Responsible AI.

An AI system deployed into a business is an engineering artefact with consequences, not a demonstration. It should be evaluated before it ships, monitored after it ships, accountable to a named human, transparent to the people it affects, and cheap enough to justify its existence. Most of what goes wrong with AI in production is not exotic. It is a system nobody tested, nobody watched, and nobody owned.

Evaluation before deployment. Every agent we put into production has a test suite. We build golden datasets and regression harnesses so that changes to a prompt, a tool, or an underlying model are measured rather than debated. Exploratory prototypes are a normal part of the work and are run under human review, clearly labelled as prototypes, and not relied on for decisions until they have been evaluated.

A named human owns every system. Before an agent goes live, there is a person on the client’s side accountable for its outputs, with the authority to pause it. Autonomy is scoped deliberately: we define what an agent may do without approval, what requires a human decision, and what it must escalate.

Transparency to the people affected. Where a system converses or interacts directly with a person, that person should be able to tell they are dealing with an AI system rather than a human. We design to this by default, and build additional disclosure where a client’s regulatory position requires it. How a client labels AI-assisted content it publishes under its own brand is the client’s editorial decision, and we build to whatever standard they set.

Data minimisation. An agent gets the least data required to do its job, with the client’s access controls carried through rather than bypassed. A common failure in AI deployments is an agent that can see more than the person asking it. We design against that specifically.

Monitoring after launch. Systems we build emit traces: what was called, what came back, what was decided, what it cost. Drift, failure rates, escalation volumes and cost per task are visible from day one, because a system nobody is watching is a system nobody can be accountable for.

Model-agnostic by design. We do not tie a client’s system to one model provider. Models change, prices change, and providers change their terms. Our architectures let a model be swapped without rebuilding the system around it.

Client data is never used to train foundation models, and is never used to improve our own systems or other clients’ systems. We enforce this two ways: contractually, through data processing agreements with our clients and with every AI provider we use; and technically, by using enterprise service tiers whose terms exclude customer data from training, with zero-retention configuration where the provider offers it.

Where a client wants a model trained or fine-tuned on their own data, that is a deliberate, contracted decision with defined scope, and the resulting model belongs to them.

We use AI tools in our engineering and business operations, because it would be strange to sell agentic engineering and not practise it. That use is governed by an internal acceptable use policy. In short: approved enterprise tools only, on service tiers that exclude our data from training; client data only within the client’s approved environment and tooling; no client confidential information in personal or consumer AI accounts; and a human reviews anything that reaches a client.

AI writes a meaningful share of the code we ship. All of it is reviewed by the engineer accountable for it, and it goes through the same tests, scanning and review as anything else. We do not treat generated code as trusted input.

Some things we decline regardless of budget.

• General-purpose systems that score or rank people’s trustworthiness or social behaviour.
• Systems that infer or score individuals’ emotional state in order to evaluate them as employees or students.
• Biometric categorisation by race, political opinion, religion, or sexual orientation.
• Systems designed to manipulate people into decisions against their own interests, or that exploit vulnerability.
• Systems that impersonate a specific real person without their consent.
• Fully automated decisions with legal or similarly significant effects on a person, where no human can review the decision and there is no route of appeal.

To be clear about what this does not exclude: risk, fraud and credit models, sentiment and voice-of-customer analysis, propensity scoring, and branded AI assistants that are plainly identified as AI are all normal parts of our work.

We also decline work where the intended use requires an accuracy we cannot demonstrate. We would rather lose the engagement than ship a system that appears to work.

The regulatory position is moving and varies by jurisdiction. In the United States, sector regulators in financial services, healthcare and employment are setting expectations faster than any general AI statute. In the European Union, the AI Act’s transparency obligations are already in force, with obligations for high-risk systems applying from December 2027. State privacy and automated-decision rules are tightening in parallel.

We track this because our clients operate in the sectors where it lands first, and because a system designed against next year’s requirement is cheaper than one retrofitted to it. We are not a law firm, and on anything consequential we work alongside your counsel rather than in place of them.

If you believe a system we built is producing harmful, biased, or incorrect outputs, tell us: responsible-ai@augmentedx.ai. We would rather hear it early and awkwardly than late and formally.