It depends on the service. For document classification we need at least 500 labelled examples per category. If you have the documents but they are not labelled, we can help you label a sample during the discovery phase. Demand forecasting requires a minimum of 18 months of timestamped records so we can capture seasonal patterns. Chatbots need a knowledge base of at least 40 distinct articles, help pages or policy documents. If you are unsure whether you have enough, send us a brief description and we will tell you straight away.
Getting started
All data stays on UK-based infrastructure that you control. During the project we access it under a data-processing agreement that specifies who can see what, for how long, and for what purpose. We delete our working copies within 30 days of project completion. We never use client data to train models for other clients, and we never send it to third-party APIs unless you explicitly request it and sign a separate consent.
No. We design every system so that a single technically confident staff member can manage day-to-day operations using a web-based admin panel. That person does not need programming skills; they need to be comfortable with a web browser and a spreadsheet. We include twelve months of support after go-live, during which we handle bug fixes, performance checks and any questions your team raises. After that, ongoing support is available on a monthly retainer starting at £400 per month.
Yes, and we encourage it. The proof-of-concept stage exists precisely for this purpose. We take a slice of your data, build a working prototype, and measure its accuracy against the metric we agreed during discovery. If the results are not good enough, you walk away having spent between £2,000 and £5,000 instead of the full project cost. About one in five pilots does not progress, usually because the underlying data turns out to be too inconsistent. We would rather find that out early.
Technical and security questions
Our chatbots use retrieval-augmented generation, which means the language model only produces an answer when it can ground it in a specific document from your knowledge base. If no relevant document is found, the bot says so and offers to connect the resident with a human agent. We also run automated checks that compare the bot's output against the source text and flag any response where the similarity score falls below a threshold. In testing across three council deployments, the hallucination rate was below 1.5 per cent of conversations, and every hallucinated response was caught by the flagging system before reaching a second user.
Model drift is real and we plan for it from day one. Forecasting models retrain automatically each week on fresh data. Document classifiers retrain monthly. We set up alerts that fire when accuracy drops more than five percentage points below the baseline measured at go-live. During the twelve-month support period we investigate any alert within 48 hours. If the drift is caused by a genuine change in your operations (a new document type, a restructured team), we update the model and its training data at no extra cost during that first year.
In most cases, yes. We have built integrations with Civica, Northgate, Jadu and several bespoke Access and SQL Server databases. If your system exposes an API or allows ODBC connections, integration is straightforward. If it does not, we can usually work with file exports (CSV or XML) on a scheduled basis. We scope the integration during discovery and include it in the fixed-price quote so there are no surprises later.
Every interface we build meets WCAG 2.1 AA as a minimum. Dashboards support screen readers and keyboard navigation. Chatbot widgets include ARIA labels and work with assistive technology. We test with both automated tools (axe, Lighthouse) and manual keyboard-only walkthroughs before delivery. If you need AAA compliance for specific components, let us know during discovery and we will include it in the scope.
Commercial questions
You do. Once you have paid the final invoice, all code, models and documentation transfer to you under a perpetual, royalty-free licence. We retain no proprietary claim. The only exception is generic open-source libraries we use (such as scikit-learn or FastAPI), which remain under their original licences. We list every dependency in the project handover pack so your IT officer knows exactly what is running.
We are familiar with public-sector procurement in the UK. For engagements under the direct-award threshold we can work on a purchase-order basis. For larger projects we are happy to respond to formal tenders and can provide references from previous public-sector clients. We are also registered on the Digital Marketplace (G-Cloud) under the cloud-support category, which simplifies procurement for central and local government buyers.
The proof-of-concept stage is specifically designed to catch this early. If the prototype does not meet the agreed success metric, you are not obliged to proceed. If a problem emerges after go-live, we fix it under the twelve-month support agreement. We have never had a client request a full refund, but our terms of service allow for partial refunds if deliverables are materially different from the agreed specification.
Yes. We run a full-day workshop called "AI for civic teams" that covers what machine learning can and cannot do, how to evaluate vendor claims, and how to identify good candidate projects within your organisation. The workshop costs £1,200 per session and accommodates up to 20 participants. We also provide hands-on training specific to the tools we build for you, included in every project at no additional charge. That training covers daily operation, troubleshooting common issues and interpreting model outputs.
Still have questions?
Drop us a line at [email protected] or call +44 842 505 3759. We reply within two working days.
Contact us