The six phases
We developed this framework over three years and more than sixty client projects. It keeps timelines predictable and gives you clear decision points along the way.
Discovery call and problem framing
We start with a 45-minute video call. You describe the business problem; we ask questions about your data, systems and team. By the end of the call we can usually tell you whether AI is the right tool or whether a simpler approach would serve you better.
Within two working days you receive a short brief that restates the problem in our language, lists the data we would need to see and proposes a rough timeline. No jargon, no 40-page slide deck.
Data audit
Before we build anything, we need to know what your data actually looks like. We sign an NDA, then connect to a sample of your data through a secure transfer or a read-only database credential.
Our data engineers check completeness, consistency and volume. They flag gaps, quality issues and any fields that could introduce bias. This audit typically takes five to eight working days, depending on the number of data sources involved.
You get a written data-readiness report. If the data is not yet good enough to train a reliable model, we tell you what needs to change before we proceed, and we can help you fix it if you want.
Pilot build
This is where things get tangible. We agree on a single, tightly scoped deliverable: for example, a demand-forecast model for one product category, or a document-classification tool for one type of contract.
The pilot runs four to six weeks. We work in weekly sprints and share progress every Friday in a short demo call. You see real outputs on real data, not synthetic examples.
At the end of the pilot, we present accuracy metrics, compare them against the success threshold we agreed at the start, and give you a clear recommendation: proceed, iterate or stop.
Production deployment
Once the pilot is approved, we move the model into your production environment. That could be an Azure ML workspace, an AWS SageMaker endpoint or a container running on your own servers.
We write the integration code that connects the model to your existing systems: your ERP, your CRM, your internal dashboards. We also build monitoring that tracks prediction accuracy and alerts us if the model starts drifting.
Deployment usually takes three to four weeks. We run the new system in shadow mode alongside your current process for at least one week before switching over, so you can compare outputs side by side.
Team training
A model is only useful if the people who rely on it understand what it does and what it cannot do. We run a half-day workshop for the team members who will interact with the system daily.
The workshop covers how to interpret the model's outputs, when to override its recommendations and how to flag issues. We leave behind a short reference guide written in plain English, not a technical manual.
Ongoing support and retraining
AI models are not "set and forget". Customer behaviour changes, product ranges shift and data distributions evolve. We offer monthly or quarterly retrain cycles as part of a support agreement.
Each retrain cycle includes a performance review, a fresh round of validation and a brief report summarising what changed. If accuracy drops below the agreed threshold, we investigate the cause and adjust the model architecture or training pipeline as needed.
Support clients also get a dedicated Slack or Teams channel with a four-hour response-time SLA during business hours.
What you can expect at each milestone
Concrete deliverables, not vague promises.
After discovery
A two-page problem brief, a list of data requirements and a go/no-go recommendation. Delivered within two working days of the call.
After the data audit
A data-readiness report covering completeness, quality scores and bias risks. If remediation is needed, a prioritised action list with estimated effort.
After the pilot
A working model, accuracy benchmarks against the agreed success metric, a cost estimate for production deployment and a candid recommendation.
After deployment
The model running in your infrastructure, integration code, monitoring dashboards and a shadow-mode comparison report.
Common questions about our process
If your question is not here, call us on +44 113 300 7115 or email [email protected].
Most projects reach production deployment within 12 to 18 weeks. A straightforward predictive-analytics project with clean data can be faster. Complex multi-source integrations or projects that require significant data remediation may take longer. The pilot phase is always fixed at four to six weeks, which gives you an early exit point if needed.
Pilot pricing depends on the scope, but most pilots fall between £8,000 and £18,000. We quote a fixed price before any work begins, so there are no surprises. If the pilot does not meet the agreed success metric, you are free to walk away with no further commitment.
No. Many of our clients have no data scientists on staff. We handle the model development, deployment and retraining. What you do need is at least one person who understands the business problem well enough to validate the model's outputs and give us feedback during the pilot.
We deploy on Azure, AWS and Google Cloud. We can also deploy on-premise if your data-governance policy requires it. About a third of our clients run models on their own servers for compliance reasons, and we have well-tested deployment scripts for Docker and Kubernetes environments.
Our monitoring system detects accuracy drift automatically and alerts both our team and yours. Under a support agreement, we investigate the root cause and retrain the model within the agreed SLA. Common causes include changes in customer behaviour, seasonal shifts or upstream data-quality issues. Most drift events are resolved within one retrain cycle.