Direct answer: before expanding an AI workflow, measure the work it sends to people. For every exception, record the trigger, available evidence, queue and owner, wait time, human touches, resolution, downstream effect, and whether the pattern returned. Compare total cost per accepted output—model, integration, review, rework, and escaped error—with the previous process. If automation rate rises while the queue ages, interruptions increase, or the same case returns, the system shifted work; it has not yet proved productivity.
The average can mislead. A vendor reports 82% straight-through processing, so the CMO assumes only 18% of the work remains. That 18% may contain incomplete briefs, unsupported claims, contacts without permission, duplicate accounts, unavailable offers, unexpected languages, or broken integrations. Each case requires context from three systems, interrupts planned work, and produces an explanation. A small queue by volume may dominate cost through variability.
The first-workflow plan at https://makinai.co/insights/en/where-start-ai-marketing-90-day-first-workflow includes baseline and exceptions; https://makinai.co/insights/en/marketing-team-uses-ai-results-unchanged-value-leak follows value to an outcome. This article addresses a narrower decision: expose the manual work created by the workflow itself and decide whether to eliminate, redesign, accept, or stop it.
Build an Exception Work Ledger
- Case and decision — what unit entered, what accepted output was expected, and which stage rejected straight-through processing.
- Trigger — rule, missing value, conflict, low confidence, technical failure, policy, audit sample, or human request.
- Evidence — original input, version, source, missing fields, and observable reason; “AI uncertain” is not a diagnosis.
- Route — queue, owner, priority, service target, authority, and fallback. Routing is not resolution.
- Work — wait, active minutes, systems opened, handoffs, interruptions, edits, and reopenings.
- Resolution — accept, correct, complete, reject, revert, or escalate, with a coded reason.
- Effect — accepted output, delayed campaign, lost lead, improper contact, cost, revenue, or another mature outcome when observable.
- Prevention — rule, data, interface, training, or scope change; a legitimate exception may remain human.
Do not begin with a dashboard. Manually sample complete cases for two weeks. Separate active time from waiting and classify the case before fixing it. One “other” category hides investment; twenty free-text labels defeat comparison. Start with six to ten classes and revise weekly.
Separate valuable judgment from shifted toil
Google SRE describes toil as operational work that tends to be manual, repetitive, automatable, tactical, without enduring value, and linear with service growth. Its source is reliability engineering, not marketing, so do not import Google's capacity target. Use the attributes as questions.
- Valuable judgment — approving a novel claim, interpreting a strategic account, or resolving an ambiguous preference may require context and create durable learning.
- Shifted toil — copying CRM data into a spreadsheet, reformatting every output, reopening the same error, or searching for a source the workflow could attach is repetitive design debt.
- Preventable exception — a missing required field, incompatible taxonomy, unreconciled identity, or late eligibility rule belongs upstream.
- Necessary exception — a high-impact commitment, rights conflict, or genuinely new case may deserve review despite cost.
The goal is not “zero humans.” Reserve attention for decisions where judgment changes the outcome and remove predictable work that exists because the system is poorly designed.
Four marketing examples
Campaign production: AI creates variants, but 35% return because claims lack sources. Adding reviewers treats the queue; linking every claim to approved evidence and blocking generation when evidence expires treats the cause.
CRM enrichment: proposed industry and size conflict with the master record and create tickets. Keeping observed value, proposed value, and source separate enables risk-based review instead of overwriting all records.
Conversion routing: requests without territory go to a general queue. When staff check ZIP code, coverage, and account ownership across three screens, eligibility belongs before routing.
Paid media: an agent prepares budget changes, but every recommendation requires reconstructing the window, margin, and active test. The proposal should carry this evidence; model confidence alone does not measure commercial consequence.
These are hypothetical examples, not MAKINAI clients or measured outcomes.
Decide on cost per accepted output
Calculate direct automation cost + human minutes × loaded cost per minute + rework + expected cost of escaped errors, divided by accepted outputs. Keep the components visible; one dollar total must not erase risk, customer experience, or constrained expertise.
Track straight-through rate; exceptions by class; median and tail age; active minutes; touches; reopen rate; recurrence; accepted output; service-target attainment; escaped error; downstream effect; and hours actually redeployed. Guardrails include cases incorrectly forced through automation, hidden email queues, quality loss, specialist concentration, burnout, complaints, and conversion without a comparator.
Google's workbook recommends an objective unit and continuous measurement, then comparing removal benefit with the cost of building and maintaining automation. NIST recommends observing production differences and defining reviewer responsibilities. AWS shows reviews activated by conditions or sampling. Do not turn those mechanisms into blind policy: low confidence may be harmless, while high confidence can still drive an expensive action.
Run a six-week pilot
- Week 1 — choose one frequent workflow, preserve the baseline, and define accepted output, case unit, and downstream outcome.
- Week 2 — sample complete cases and record trigger, class, time, queue, resolution, and recurrence before redesigning.
- Week 3 — classify judgment, toil, input failure, and technical failure; find the three classes consuming the most minutes or risk.
- Week 4 — remove one upstream cause, attach evidence, adjust routing, and retain a random audit sample.
- Week 5 — release a cohort, compare with the previous process, and monitor backlog, age, escaped errors, and load by person.
- Week 6 — keep, narrow, redesign, or expand; document owner, capacity budget, alert, and monthly review.
Deliverables are the actual-process map, ledger, taxonomy, baseline, economic model, routing rules, test set, dashboard, runbook, capacity limits, and prevention backlog. Cost grows with volume, variability, languages, markets, sources, integrations, permissions, consequence, data quality, audit, retention, observability, and coverage hours.
Marketing owns the decision and accepted output. Operations measures queues and labor. Data and engineering repair evidence and integration. Domain specialists approve high-impact classes. Finance validates cost and capacity, and the people doing the work participate in design. https://makinai.co/insights/en/who-owns-ai-marketing-results-agency-platform-crm separates those roles; https://makinai.co/insights/en/copilot-workflow-ai-agent-marketing-what-build-first helps choose copilot, workflow, or agent.
If a workflow looks fast on the dashboard but slow inside the team, bring MAKINAI at https://makinai.co/services/en/ai-strategy-transformation-consulting two anonymized weeks of exceptions, the prior baseline, and the three queues involved. We can scope a diagnostic and pilot that removes the most expensive cause before adding more automation.