Skip to content
PHAROS
← Journal
03 / Field guide

One workflow. One owner. One stopping rule.

A practical way to begin an AI programme without pretending the whole organisation is ready.

A technician guides a metal component into a calibration jig with a physical stop.
PHAROS JOURNAL / 03A boundary for the work · Film still

Start where the work can push back

Imagine a team reviewing supplier contracts.1 The documents arrive in different formats. A reviewer checks clauses against an approved position, flags exceptions and prepares a recommendation for someone with authority to accept the risk.

An AI demonstration can make the first pass look impressive. It finds clauses, compares language and drafts a summary. The temptation is to call that success and widen the programme.

The real workflow begins where the demonstration ends. What happens when two clauses conflict? Which source wins? Who can decide that an exception is acceptable? Does the correction stay local, or is it safe to reuse? If the answer reaches a supplier, who authorised the message?

A transformation brief is too large to expose these questions. One workflow can.

Choose the boundary before the tool

Map the present path from source to action, including the unofficial handoffs and corrections that keep the work moving. Sociotechnical design research has long argued that technical and organisational systems should be designed together.5 The model is only one participant in that path.

A useful first workflow passes five tests:

  • the work repeats often enough for the team to learn from it;
  • the current problem is visible without inventing a future return;
  • the permitted source set is narrow enough to inspect;
  • one named owner can approve, reject or pause the result;
  • the first run can be withdrawn without destabilising core operations.

Decide what the first cycle must return

Before the test, record how the workflow behaves now. The useful measure might be elapsed time, rework, missed exceptions, unsupported statements or reviewer effort. Choose one because it can change the next decision, not because it will decorate a dashboard.

During the cycle, keep a compact trace: the evidence available, the system's recommendation, the human correction, the authority exercised, the action taken and the result observed later. NIST's AI Risk Management Framework treats mapping, measurement, governance and ongoing management as part of the AI lifecycle.4 Pharos's trace turns those responsibilities into something a team can inspect inside one piece of work.

The cycle should end with a decision brief, not a victory lap. The brief records what happened, what remains unknown and the next authorised move.

Give stopping the same status as scaling

Public digital-service guidance separates discovery, prototype testing, limited rollout and wider release so that each stage answers a different question before commitment grows.23 An enterprise AI programme needs the same permission to stop.

Give the four possible decisions equal status:

  • Stop when the problem is not material, the evidence cannot be used safely or the harm cannot be controlled.
  • Revise when the workflow matters but the scope, data, model role or human review is wrong.
  • Repeat locally when the result is promising but the evidence remains narrow.
  • Expand only when the result is worth preserving and the next context is similar enough to justify another test.

One workflow will not transform an organisation. It will show whether the organisation can learn before its commitments harden.

Sources

Source record updated 5 November 2026. External findings stay within their published scope. First-party Pharos observations and interpretations are identified.

  1. 1

    Pharos. Illustrative supplier-contract review. Constructed example, 2026.

    The workflow is hypothetical. It is not a customer case, deployed product claim or measured outcome.
  2. 2

    Government Digital Service. How the discovery phase works. GOV.UK Service Manual, n.d..

    Official guidance on understanding users, constraints, problems and measures before committing to build; includes the option to stop before alpha.
  3. 3

    Government Digital Service. How the alpha and beta phases work. GOV.UK Service Manual, n.d..

    Official guidance on testing risky assumptions before limited rollout and wider release. The transfer to enterprise AI is Pharos synthesis.
  4. 4

    National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST, 2023.

    Voluntary, use-case-agnostic framework organised around Govern, Map, Measure and Manage; not a certification or sector-specific rule.
  5. 5

    Chris W. Clegg. Sociotechnical Principles for System Design. Applied Ergonomics 31(5), 2000.

    System-design principles integrating technical and organisational considerations; conceptual and practice-oriented, not AI-specific.