How we deliver AI software, step by step
Every engagement follows the same six-phase structure. The phases overlap, timelines flex, but the order stays consistent because it works.
The six phases
Most projects move through all six phases in 8 to 14 weeks. Shorter pilots that only test feasibility can wrap up in two.
Scoping call
A 45-minute video call where you describe the business problem and we ask pointed questions about your data, systems, and team. We want to understand what a successful outcome looks like in concrete terms: fewer manual hours, lower error rates, faster turnaround. By the end, both sides know whether this is worth pursuing. If it is, we send a one-page scope document within 48 hours. No cost for this step.
Data audit
Before writing any model code, we need to see your data. We sign an NDA, receive a sample dataset (or connect to your staging database), and spend three to five days checking volume, quality, labelling, and gaps. The output is a short report that tells you exactly what is usable, what needs cleaning, and whether the data supports the goal from phase one. If the data is insufficient, we say so and suggest what to collect before trying again.
Paid pilot
This is a two-week fixed-price engagement. We build a minimal version of the model or pipeline, test it on a slice of real data, and present the results. The pilot answers one question: does this approach meet the accuracy or speed threshold you need? If the answer is no, you walk away having spent a fraction of a full project budget. If yes, we proceed with a clear baseline to improve on.
Full build
The main development phase runs four to ten weeks depending on complexity. We work in one-week sprints. Each Friday you receive a short written update and a demo of whatever was completed that week. You can test intermediate versions in a staging environment. Our engineers handle data pipelines, model training, API development, and front-end interfaces if needed. We use your preferred cloud provider or host on ours.
Integration and go-live
We connect the finished system to your production environment. That might mean writing an API adapter for your ERP, setting up webhook triggers in your CRM, or building a simple internal dashboard. We run the system in shadow mode (processing real data but not acting on it) for at least one week to catch edge cases. Go-live happens only after you sign off on the shadow-mode results.
Monitoring and support
Every project includes 90 days of post-launch support. During this period we monitor model accuracy, retrain if performance drifts, and fix bugs at no extra charge. After the 90 days, you can continue with a monthly support plan or take full ownership. We hand over all code, documentation, and trained model artefacts regardless of which path you choose.
Why this order matters
We learned the hard way that skipping the data audit leads to wasted build time. One early client asked us to build a demand forecasting model for their retail chain. We jumped straight into development, spent three weeks training a model, and then discovered that 40% of the historical sales records had been entered in the wrong currency. The model was useless.
Since then, phase two has been non-negotiable. It adds a few days up front but saves weeks of rework later. The paid pilot serves a similar purpose: it forces both sides to confront reality early. If the data cannot support the desired accuracy, it is far better to find out in week three than in week twelve.
Shadow mode before go-live is another lesson from experience. Production data is messier than staging data. Formats change, edge cases appear, upstream systems behave differently under load. A week of shadow running catches most of these surprises before they affect your users or customers.
Questions we hear often
Paid pilots start at €4,500. Full builds range from €15,000 to €65,000 depending on data complexity, the number of integrations, and whether a custom front-end is required. We quote a fixed price after the data audit, so there are no surprises.
It depends on the task. For classification and extraction tasks, labelled examples dramatically improve accuracy. If you have no labels, we can set up a lightweight annotation workflow during the pilot phase using tools like Label Studio. We have also used semi-supervised techniques when labelled data is scarce but unlabelled data is plentiful.
Yes. About a third of our clients require on-premise deployment for regulatory or data-residency reasons. We package models in Docker containers and provide Helm charts for Kubernetes clusters. If your infrastructure is simpler, a plain Docker Compose setup works too.
You receive the pilot report, all code written during the pilot, and a clear explanation of why the approach did not meet the target. Sometimes the fix is straightforward: more data, different features, a revised target metric. Other times the problem genuinely is not solvable with the data available. Either way, you pay only for the pilot and nothing more.
We sign a data processing agreement before any data transfer. All data in transit uses TLS 1.3. Data at rest is encrypted with AES-256. We can work within your VPN if needed and delete all copies of your data from our systems within 14 days of project completion, with written confirmation.
Ready to start with a scoping call?
Describe your problem in a few sentences and we will schedule a 45-minute call within two working days. No charge, no obligation.
Get in touch