Lesson 86.3 gave every campaign a brief with one promise. Now comes the production question: the brief calls for three emails, two variants, a research summary, and formatted assets — by Friday. AI can do most of that drafting in an afternoon. The lesson is knowing which half it must never decide alone.

Where AI genuinely helps

Give AI the preparation jobs: high-volume, reversible, checkable work where a human judge stays cheap and fast.

• Draft variations. Three subject lines, two email bodies, five call-to-action phrasings from the approved brief — same promise, different wording. The brief constrains; AI multiplies.
• Classify feedback. Sort 200 replies into "booked / question / objection / unsubscribe / out-of-office" so a human reads the twenty that matter.
• Summarize performance. Turn the week's sends, opens, clicks, and bookings into three plain sentences plus one anomaly worth investigating.
• Prepare research. Collect what competitors promise on their landing pages, with links, so the human positions honestly.
• Format approved assets. Resize, lay out, convert the approved copy into the email template, the page box, the PDF — no wording changes during formatting.

The pattern from Class 84 holds: AI proposes, evidence constrains, the human disposes. Every AI output enters through a review queue; nothing it drafts reaches a reader unapproved.

What a human owns — non-transferable

Five things stay human, no matter how good the draft:

1. Claims. "Cut inference costs 40%," "the only checklist that…," any number, superlative, or guarantee. A human verifies against evidence or deletes it. Fabricated proof is the fastest way to burn the trust Parts XIII–XIV built. 2. Budgets. How much spend, on which channel, with what kill threshold. Money movement is a red-line decision (Lesson 84.4). 3. Partners. What goes out under someone else's name — guest slots, co-branded assets, affiliate copy. Their reputation, their rules, their sign-off. 4. Audience selection. Who gets contacted, and who is suppressed. Mailing the unsubscribed or the already-converted is a human failure, not a tool glitch. 5. Brand-sensitive replies. Frustration, complaints, billing confusion, public mistakes. Lesson 87.2's handoff rules apply here too: ambiguity and emotion route to a person.

A useful test: if the output is wrong, who apologizes? If the answer is a human with a name, that human approves before sending.

The contract plus the gate

Apply the Class 84 automation contract to the campaign. Name the trigger, the approved input, the bounded action, the observable output, the owner — and the permissions the workflow must never receive (no list deletion, no suppression editing, no unsupervised sending, no budget changes).

Then add the review gate: the explicit checkpoint between AI preparation and reader delivery.

brief (human) → AI draft → review queue (human) → approve/edit/kill → send → log

Make the gate concrete: who reviews (named owner, plus a backup), what they check (claims against evidence, promise matches the brief, suppression applied, links work), and where the decision is logged. A gate without a log is a rumor — when a complaint arrives in month three, the log says who approved which wording and why.

Walk it with Sonariq. Trigger: "Day 6 campaign touch due for stalled-finishers segment." Input: the approved brief plus last week's reply classifications. Bounded action: AI drafts two Day-6 variants and a performance summary. Output: drafts in the review queue — never in the sender. Owner reviews claims ("fitted in 25 minutes" — is that still true?), checks suppression (12 newly booked, 3 unsubscribed — excluded?), approves one variant, logs the decision. The workflow may *prepare*; only the owner *sends*.

Exercise: contract one campaign step

Take your Lesson 86.3 brief and put its riskiest AI-assisted step under contract.

# AI-CAMPAIGN-CONTRACT.md — [Campaign + step]

## The step
- Workflow: ___ (e.g. "draft Day-6 variants + summarize week-1 performance")
- Trigger: ___ / Approved inputs: ___ (brief version, data sources)

## Bounds
- AI may: ___ (draft / classify / summarize / format — from the safe list)
- AI must never: ___ (send / edit suppression / change claims / move budget / reply to complaints)
- Output lands in: ___ (review queue location, never the sender)

## The gate
- Reviewer + backup: ___ / checks: claims vs ___ / promise vs brief ___ / suppression ___ / links ___
- Decisions logged where: ___ (date, approver, variant chosen, reason)

## Failure paths
- Ambiguous reply → ___ / complaint → ___ / AI invents a claim → ___ (kill + rewrite rule)

Finish line: one step of your campaign under a written contract with a named reviewer, a checkable gate, and logged decisions.

Verify quickly: ask "if the AI quietly rewrote our guarantee overnight, what stops it reaching readers?" If you can point to the queue, the reviewer, and the log — the gate is real. If the answer is "we'd probably notice" — it is not.

Common failure mode: the unsupervised sender — AI connected directly to the mailing tool "to save time," suppression edited by automation, a fabricated statistic mailed to 4,000 people. Recovery costs more than every hour the shortcut saved. Prepare fast; approve deliberately.

Check your understanding

1. Name AI's five safe marketing roles and the five things a human must own. 2. Draw the gate flow from brief to log. Where does AI output land, and why there? 3. What three permissions must a campaign workflow never receive?

Next

Drafts are out, approvals logged, sends recorded. But did anything work — and what should change? Lesson 86.5 turns the week's numbers into reporting that decides: result, plausible cause, evidence, uncertainty, and one next test.