← Interactive field guide · David Seymour — Open Work
Version 3 · Proposed framework for practitioner testing · Reference edition updated 20 October 2026.
This page contains the guide’s explanatory material, all fictional cases, evidence prompts and record fields without requiring JavaScript. It does not contain the separately supplied companion paper or anyone’s working record.
The unit of review is a delegation in a workflow: a task, population, configuration, owner and period of use. Apply it to agents, copilots and consequential automation. A label does not determine the need for oversight.
Distinguish reading, recommending, drafting, submitting, changing and delegating. Record when each is permitted and when a person must approve. Capability does not confer permission.
Stable identity, configuration, individual execution, credential and delegated authority are separate records. Rotating a credential does not create a new agent.
A booking can be restored after a time-sensitive opportunity has passed. A draft can influence a decision before it is corrected. Assess harm, timing, scale and recovery, including effects mediated through people.
Approval requires a reviewer with context, time, competence and authority. Credential revocation is one containment control, not proof that all activity has stopped.
Describe the service problem, people affected, task boundaries, alternatives and current baseline. Involve those who will receive the service as well as those operating it.
Output: a mandate and an evaluation plan.
Use representative and difficult cases, misleading inputs, failures and recovery exercises. Record case selection, outcomes, limitations and the human work required.
Output: evidence with scope, date, version and unresolved concerns.
State allowed actions, approval conditions, data boundaries, duration, stop rules and the accountable role and holder. A pilot requires its own authorisation.
Output: a human decision with limits, not a badge generated by this guide.
Verify completed actions. Track outcomes, exceptions, review capacity, patient effort and unexpected interactions. Keep a meaningful route to a person.
Output: attributable activity and service evidence.
Compare benefit and burden, examine subgroup differences and inspect incidents. Reassess after material changes to models, tools, policies, data, memory or workflows.
Output: an explicit judgement about continued use.
Renew or narrow the delegation, or withdraw authority and resolve pending work. Preserve appropriate records and ensure a dependable human or alternative service.
Output: an updated mandate or a completed retirement and handover.
A lens on the people, institution and working practices behind AI. Confidence should be proportionate to evidence for a defined use.
The original Trust Equation: Trustworthiness = (Credibility + Reliability + Intimacy) / Self-orientation.
From Maister, Green and Galford, The Trusted Advisor (2000). Their associated materials emphasise the human basis of intimacy and self-orientation. Original framework.
No trust score. This OPL adaptation uses the dimensions as questions. It does not measure a model’s intentions or allow strengths to cancel serious failures. No endorsement by the original authors is implied.
Look for: Demonstrated competence and sources that support the actual statement.
Test the task and population. Verify material facts and calculations. Separate observations, inferences and assumptions.
Do not substitute: Fluency, confidence, citations that do not support the claim, or agreement among models.
Evidence prompt: what would make this conclusion false?
Look for: Verified outcomes and a dependable process under stated conditions.
Check the resulting state, not just a promise. Track failures and acceptable variation across runs. Reassess after material changes.
Do not substitute: Availability, repeated wording, or a log entry saying an action was attempted.
Evidence prompt: where is the completion check and what happens when it fails?
Look for: Context fidelity, appropriate disclosure and demonstrable data handling.
Explain what is collected, where it goes, who can see it and how long it remains. Respect constraints and provide access to a person.
Do not substitute: A warm tone, simulated empathy or an assumption that nobody else can access the conversation.
Evidence prompt: would the person endorse the actual data flow?
Look for: Visible objectives, incentives and consequences across stakeholders.
Examine supplier and organisational incentives. Record conflicts and hidden burdens. Allow a conclusion that less AI or a different process is better.
Do not substitute: Assuming absence of an ego means absence of conflicting incentives.
Evidence prompt: who gains and who supplies uncounted effort or bears risk?
Fictional design exercises, not deployment recommendations or evidence from real services. Hidden consequences are included here for inspection.
Net staff hours = cases × (gross staff minutes released − review − rework − coordination) ÷ 60. Additional patient hours = cases × additional patient minutes ÷ 60. Running cost is reported separately in GBP, not converted to hours. A positive balance does not establish overall benefit, safety or financial return.
