Responsible AI

AI you can put your name on.

Every system we ship carries its own guardrails: evaluation before launch, monitoring after, human oversight where judgment matters, and audit trails throughout. Not a compliance slide — engineering practice.

What this means in practice

  • Evaluation as a deliverable

    Accuracy, grounding, and failure modes measured on your data before go-live, and monitored after.

  • Guardrails by design

    Input validation, output constraints, permission enforcement, and escalation paths defined with your team.

  • Human oversight where it matters

    Confidence-based routing to people; autonomy expands only as the evidence supports it.

  • Auditability

    Decision logs and traceable outputs, designed around your security and compliance requirements.

Proof

Our own hiring

The agent that writes our candidate profiles ran behind a human gate for its first weeks; every correction became a rule, and the gate opened only when the corrections stopped. A year in production since.

Read the field note
Human gate first
Rules from corrections
In production since early 2025

When teams come to us

  • An AI pilot that legal or security won't sign off
  • A regulated environment where “the model said so” is not an answer
  • A board asking what could go wrong

How we build it

  1. Risk map

    what can go wrong, and who it affects

  2. Guardrails design

    constraints, permissions, escalation paths

  3. Evaluation before launch

    on your data, your cases

  4. Oversight in production

    monitoring, audit trails, human review where it matters

What can go wrong, by kind

Responsible is not one property. It is four different failures, each with its own defence:

Wrong

the model answers with confidence and the answer is false. Defence: grounding on your data, citations a reader can check, and "I don't have that" as a legitimate output. Measured by an evaluation set of real cases, run on every change.

Unauthorised

the system reads or does something the user could not have done by hand. Defence: permissions enforced at query time, not by prompt; an explicit list of the actions it may take.

Unexplained

a decision nobody can trace after the fact. Defence: every output logged with the input, the model, the retrieved sources and the rule that allowed it.

Unsupervised

an action with consequences that no person approved. Defence: a gate that a human holds until the action class has a track record, and that closes again when the surprises return.

A slide about ethics prevents none of these. Engineering prevents all four.

How autonomy gets earned

Every system we ship starts supervised: a person approves what it proposes. That is not a limitation to apologise for. It is how the system learns the edge cases and how your team learns to trust it. Each correction becomes a rule; the log of corrections is the most useful artifact of the first month, and it is the reason the rules are right about where the problems actually are rather than where someone guessed they would be.

When corrections approach zero for one kind of action, the gate opens for that kind — and stays closed for the ones that still surprise you. Autonomy is granted per class of work, on evidence, and it is reversible: monitoring in production watches for the day the inputs drift, and a rise in errors closes the gate again. This is the shape we run on our own operations, and the shape we build into every client's.

Works well with

Built into every system, never bolted on.

FAQ

  • Is this a compliance checkbox?

    It's engineering practice: measurable accuracy, traceable decisions, defined escalation. Paper doesn't supervise anything.

  • Will guardrails make the AI useless?

    Done well, constraints raise trust and adoption. The useless assistant is the one nobody trusts.

  • Who is accountable when the AI errs?

    A named human path. That’s what escalation design means.

  • Does this slow the project down?

    The evaluation set and the gate add about a week to a first build, and they take months off the part nobody budgets: the incident, the rollback, the stakeholder who stops trusting the system after one confident wrong answer. Trust is the expensive thing to rebuild.

  • How do you handle data that can't leave our environment?

    By not moving it. The system runs inside the access you grant — your cloud, your permissions, your audit log. Where a commercial model is the right tool, we use terms that exclude your data from training; where that isn't enough, open-weight models run inside your environment.

  • What do we get that we can show a regulator or a board?

    The evaluation results on your cases, the decision logs, the list of actions the system may take and who approved each class of them, and the runbook for what happens when it fails. Not a policy document — the evidence a policy would have to point at.

Ship AI your stakeholders can trust.