fixed costs. Clients often tell us this report alone was worth the fee because it gave them a clear picture of what was realistic.
Prototype build
If you approve the plan, we build a working prototype in three to five weeks. This is not a demo with fake data. It runs on a subset of your real data and produces real predictions or extractions you can evaluate.
We use Python, PyTorch and scikit-learn for most model work, with FastAPI for serving. Your team gets access to a staging environment where they can test the prototype and give feedback. We iterate quickly: most prototypes go through two or three rounds of adjustment before we lock the approach.
Production engineering
Once the prototype proves itself, we harden it for production. That means containerised deployment (Docker, Kubernetes or your existing infrastructure), automated testing, logging, error handling and security review.
We also build a monitoring dashboard so your team can see model performance metrics in real time: prediction volume, latency, accuracy drift. If accuracy drops below an agreed threshold, the system sends an alert.
Handover and training
We do not just throw code over the wall. Every project includes a half-day training session for your technical team covering the model architecture, the retraining pipeline and the monitoring setup. We also deliver written documentation: architecture diagrams, API reference and a runbook for common issues.
If your team does not include ML engineers, we can arrange a longer knowledge-transfer programme or an ongoing support contract (see stage 6).
Ongoing support
AI models are not static. Data distributions shift, business rules change, new edge cases appear. We offer monthly support packages that include model retraining, performance reviews and priority access to our engineering team.
Support clients get a quarterly review meeting where we look at model metrics, discuss upcoming data changes and recommend improvements. About 70% of our clients choose some level of ongoing support.
Why a structured process matters
AI projects fail for predictable reasons. The most common one is skipping the data diagnostic and jumping straight to model building. When you do that, you discover data problems late, after you have already spent weeks of engineering time. Our diagnostic stage catches those issues early, when they are cheap to fix.
The second most common failure is building a model that works in a notebook but never reaches production. Our process separates prototyping from production engineering on purpose. The prototype proves the concept; the production phase makes it reliable. Mixing the two leads to fragile systems.
A third risk is model decay. A model trained on 2023 data may perform poorly on 2025 data if customer behaviour or market conditions have shifted. Our monitoring and retraining stage addresses this directly. We set measurable thresholds and automate alerts so problems surface before they affect your business.
Frequently asked questions about our process
The diagnostic is a fixed-fee engagement. For most projects it ranges from €3,500 to €6,000 depending on the number of data sources and systems involved. We quote the exact figure after the discovery call.
We strongly advise against it. The prototype stage is where we validate assumptions, test different model architectures and get feedback from your domain experts. Skipping it increases the risk of building something that does not solve the actual problem. In rare cases where a client has already validated the approach internally, we can compress the prototype phase to two weeks rather than eliminating it.
Most data is messy. Missing values, inconsistent formats, duplicates: we deal with these on every project. The diagnostic tells us how much cleaning is needed and whether the remaining data is sufficient for the use case. Sometimes we recommend a short data-engineering sprint before model work begins.
Yes. About a third of our clients are based in the UK, Germany or the Netherlands. We work remotely with them and travel for on-site workshops when needed. All our contracts comply with GDPR.
We set accuracy targets during the diagnostic phase based on what the data can realistically support. If the prototype does not meet those targets, we discuss options: additional data collection, alternative modelling approaches or, in some cases, a recommendation not to proceed. You are never locked into a build phase if the prototype results are unsatisfactory.
Ready to start?
Book a free 30-minute discovery call and we will tell you whether AI can help with your specific problem.
Get in touch