PlayP10 Automation · Operating work · Editorial rank 9 of 12

Redesign and test one AI-assisted workflow

A recurring workflow consumes material effort, delays useful decisions or produces avoidable errors.

Decision: Should this task remain human, use rules/native tools, add AI, or justify dynamic tool use?
Primary commercial metric: Incremental realised operating savings or useful contribution net of full added cost; capacity reported separately.
Readiness: Basic; execution Connected
Agent role: Optional only when dynamic tool selection beats the fixed process; one candidate with bounded tools, data, time and cost.

Fit

When to use this Play

The work has enough volume or value, inspectable outputs and a credible conventional improvement baseline.

When not to use it

Do not add an AI comparison when stable native automation suffices, volume is too low, the output cannot be evaluated or scarce review capacity is absent. The mechanism-selection step may still conclude with human work or deterministic automation.

Why it should work

Process repair removes unnecessary work; selective AI may improve unstructured steps; each additional layer must earn its cost.

Basic version

One draft/synthesis step and human review, or deterministic automation when it wins.

Prepare

Required maturity

Basic for draft assistance; Connected for system execution. Ready with baseline, reviewer and allowed data. Not ready for actions without a tested authority/stop contract.

P01 Establish decision-specific customer economics — Only when the selected intervention depends on unresolved customer-cohort economics. For workflow cost or other non-customer economics, use the relevant T2 view; P01 is not a prerequisite.Check readiness for P10 (D02)

Inputs

Observed task samples, accepted-output standard, current process A, repaired process B, costs, errors, permissions and staff capacity; candidate AI C only if an additional comparison is justified.

Audience or customer

Internal workflow users; external customers only through an independently approved intervention and execution contract.

Human role

Functional owner owns work and capacity realisation; Finance checks benefit; Operations owns reliability; reviewers have authority and time.

Systems required

Current workflow and approved model access if justified; no custom orchestration required for a fixed task.

Guardrails

Data access, grounded output, exact authority, no model arithmetic, workload limits, manual fallback and independent stop.

How to configure a Play

Copy the Play you need and complete only the relevant parts of T1–T5. Name the people responsible, define the customer group, agree economic and customer limits, and set a review date before starting. An empty decision threshold is not permission to act. Basic, Connected and Advanced describe how the work is configured; Experimental describes its evidence. AI is optional: compare each AI step with a simpler way of doing the same work.

Do the work

Intervention

One redesigned workflow. Retain a sufficient human or deterministic mechanism; add a bounded comparison of B with C only when justified. Keep consequential authority unchanged during initial evaluation.

Workflow

Observe tasks.
Remove redundant steps.
Establish B.
If B is sufficient, record the mechanism decision and move to adoption and retirement; otherwise define one justified AI step and evaluation.
Compare accepted quality/time/cost.
Test edge and failure cases.
Decide.
Retire replaced work only after reliable adoption.

AI role

Workflow under test A specified interpretation or production step whose output a human can verify; no open-ended commercial goal.

Automation role

Stable routing, arithmetic, checks, state, logs and exact approved actions.

Agent role

Optional only when dynamic tool selection beats the fixed process; one candidate with bounded tools, data, time and cost.

Failure modes

Comparing AI with broken A only
counting draft speed instead of accepted work
savings not realised
duplicate processes never retired.

Measure and decide

Measurement

Separate B−A, C−B and any later D−C; include rework, rejected outputs, review, integration, support and incidents.

Primary commercial metric

Incremental realised operating savings or useful contribution net of full added cost; capacity reported separately.

Leading indicators

Accepted-work yield, end-to-end time, error by risk slice, review minutes and queue age.

Experiment

Matched or randomised comparable tasks with blinded quality review where feasible; production actions need offline, shadow and capped canary evidence.

Guardrails

Data access, grounded output, exact authority, no model arithmetic, workload limits, manual fallback and independent stop.

Scale when

Retain or adopt a sufficient conventional improvement after its quality, cost and realisation checks. Adopt an AI addition only if C beats B on the declared quality/cost standard and a benefit owner can realise value; expand authority through a separate gate.

Stop / reassess when

Stop or fall back for unacceptable error, policy breach, review overload, no additional benefit or no maintainable owner.

Learn and progress

Learning

Task boundary, mechanism, evaluation slices, accepted-work economics, failure cases, version and realisation decision.

Practical tool

T4 work design/agent card, T2 benefits and T5 retirement.

T2 Economics and benefits ledger T3 Decision experiment and evidence record T4 Action authority and operating contract T5 Delivery and retirement plan K02 Net intervention value

Next action

sample the full workflow including review and correction.

Open T2 Economics and benefits ledger

Curated next Play

Optional depth

Advanced versionPrerequisites: see Required maturity above.

Bounded dynamic investigation after baseline superiority; autonomy requires a separate same-action delay/review-value comparison.

Evidence for this Play5claims reviewed, with confidence ratings, limits and sources.

Claim classifications

Accepted means a principle or mechanism survived critical review, not that it guarantees a return. Accepted with conditions means prerequisites apply. Qualified means the original claim has been narrowed. Advanced needs substantial readiness. Experimental needs a capped comparison. Unresolved means the evidence cannot settle the claim. Rejected means the recommendation is excluded.

High, Medium and Low describe confidence in the stated target, not the chance of a positive ROI. For a Qualified or Rejected historical claim, confidence applies to the decision to narrow or reject it and to the supported limit—not to the original headline. Read the headline and classification together. An accounting identity or design principle is not empirical evidence of a commercial outcome. Follow the linked chapters, Plays and tools to see where each claim applies.