Delegating Work to AI

A practical test for deciding what to hand off, what to review, and what humans still own

Philip Budden

SpecifyWhat are we asking for?
VerifyCan we check it?
JudgeWhat stays human?
OwnWho stands behind it?

About Me

Phil Budden
Senior Data Engineer & Data Platform Lead, Codi Group

  • Working across data, AI strategy and responsible AI adoption in social housing
  • Technical practitioner increasingly using AI to delegate substantial knowledge and engineering work
  • Interested in the gap between what AI can produce and what we can dependably hand over to it

The Practical Delegation Test grew out of a practical question: how much of this work should I actually hand over?

The Productivity Story Is Too Narrow

AI can generate plausible work quickly, but generation is only one stage in the workflow.

Defineintent, audience, constraints
Generatenow much cheaper
Verifystill needs attention
Integratefit with the real workflow
Ownaccountability remains

The Better Question

Too narrow
Can AI produce this?
More useful
Can this human-and-AI system deliver this dependably?

Count specification, verification, judgment and accountability.

Review Debt

AI can create candidate work faster than qualified people can check it.

Generatecandidate work arrives quickly
Queueunchecked work accumulates
Reviewexpert attention remains finite

Where The Test Sits

Control stack showing the Practical Delegation Test between strategic direction and formal governance controls

The Practical Delegation Test

01Can I describe the outcome clearly?success, failure, audience, constraints
02Can I verify the result cheaply enough?time, expertise, detectability, reversibility
03What judgment must remain human?belief, risk, commitments, trade-offs
04Where does the uncertainty go?absorbed, named, escalated, or dumped

Question 1: Outcome Clarity

AudienceWho is the output for?
AcceptableWhat would make this good enough?
UnusableWhat would make it unsafe or wrong?
ConstraintsWhat must it include, exclude, or obey?

If the brief is vague, use AI for exploration rather than execution.

Outcome Clarity Failure Pattern

Weak task
Write something good about this.
->
Stronger task
Draft a two-page decision brief for this audience, covering these points, excluding these claims, and naming remaining uncertainty.

Weak clarity moves the task toward exploration.

Question 2: Verification Cost

TimeHow long does meaningful review take?
ExpertiseWho can detect important failures?
DetectabilityAre errors visible or hidden?
ReversibilityCan missed errors be corrected?
Check structureIs there a repeatable test?

Verification Is Not One Thing

Cheap to check

  • Narrow tests
  • Reconciliations
  • Known examples
  • Objective criteria

Expensive to check

  • Subtle omissions
  • Tacit context
  • Live consequences
  • Persuasive but untested claims

Question 3: Retained Human Judgment

DraftAI can produce candidate wording
CompareAI can surface options
Commithumans own promises
Accept riskhumans own trade-offs

`Human in the loop` is too vague.

Judgment Is Not Always One Person

Originatorwhat is being asked for
Executorwhat has been produced
Reviewerwhat must be challenged
Approverwhat can be released
Ownerwho stands behind the outcome

Question 4: Where Does The Uncertainty Go?

Diagram showing where uncertainty moves through a handoff

Good handoffs absorb, name, or escalate uncertainty instead of passing it on as polish.

Legitimate Review Vs Hidden Transfer

Legitimate review

  • Expected, scoped and resourced
  • Residual uncertainty named
  • Reviewer can challenge

Hidden transfer

  • Implicit, improvised or dumped
  • Uncertainty obscured by polish
  • Reviewer pressured to accept

From Questions To Delegation Levels

Direct delegationclear, testable, low residual judgment
Checkpointed delegationuseful execution, important review points
Human-led with AI supportambiguity or judgment remains central
Do not delegatetoo unclear, sensitive or hard to verify

Three Modes Of AI Use

Diagram comparing three modes of AI use

Worked Examples

Meeting summaryclear enough, but decisions, actions and tone need reviewCheckpointed
Unit testsnarrow task, objective execution path, quick failure signalsDirect + review
Complaint responseformat can be supported, but remedy and commitments stay humanHuman-led
Redundancy rankingsevere consequences, contestable assumptions, legitimacy dominatesDo not delegate

Agentic Workflow Example

Classifycase type and risk
Retrievepolicy and evidence
Draftcandidate response
Routeexceptions and escalation
Reviewfinal visible output

Apply the test at workflow or sub-workflow boundaries.

Final review may hide routing, retrieval, classification and escalation failures.

Accountability Controls

Approvalexplicit release points
Ownershipnamed accountable person
Evidencewhat was checked
Challengereviewers can reject
Traceabilityproportionate record

Review Power Matters

Timeenough space to check
Contextenough information to judge
Authoritypermission to reject
Safetycover to challenge speed pressure

A review step without power is often just a rubber stamp with nicer paperwork.

Maturity Over Time

Diagram showing delegation maturity over time

A Lightweight Team Worksheet

Taskname the work and audience
Checksverification path and reviewer
Judgmentwhat stays human
Decisionlevel, uncertainty, controls

The worksheet should fit in a meeting note. If it becomes ceremony, use a team default or split the task.

What Good Use Looks Like

Cleaner handoffsless work arriving as polished uncertainty
Faster safe delegationstable low-risk tasks move more easily
Better disagreementborderline tasks expose the real issue
Less cleanupreview debt is noticed before it compounds

The Leadership Ask

Do not govern AI only through tool approval.
Workflowwhere the work happens
Review debtwhere checking accumulates
Uncertaintywhere doubt is carried
Accountabilitywho can answer for it

Closing

Not the goal
Maximum delegation
The goal
Dependable work with clear ownership of remaining judgment

Use the four questions before cheap generation creates expensive ambiguity.

Q&A

What kind of task are you currently unsure whether to delegate?

01Can I describe the outcome clearly?
02Can I verify the result cheaply enough?
03What judgment must remain human?
04Where does the uncertainty go?

Find out more

philipbudden.co.uk/practical-delegation-test/

  • Canonical whitepaper for senior leaders, sponsors and accountable owners
  • Practitioner guide for applying the test to your own work
  • Feedback form, with an option to respond anonymously

Please share successes, difficulties, disagreements with the framework or its questions, and work the four questions did not fit.

Return to the Delegation Test