Skip to content

03

Grounded AI delivery

Most AI features are built demo-first and governed later, if at all. This is the reverse: the guardrails are designed before the feature, because retrofitting them means rewriting the feature.

Commercial terms

Fixed-fee phases. The first phase is usually a narrow, high-value workflow rather than a general assistant — narrower scope is what makes evaluation possible.

Not the right fit if

  • A general-purpose chatbot for a website — that is a content problem, not this
  • A use case where nobody can define a correct answer, because it cannot be evaluated
  • Research or fine-tuning work — this is applied delivery, not model development

The problem

A model that is right 95% of the time is not a business system. The five per cent is silent, plausible, and arrives with the same confidence as the rest.

In a regulated or audited process, an answer nobody can trace is worse than no answer. It creates liability the organisation did not know it had.

The failure is rarely the model. It is the absence of a contract between the model's output and the evidence it is allowed to rely on.

What we actually do

  • Bind generation to selected source facts, and reject invented references before anything is persisted
  • Design refusal semantics: a typed error and a clear message rather than a confident guess
  • Tripwire prohibited claims in code, so a policy breach fails a test rather than reaching a customer
  • Build an evaluation harness with cases drawn from real work, run on every change
  • Cap cost per call and per month in the application, not in a dashboard nobody watches
  • Log the route, model, prompt version and cost of every generation

What you get

What you get.

Everything produced is written down and handed over. Nothing lives only in someone's head, and nothing depends on us to keep running.

  • The feature, with guardrails as code rather than as prompt instructions
  • An evaluation suite covering the failure modes that matter, wired into CI
  • A written account of what the system will not do, and why
  • Cost telemetry with enforced ceilings
  • A review process for changing prompts and models, with the evidence retained

Who it is for

A business unit that wants AI assistance in a process with real consequences
A risk or compliance function that has blocked an AI initiative for good reasons
A product team that has a demo and needs something an auditor will accept
An organisation where a wrong answer costs money, not just credibility