We refused a £400k contract in 2023 because the client wanted to deploy facial recognition without consent protocols. That decision shaped everything we do.
Artificial intelligence built on principles you can verify
Ethical AI firm designs, audits and deploys AI systems for organisations that need results they can defend publicly. Based in Wales, working across the UK.
Founded on a principle
Two former data scientists from the NHS left stable roles after watching a predictive model quietly deprioritise patients from certain postcodes. They incorporated Ethical AI firm with one rule: every model we build must be explainable to the person it affects.
First bias audit for a Welsh council
A local authority asked us to review their benefits-allocation algorithm. We found that single-parent households were scored 18% lower on average due to a proxy variable the vendor never disclosed. The council paused the system within a week.
Open-source fairness toolkit
We released a Python library for demographic parity testing. It now has over 2,600 stars on GitHub and is used by three UK universities in their MSc programmes. We maintain it with quarterly updates.
Turned down the surveillance contract
A retail chain offered us a six-figure project to build in-store emotion detection. We said no. The board published a blog post explaining why, which generated more inbound enquiries than any marketing campaign we have ever run.
ISO 42001 certification
We became one of the first fifteen firms in the UK certified under the new AI management system standard. The audit took four months and required documenting every model lifecycle from scoping to decommission.
Today: 23 people, same principle
Our team includes machine-learning engineers, a philosopher, two data-protection lawyers and a former regulator. We still apply the same test to every engagement: can we explain this model to the person it affects?
What we actually do
We do not sell generic "AI strategy". Every engagement falls into one of these capability areas, each with defined deliverables and timelines.
| Capability | What you receive | Typical duration |
|---|---|---|
| Bias and fairness audit | Quantified bias report across protected characteristics, remediation roadmap, executive summary | 3–6 weeks |
| Model governance design | Policy documents, approval workflows, risk registers, board-ready dashboards | 8–12 weeks |
| Responsible ML engineering | Production model with explainability layer, monitoring pipeline, drift alerts | 10–20 weeks |
| Regulatory readiness | Gap analysis against EU AI Act and UK framework, compliance action plan | 4–8 weeks |
| Team training | Workshops for engineers, product managers and leadership (in-person or remote) | 1–3 days per cohort |
Are we the right fit?
We are selective about the projects we take on. Not because we are precious, but because ethical AI work done badly is worse than no AI work at all.
If your organisation matches three or more of the criteria on the right, we should talk.
Start a conversationYou handle sensitive personal data
Health records, financial decisions, employment screening, housing allocation or criminal justice.
You face regulatory scrutiny
FCA, ICO, CQC or sector-specific bodies have asked questions about your automated decision-making.
You have existing models in production
Something is already running and you are not confident it behaves fairly across all user groups.
Your board asks hard questions
Leadership wants evidence that AI investments are defensible, not just profitable.
Our method, step by step
Scoping interview
A 90-minute session where we map your AI systems, data flows and the people affected by automated decisions. No slides. We ask questions and listen.
Risk and impact assessment
We classify each system against the UK government's AI assurance framework and the EU AI Act risk tiers. You get a prioritised list, not a generic matrix.
Technical deep-dive
Our engineers examine model architecture, training data provenance, feature importance and performance across demographic slices. We test for disparate impact using our open-source toolkit.
Stakeholder translation
Findings are written in plain language for three audiences: the engineering team, the legal/compliance function and the board. Each gets a tailored document, not a rephrased version of the same report.
Ongoing monitoring
For production systems, we set up automated fairness dashboards that flag drift before it causes harm. Alerts go to a named responsible person in your organisation, not a shared inbox.
What clients have said
"They found a proxy-discrimination issue in our lending model that two previous vendors missed. The fix improved approval rates for under-served groups by 11% without affecting default rates."
— Head of data, UK challenger bank
"The board training session was the first time our non-executive directors genuinely understood what our algorithms do. That changed the quality of oversight overnight."
— Chief operating officer, housing association
Questions we hear often
How much does a bias audit cost?
It depends on the number of models, the volume of training data and how many protected characteristics we test. A single-model audit for a mid-size organisation typically runs between £12,000 and £28,000. We provide a fixed quote after the scoping interview.
Do you build AI systems from scratch?
Yes, but only when the use case passes our ethical review. We will not build surveillance tools, social scoring systems or anything designed to manipulate behaviour without informed consent. For qualifying projects, we handle everything from data engineering to production deployment.
Can you work with our existing engineering team?
That is our preferred model. We embed with your engineers for the duration of the engagement, transfer knowledge through pair programming and code reviews, and leave your team capable of maintaining fairness standards independently.
What industries do you work with?
Financial services, healthcare, local government, housing and education make up most of our portfolio. We have also worked with two recruitment platforms and a criminal-justice analytics provider. If your AI system affects people's access to services or opportunities, we are likely a good match.
Get in touch
Tell us what you are working on. We respond to every enquiry within two working days.
116 East Avenue, Hayesridge, Wales, AS6 9RJ, United Kingdom
Responsible AI is not a feature. It is a practice.
Read our story from the beginning