Artificial Intelligence: frequently asked questions

Most projects reach a production-ready state in 8 to 14 weeks. The discovery phase takes one to two weeks, the proof of concept another three to four, and the remaining time goes to iteration, integration, and testing. Projects with clean, well-structured data move faster. If your data is scattered across multiple systems with inconsistent formats, expect the upper end of that range or a separate data-engineering phase before modelling begins.

We need access to the historical data relevant to the problem. For a demand forecasting model, that means past sales transactions, stock levels, and any external signals like promotions or weather data. For a chatbot, we need your internal documentation. For computer vision, we need labelled images or video clips showing the objects or defects you want the model to recognise.

During the discovery phase we produce a data requirements document that lists exactly which tables, fields, and volumes we need. If some of the data does not exist yet, we help you set up collection processes before model training starts.

Not necessarily. Many of our clients have no in-house ML team at all. We handle the full lifecycle: data preparation, model training, deployment, and monitoring. You need someone on your side who understands the business problem and can validate whether the model outputs make sense, but that person does not need to write Python.

If you do plan to build an internal team eventually, we offer a knowledge-transfer package where we pair-programme with your new hires during the project so they can maintain the system after handover.

Yes. We have worked with clients in healthcare and financial services who cannot move data to the cloud for regulatory reasons. In those cases we set up secure VPN access to your on-premises environment and run training jobs on your hardware. We can also deploy models on local servers or edge devices rather than cloud endpoints.

It depends on scope. A focused proof of concept for a single predictive model typically costs between £12,000 and £25,000. A full production deployment with integration, monitoring, and three months of support ranges from £30,000 to £80,000. Enterprise-scale projects with multiple models and custom infrastructure can exceed that. We publish indicative pricing on our pricing page, and every engagement starts with a fixed-price discovery phase so you know the full cost before committing.

We define performance thresholds during the discovery phase. If the proof of concept does not meet those thresholds, we present an honest assessment: sometimes the fix is more data, sometimes a different modelling approach, and sometimes the problem is not well-suited to ML at all. In the last case, we will tell you that and you only pay for the discovery and PoC work completed.

We have walked away from three projects in the past two years because the data was insufficient to produce reliable predictions. We would rather lose a contract than deliver a model that misleads your team.

Yes. Every project includes 30 days of post-launch support at no extra cost. After that, we offer monthly retainer plans that cover model monitoring, retraining, bug fixes, and a set number of engineering hours for feature requests. Most clients choose the mid-tier retainer, which includes weekly monitoring reports and up to 20 hours of engineering time per month.

You do. All custom code, trained model weights, and documentation produced during the engagement are your intellectual property. We retain no licence to use your data or models after the project ends. The only exception is our internal tooling and libraries, which we licence to you for the duration of the support contract.

Yes. Most of our work is remote, and we have delivered projects for clients in Ireland, the Netherlands, and Germany. Time-zone overlap matters for collaboration, so we work best with teams in UTC-1 to UTC+3. If you are further afield, we can adjust meeting schedules, but async communication via Slack or email becomes the primary channel.