All defaults below are illustrative. Use one comparable case population and period; avoid double counting. Inputs must be finite and non-negative, at most 1 billion; case count must be a positive whole number. Input basis is illustrative, estimate or observed; observed inputs require a scope and source. Display rounds to one decimal, after calculation at full precision.
| Field | Meaning | Unit | Default |
|---|---|---|---|
| volume | Number of comparable cases | cases | 100 |
| gross | Gross staff minutes released per case | minutes | 6 |
| review | Human review per case | minutes | 2 |
| rework | Correction and rework per case | minutes | 1 |
| coord | Coordination and exceptions per case | minutes | 1 |
| patient | Additional patient effort per case | minutes | 3 |
| cost | Additional running cost for this period | GBP | 0 |
Full reference content (JSON) · Record structure (JSON Schema) · Blank editable record (JSON)
There is no submission API, cloud saving or automatic transmission by the guide. Reloading discards unsaved changes. JSON is a local interchange format, not an authorisation. Structure validation cannot verify evidence, calculations or decision-making authority. Feedback can be raised through Open Work’s contact route.
name — Name of this delegationreviewBy — Review duepurpose — Purpose and intended outcomebeneficiaries — People affected and intended beneficiariesscope — Setting population and exclusionscomparator — Current process or simpler alternativerole — Accountable roleholder — Current human holder of that roleops — Operational owner and contact routeexperience — Service or experience owner and contact routesuccess — Benefit measures and acceptable limitsescalation — Route to a human and ownership coveridentity — Stable identity and execution attributionconfig — Model tools policy and configuration versionsactions — Allowed actions and conditionsexclude — Prohibited actionsdata — Data access disclosure and retention boundariesapproval — Approval points and reviewer capacitydelegation — Subtasks delegation chains and shared dependenciesstop — Stop recovery and service fallbackchanges — Changes that trigger reassessmentbudget — Cost rate and resource limitsdecision — Decision under considerationdecider — Authorised decision-makerdecisionDate — Decision dateconditions — Reasons conditions residual uncertainty and next actionEvidence states: unknown = Unknown, declared = Declared, evidenced = Evidenced, tested = Tested, concern = Concern or failure, na = Not applicable. Evidenced and tested require a reference, reviewer, date and finding. Not applicable requires a reviewer, date and reason. Review dates must be valid and not in the future. These are user-reported states, not verified grades or a total score.
purpose — Purpose and comparisoncompetence — Task competence and claim supportauthority — Authority identity and attributionsafety — Consequences oversight and recoverydata — Context and data handlingreliability — Operational reliability and completionbenefit — Benefit and displaced burdenincentives — Interests and incentivesportfolio — Interactions and dependencieschange — Lifecycle and reassessmentDecision values: Not recorded; Gather more evidence; Redesign or use a simpler process; Authorise a bounded pilot; Continue within stated conditions; Restrict or suspend; Retire.
This is an AI-assisted conceptual synthesis, revised through red-team review. It has not been validated through a practitioner study. Agreement among model outputs is not independent expert consensus.
Anatomy reveals dependencies; ecology reveals interactions and resource burdens. We have dropped the taxonomic tree. In biological usage, parasitism is a type of symbiosis; mutualism describes mutual benefit.
Here, beneficial coexistence concerns the people and institutions using AI. It does not assume that the AI has welfare interests. Ask whose work improves, whose burden increases and who bears risk.
Compare decisions made with existing language and with this guide. Use comparable cases, counterbalance order, and include staff and people affected. Measure missed hazards, false reassurance, responsibility clarity, time and agreement on next steps.
Check whether the metaphors add value beyond the record. Preserve contrary findings. A successful small exercise supports usability, not universal validity.
Delegated work replaces taxonomy. Purpose and benefit join authority and accountability. Identity is separate from credentials. Action-specific consequences replace automatic autonomy ceilings. Evidence states and unknowns replace deployment verdicts. The Trusted Advisor adds questions about warranted reliance. Earlier versions remain intact.
The companion v3 paper is supplied separately and covers the argument, examples, limitations and operating cycle. Sources checked 19 October 2026; inclusion does not imply endorsement. The original shared chat was unavailable; supplied local papers and Trusted Advisor reflections informed the revision.
[1] Figlioli et al. · Healthcare’s AI agent management challenge — practitioner hook.
[2–3] Trusted Advisor Associates · Trust Equation and core concepts — attribution and limits of the analogy.
[4] Anthropic · Building effective agents — architecture and workflow patterns.
[5] Feng et al. · Levels of Autonomy for AI Agents — proposed interaction framework.
[6] IETF · Token revocation — containment constraints.
[7] HHS · Minimum necessary requirement — scope and exceptions.
[8] NHS · Make your service clinically safe — applicable safety processes, including DCB0129 and DCB0160.
[9] OpenStax · Relationships and microbiomes — biological terminology.
[10] NIST · AI RMF Playbook — established risk-management functions.
Suggested citation: OPL Advisory. Managing delegated AI work, field guide version 3 (19 October 2026), reference edition updated 20 October 2026. Include the relevant section URL and your access date. This proposed framework is not practitioner-validated; cases are fictional and evidence remains user-reported. Source inclusion does not imply endorsement. This reference does not grant a new reuse licence for the guide or linked sources.