The Acceptance Gap

The Acceptance Gap is our model for where engineering effort goes in the age of agents. As generation becomes abundant and cheap, the binding constraint moves to the next scarce step: acceptance — the judgement that a change is correct, safe and worth shipping.

The four stages

• Generated — an agent produces a candidate change (abundant, fast, cheap).
• Reviewed — humans and deterministic gates inspect intent, architecture and correctness.
• Trusted — the change clears governance: security, compliance, operability, ownership.
• Shipped — it reaches production and creates value.

Everything in agentic engineering — context engineering, the harness, evaluation, governance, the operating model — is a way of narrowing the gap between Generated and Shipped. You close it by raising acceptance, not by generating faster.

You measure it the same way: acceptance rate and rework rate, alongside the DORA flow and stability metrics — never lines of code generated.

Anti-patterns

• Measuring generation. Counting AI-written lines or PRs opened — celebrating the abundant step while the scarce one, acceptance, goes untouched.
• Scaling reviewers. Trying to close the gap with more human review stamina instead of better gates. Reviewer fatigue is not a strategy.
• Trusting the demo. Accepting because it looked right in a demo, not because it cleared deterministic checks and governance.
• A uniform gate. One heavy review applied to everything, regardless of reversibility or blast radius.

A maturity model: how the gap behaves as you mature

• Experimentation — large gap. Generation happens in pockets; acceptance is ad hoc and personal.
• Assistance — large gap. Copilots everywhere, productivity assumed, acceptance unmeasured.
• Automation — visible gap. Generation outpaces review and the verification tax becomes obvious.
• Agentic Engineering — managed gap. The harness, context and deterministic gates close most of the distance; humans accept only at the doors that matter.
• AI-Native — optimised gap. Acceptance is engineered, measured and continuously narrowed; the gap is small and known.

Diagnose your acceptance gap

• Do you measure acceptance rate and rework — or output (lines, PRs)?
• When an agent's change is accepted, what evidence cleared it beyond a human reading the diff?
• Which deterministic gates — tests, types, evals, policy — run before a human ever looks?
• Is review proportioned to blast radius, or applied uniformly?
• Can you say where the gap is widest in your pipeline today?

Illustrative — a composite, not a specific client: a team that doubled merged AI-authored PRs in a quarter and called it a productivity win. Rework rose in lockstep and the net delivered change was flat. Generation had doubled; acceptance had not. The gap, not the output, was the story.

From insight to action

The Acceptance Gap is the lens beneath our AI work, and it is deliberately a delivery idea, not an AI one. We are not claiming to have closed it in production at scale — the field is too young for anyone to claim that honestly. What we bring is the engineering discipline that narrows it: small reviewable scopes, feedforward context, deterministic gates wired in as feedback, and measurement that counts what lands rather than what is typed. The The AI Engineering Maturity Model locates where your gap sits today; Designing the Human Acceptance Loop: Where People Must Stay in the System as Agents Take Over the Build is how you govern the acceptance itself.