ChapterCH09 Chapter 9 of 12

AI and automation

Why AI here, and why would an agent be better than a fixed workflow?

AI belongs where interpretation, production or a genuinely variable investigation is a valuable bottleneck. It must improve accepted work, decision usefulness or total operating cost after checking and correction. The statement “AI can do it” is insufficient.

Select the mechanism

ModelM08Model

If the logic is stable and explicit, use a rule, query or native workflow.
If the need is prediction, compare an appropriate statistical model with a simple baseline.
If unstructured evidence needs synthesis or a draft, test a bounded generative step.
If valuable next steps and tool selection change with discovered context, consider an agent comparison.
If the outcome cannot be evaluated or the consequence is ambiguous and material, retain human judgement and execution authority.

Manual → standardise → deterministic automation → optional AI step → justified agent → separately justified authority expansion is a dependency sequence, not a compulsory destination.

Manual → standardise → deterministic automation → optional AI step → justified agent → separately justified authority expansion is a dependency sequence, not a compulsory destination.Dashed: an optional AI or agent step.

Text version

Rule: If the logic is stable and explicit, use a rule, query or native workflow.
Model: If the need is prediction, compare an appropriate statistical model with a simple baseline.
AI: If unstructured evidence needs synthesis or a draft, test a bounded generative step.
Agent: If valuable next steps and tool selection change with discovered context, consider an agent comparison.
Human: If the outcome cannot be evaluated or the consequence is ambiguous and material, retain human judgement and execution authority.

Mechanisms within tasks, not a ladder. Use this model with the conditions and limits described above.

When AI is the wrong tool

Use deterministic software for Finance arithmetic, canonical metric retrieval, customer identity, consent enforcement, assignment, caps and exact approved transactions. Use qualified human judgement for strategy, major investment, pricing, novel claims and complex customer consequences. Native platforms may already provide bidding, routing and campaign execution. A low-volume task may be cheapest to leave manual.

Generative AI can assist customer theme extraction, comparison of market evidence, performance hypotheses, creative drafts and signed-result briefing. It cannot infer a causal answer just because its explanation is plausible. It cannot certify legal compliance. The Prompt Library is limited to six such workflows; calculations and production authority remain elsewhere.

What an agent changes

An agent selects permitted next steps or tools in pursuit of a bounded task, using feedback. A scheduled prompt, report generator or fixed chain of known calls is not an agent merely because it runs unattended. Remove dynamic selection: if quality remains comparable, use the fixed workflow.

No Play requires an agent. A justified candidate starts read-only, with purpose-scoped data, allowed tools, a version, owner, limits and a stop condition. Multiple agents need evidence that isolation, context or independent review improves end-to-end work enough to cover handoff and operating cost. Agreement among similar models is not independent evidence.

Authority belongs to actions

Mode	Permitted role	Required boundary
Assist	Draft or interpret	Human validates before use
Recommend	Assemble options and an evidence-backed proposal	Named owner decides
Human-approved execution	Prepare and perform an exact authorised action	Approval bound to action/version, current-state recheck
Bounded autonomy	Execute an explicitly delegated action	Demonstrated value, empirical reliability, limits, monitoring and stop
Human-only authority	Own the consequential judgement	AI may support evidence but cannot approve

Test normal, missing, conflicting, stale and malicious inputs. External documents are data, not instructions. Keep credentials outside model context; validate arguments, destination, identity, current permission and cumulative exposure outside the model. Record what actually happened. A send cannot be unsent even if later sends can be stopped.

Offline evaluation tests task quality and controls. Historical replay must use only what was known at the time and cannot reveal every counterfactual. Shadow mode has no external writes. A small production canary needs a cap, operator, comparison and rehearsed stop. Assess task correctness, decision usefulness, commercial effect and full cost separately. Zero failures in a small test is not proof of zero risk.

Autonomy has a separate economic test: additional value from faster action plus realised review savings, less extra evaluation, monitoring, recovery and expected loss. Compare the same intervention with and without the approval step. Do not attribute all workflow redesign benefits to autonomy. Model self-confidence never grants permission.

Your next move

Use P10 and T4 to compare one fixed workflow with a bounded AI variant. Keep authority unchanged while testing work quality and cost.

Evidence for this chapter9claims 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.