MUSYG · AI ADOPTION

Organisation tracks

Public-service track — formal gates

Adapt responsibilities and your first pilot to your situation.

Build my route

Principle

Efficiency gains alone do not legitimize a public system. Legal mandate, proportionality, fundamental rights, transparency, accessibility, recourse, and archiving are part of the outcome.

The Mont-Rive City synthetic case applies these gates to an A2 administrative agent that checks planning dossiers. It preserves the counter and postal channel, makes the low case economically non-viable, and keeps 100% of public decisions with qualified staff.

Roles to name before P0

Name each role in writing, with a deputy. One person may hold several roles in a small authority, but the decision authority must stay separate from the evaluator.

Role Responsibility
Decision authority Competent head of service or body that accepts residual risk at P5
Process owner Owns the service, the baseline, and the day-to-day use
Data protection adviser Reviews purpose, data, impact assessment, and retention
Legal service Confirms mandate, legal basis, procurement rules, and recourse
Security officer Reviews access, logging, supplier, and incident handling
Records and archives Defines what is kept, for how long, and in what form
Independent evaluator Reviews the test set, thresholds, and results, and has no stake in the outcome
Staff and user representatives Speak for the people who work with the system and those it affects

Gate summary

Each gate is passed only when its exit evidence exists in writing. The durations are working hypotheses to replace with your own calendar, not published benchmarks. Procurement (P2) and the pilot (P4) usually set the pace.

Gate Enter when Exit evidence Indicative duration
P0 A named problem and owner exist Mandate, baseline, non-AI options, decision authority 1–2 months
P1 P0 is signed Impact assessments, recourse design, consulted services 2–3 months
P2 P1 is signed Contract with audit, log, exit, and change-notice terms 3–6 months
P3 A candidate system and test set exist Independent evaluation report with segment results and limits 2–3 months
P4 P3 is accepted and staff are trained Shadow-mode results against the baseline, complaint log 3–6 months
P5 P4 shows acceptable results Formal decision, registration, monitoring and retirement plan 1–2 months

A full run of about 12 to 22 months is plausible for a first system. Reuse of an already assessed supplier or pattern can shorten P1 to P3, never P4 or P5.

What published evidence supports

Public-sector time-saving figures come mostly from surveys, not from timed tasks. The UK Department for Work and Pensions evaluation had a comparison group, yet its estimate fell from about 19 to about 10 minutes a day without statistical controls. Treat such figures as context for choosing what to measure at P0, and measure your own baseline. See the task-time evidence register. To find the rules that apply to your authority, use the comparison of how administrations govern AI and, for Switzerland and the EU, the legal page.

Gate P0 — Public opportunity

Require:

  • the public-service problem and population concerned;
  • the authority’s mandate and competence;
  • consideration of non-AI options;
  • a baseline for quality, time, cost, access, and errors;
  • an administrative owner and decision authority.

Gate P1 — Legality and impacts

  • qualify purpose, data, decisions, and affected people;
  • determine the impact assessments required;
  • assess non-discrimination, accessibility, language, and digital exclusion;
  • define notice, explanation, challenge, and human recourse;
  • involve data protection, legal, security, archives, and the relevant service.

Gate P2 — Procurement and architecture

  • auditable requirements in the procurement and contract;
  • access to logs, documentation, and evidence;
  • subprocessors, regions, transfers, retention, and training;
  • notification of model changes and incidents;
  • ownership, export, deletion, continuity, and exit;
  • prohibition on supplier secrecy preventing lawful oversight;
  • accessibility and interoperability tests.

The European model contractual clauses for public procurement of AI systems are a guidance resource and must be adapted to applicable law.

Gate P3 — Independent evaluation

  • a test set representative of the population and rare cases;
  • results by segment and severity;
  • tests for security, abuse, drift, and unavailability;
  • independent review of thresholds and evidence;
  • documented limits and uncertainties;
  • no authority to act beyond what was tested.

Gate P4 — Controlled pilot

  • shadow mode before any influence;
  • bounded population, duration, data, and authorized staff;
  • human oversight with the necessary time and authority;
  • complaint channel and immediate stop mechanism;
  • public or internal comparison with the baseline, as obligations require.

Gate P5 — Production and accountability

  • a formal decision and acceptance of residual risks by the competent authority;
  • system registration and appropriate notice;
  • logging, archiving, monitoring, and audits;
  • operational human recourse;
  • service continuity without AI;
  • reassessment after a significant change;
  • a retirement date and conditions.

Stop conditions

Stop or return to the safe process in the event of an unauthorized legal effect, severe discrimination, a data breach, loss of traceability, unavailable recourse, an unevaluated supplier change, or authority exceeding the granted powers.

To print or save as PDF: Ctrl+P (⌘P on Mac).

Source and history · GitHub