Frequently asked questions about Artificial Intelligence

Most projects fall between £18,000 and £55,000. The range depends on data complexity, the number of integrations, and whether you need ongoing retraining. We quote a fixed price after the discovery workshop so you know the total before any engineering work begins. If the scope changes mid-project (it sometimes does), we agree a change order in writing before doing extra work.

It depends on the task. For tabular prediction models we typically need at least 5,000 labelled records to produce reliable results. Fewer records can work for simpler problems with clean data, but accuracy tends to suffer below that threshold. For document automation, 200 to 500 annotated documents are usually enough to fine-tune a transformer model. We will tell you during the discovery workshop whether your data volume is sufficient.

On your own infrastructure or a UK-based cloud provider you choose. We never move data to our own servers without explicit written consent. During development, our engineers connect to your environment via a VPN or SSH tunnel. If you prefer, we can work entirely within an air-gapped network.

All production models include a monitoring dashboard that tracks prediction accuracy and input-data drift. When metrics fall below a threshold you define, the system sends an alert. Under our support package, we investigate within two business days and retrain the model if needed. Retraining is included in the quarterly schedule, but urgent retrains outside that schedule are covered too at no extra charge during the 12-month support period.

Yes, but with caveats. Our document-automation models perform best on English text because that is where we have the most training data and annotation experience. For Welsh, French, German, and Spanish we have delivered successful projects with accuracy above 92%. Other languages are possible; we evaluate feasibility during the discovery workshop and may recommend a longer annotation phase.

You do. The trained model, the code we write for your project, and all documentation become your property upon final payment. We retain no licence to use your data or model after the engagement ends. The only exception is open-source libraries we use as dependencies (e.g. PyTorch, scikit-learn), which remain under their original licences.

During model development we run a fairness audit that checks performance across demographic groups present in your data (age brackets, gender, geographic region, etc.). If we find statistically significant disparities, we adjust the training process: resampling, feature exclusion, or post-processing calibration. The results of the audit are documented in the model card you receive at delivery.

Yes. A proof of concept typically takes four to six weeks and costs between £8,000 and £15,000. It covers the discovery workshop, data preparation on a sample dataset, and a working prototype model with evaluation metrics. The goal is to show whether AI can solve your problem before you commit to a full build. About 70% of our proof-of-concept clients proceed to a full project.

We deploy on AWS, Azure, and Google Cloud. If you run on-premise infrastructure, we package models in Docker containers that run on any Linux server with at least 8 GB of RAM. For GPU-intensive workloads (large language models, image recognition), we recommend cloud instances with NVIDIA T4 or A10 GPUs, which cost roughly £0.50 to £1.20 per hour depending on the provider and region.

Yes. Our contracts include a cancellation clause. You pay for work completed up to the cancellation date, calculated as a proportion of the total project fee based on milestones delivered. There is no penalty beyond that. We hand over all code, data artefacts, and documentation produced up to that point.