How our Artificial Intelligence projects run

Every engagement follows the same five-phase structure. You know what happens next, what it costs, and when it will be done.

The five phases

We developed this process over 43 completed projects. It keeps timelines honest and prevents scope from ballooning. Each phase has a defined output, and you sign off before we move on.

Discovery sprint

Two weeks. We audit your data sources, interview the people who will use the system, and test whether the data quality supports the AI approach you have in mind. The output is a written feasibility report with a recommended architecture, estimated accuracy ranges, and a fixed-price quote for the build phase. If the data does not support a viable model, we tell you and the engagement ends here. You pay only the sprint fee (typically £3,500 to £5,500 depending on complexity).

Data preparation

This is where most AI projects actually succeed or fail. We clean, label and structure your data. For document automation, that means building extraction templates and testing them against a sample of 200+ real documents. For predictive models, we engineer features, handle missing values, and create training/validation splits. Duration: two to four weeks. You receive a data quality dashboard so you can see exactly what went in.

Model development

We train, evaluate and iterate. For most projects this means testing three to five model architectures and selecting the one that best balances accuracy, speed and explainability. We do not default to the most complex option. A gradient-boosted tree that runs in 50 milliseconds and explains its reasoning is often more useful than a deep network that takes two seconds and cannot. Duration: three to six weeks depending on scope.

Integration and deployment

The model needs to live inside your existing workflow. We deploy via REST API, embed it in your internal tools, or connect it directly to your database triggers. We write monitoring dashboards that track prediction accuracy, latency and data drift in real time. You get alerts before performance degrades, not after. This phase usually runs two to three weeks and includes load testing at 2x your expected peak volume.

Handover and optional maintenance

We document everything: model architecture, training data lineage, deployment configuration, and a runbook for common issues. Your team gets a half-day training session. If you want us to handle ongoing retraining and monitoring, we offer maintenance contracts billed quarterly. About 60% of our clients choose this option. Retraining frequency depends on how fast your data changes; monthly is typical for sales forecasting, quarterly for document classifiers.

Project planning on a whiteboard during a discovery sprint

Why a fixed-price discovery sprint?

Plenty of consultancies offer free initial assessments. The problem is that a free assessment has no teeth. Nobody assigns their best analyst to unpaid work, and the output is usually a slide deck designed to sell the next phase rather than to give you an honest answer.

Our discovery sprint is a real project. We assign a senior engineer for two weeks. They look at your actual data, not a sample you selected to look good. The report includes specific accuracy estimates with confidence intervals, a list of data gaps that would need to be filled, and a realistic timeline.

If the conclusion is that AI will not deliver enough value to justify the cost, you walk away with a clear explanation of why, and you have only spent the sprint fee. About one in eight discovery sprints ends with a "not yet" recommendation, usually because the training data is too thin. In those cases we outline what data you would need to collect and how long before it would be worth revisiting.

What you get at each checkpoint

We do not send status emails that say "things are going well." Each phase ends with a concrete deliverable you can evaluate.

After discovery

A written feasibility report (8–15 pages), including data quality scores, recommended model type, estimated accuracy range, infrastructure requirements, and a fixed-price quote for phases 2 through 4.

After data preparation

A clean, versioned dataset stored in your own infrastructure. A data quality dashboard showing completeness, consistency and distribution metrics. Documentation of all transformations applied.

After model development

A trained model with evaluation metrics (precision, recall, F1, or RMSE/MAE depending on task type). A comparison table showing all architectures tested. An explainability report showing which features drive predictions.

After deployment

A live API endpoint or integrated module, a monitoring dashboard, alert configuration, load test results, and a complete runbook. Your team gets a hands-on training session before we step back.