05 · The adoption system

Start bounded.
Scale with evidence.

TADF should enter through a measurable pilot, not a practice-wide mandate. Prove value, expose friction, improve the model, and then scale.

Readiness diagnostic

Is your organisation ready to adopt?

Rate the current state honestly. This offline diagnostic calculates locally and does not save or transmit answers.

01We have named the business and user outcomes AI-supported design should improve.
02Our teams know which design decisions must remain human-owned.
03We use a shared AI-augmented process rather than individual tool habits.
04Every stage has an explicit quality gate and accountable approver.
05AI-supported decisions remain traceable to their evidence and source artefacts.
06We distinguish research-backed claims from assumptions and unvalidated targets.
07We assess judgment, validation, and orchestration, not only prompting.
08Role-based learning plans address the anchor skills each designer must own.
09We have clear policies for data, consent, accessibility, bias, and generated code.
10Overrides, exceptions, and quality drift are reviewed at a defined cadence.
11We have baseline measures for speed, quality, rework, trust, and outcomes.
12Pilot results can change the operating model rather than merely confirm it.

30 / 60 / 90-day adoption

00–30

Frame

  • Name executive sponsor and governance owner
  • Select one product team and two priority stages
  • Baseline time, quality, rework, and risk
  • Assess anchor skills
  • Agree pilot success and stop criteria
Output: signed pilot charter
31–60

Operate

  • Train stage-specific agents and skills
  • Run human checkpoints visibly
  • Capture corrections, overrides, and exceptions
  • Hold weekly quality reviews
  • Track operating friction and unintended effects
Output: evidence-backed pilot log
61–90

Decide

  • Compare outcomes against the baseline
  • Review quality and capability movement
  • Change weak standards or agent boundaries
  • Define scale conditions and investment
  • Publish a transparent adoption decision
Output: scale, revise, or stop decision

Pilot playbook

One team. Two stages. Six weeks.

Choose frequent work with observable quality signals. Avoid the highest-risk regulated use case as the first experiment.

Recommended entry

Experience Architecture + Experience Validation

  • Clear existing artefacts
  • Frequent team activity
  • Observable usability and accessibility signals
  • Enough judgment to test the checkpoint model

Pilot evidence

  • Time before and after
  • Rework and revision cycles
  • Quality-gate pass rate
  • AI suggestions corrected or rejected
  • Designer confidence and stakeholder trust

Stop conditions

  • Data or consent risk cannot be contained
  • Quality decreases despite faster output
  • Human checkpoints become rubber stamps
  • Evidence cannot be traced
  • Operating cost exceeds measured value

Leadership responsibilities

01

Set the ambition

Name the outcomes AI-supported design should improve and what must never be traded away.

02

Fund the operating change

Allocate time for capability, governance, evidence capture, and measurement, not just licenses.

03

Protect the checkpoints

Prevent speed pressure from turning human review into ceremonial approval.

04

Review the evidence

Decide whether to scale, revise, or stop based on observed outcomes and risks.

Measurement model

Measure more than speed.

A pilot succeeds only when velocity, quality, capability, trust, and outcome move together without unacceptable risk.

Velocity

Time to charter, insight, validated concept, specification, and resolved issue.

Quality

Task success, errors, clarity, accessibility, coherence, trust, recovery, and drift.

Capability

Anchor-skill maturity, corrections, overrides, escalation quality, and learning transfer.

Adoption

Checkpoint compliance, active use by stage, governance exceptions, and abandoned workflows.

Outcome

Conversion, adoption, retention, satisfaction, revenue, cost, or risk reduction, context dependent.

Risk

Privacy findings, unsafe output, bias gaps, unsupported claims, and irreversible errors.