AI Business Automation Playbook: What to Automate

Most “AI automation” roadmaps are product catalogues wearing a strategy hat. A real playbook answers three questions before it names a vendor: What is worth automating? What must never run without a human? Who owns the failure at 2am?

This is that playbook for operations-minded teams — deeper than AI automation for small business (the 30-day first-steps plan). Use the SMB guide to ship one boring win. Use this page to scale across departments without building a silent mess.

Tool spine: Zapier for glue most teams can edit; n8n when control and economics demand it (Zapier vs n8n, Make vs n8n, Zapier alternatives); ChatGPT / Claude for draft and classify steps with gates. Productivity context: complete AI productivity guide. Agent reality check: complete AI agents guide.

The operating principle

Automate work that is:

1. Frequent — weekly or more (daily is better).
2. Rule-shaped — a new hire could follow it in three sentences.
3. Low blast radius if wrong — annoying, not existential.
4. Already happening — you are paying humans to do it now.
5. Observable — you can detect failure within a business day.

Deprioritise: nuanced complaints, pricing exceptions, legal/HR edge cases, “summarise the company every morning” dashboards nobody reads, and anything that invents a customer commitment.

From the productivity guide: automation only escapes review debt when the step is truly rule-based. The moment an LLM writes external copy and auto-sends it, you reintroduced review with worse optics.

Three layers (do not buy layer 3 first)

You do not need an agent platform to stop copying form rows into a CRM.

Risk classes (label every candidate)

If a stakeholder cannot name the risk class, the automation is not ready for design review.

ROI: a simple model that survives contact with finance

Monthly value ≈ (minutes saved per run × runs per month ÷ 60) × loaded hourly cost
Monthly cost ≈ tool subscription + model API + (maintainer hours × loaded hourly cost)
Net ≈ value − cost

Maintenance is not free. A Zap nobody understands is a liability. Budget 10–20% of the saved time for ownership, or the savings are fictional.

Time-to-value gates

Ambitious multi-app “AI employees” that skip week 1 almost always fail in month one. Boring wins compound.

Platform choice (company view)

Pricing shapes: Zapier tasks (our review checked 2026-07-27: free small allowance + two-step only; Professional multi-step from about $19.99/mo annual framing). n8n: executions on Cloud or infra on self-host (n8n.io/pricing). Always model your volume; see Zapier alternatives.

Hard rule: shared credential vault standards beat five personal Zapier accounts on the company card.

Department playbooks

Each section lists Tier A (ship) and Tier C (govern heavily). Tier B sits between: useful after A is stable.

Sales and revenue ops

AI fit: classify inbound intent; draft personalised first lines from real CRM fields only — never invent case studies the company does not have.

Marketing

Same honesty bar as the writing guide: unreviewed AI copy does not ship.

Customer support

Classification prompt shape (short):

Labels: billing, tech, sales, spam, other. Return JSON only: label, confidence high|medium|low, reason. If not high confidence, label other.

Finance and ops

People / HR

Product and engineering

Meetings and knowledge (cross-cutting)

Capture is not automation. Otter and peers create transcripts; workflow tools push confirmed actions to trackers. Consent and policy first — see meeting transcription and meeting assistants. Never auto-join sensitive rooms.

Design patterns that survive production

1. Dual-run before cutover

Manual or old path stays live until outputs match for a defined sample.

2. Failure is a feature

Every production flow notifies on error. “Green but wrong” needs sampling (spot-check CRM rows weekly).

3. Idempotency

Webhooks retry. Design for duplicates (unique keys, upserts).

4. Least privilege credentials

Integration users, not founder god-mode OAuth, wherever vendors allow.

5. Human gates as explicit steps

Do not hide approvals in “we’ll watch Slack.” Use a ticket state, form approve, or queue.

6. Document in five bullets

Trigger, happy path, failure path, owner, link to run history. Future-you is a stranger.

7. One AI step at a time

Classify or draft or extract — not a twelve-node agent fantasy on day one.

Governance without bureaucracy

Shadow IT rule: personal free Zapier on company data is still company risk. Provide an approved path so people stop inventing worse ones.

90-day rollout (beyond the SMB 30-day plan)

Days 1–30 — Foundation

• Inventory repeated work (interviews beat surveys).
• Score with frequency × clarity × inverse risk.
• Ship two R0/R1 flows with failure alerts.
• Pick primary platform (usually Zapier) and logging standard.
• Read SMB automation if the team has never shipped glue.

Days 31–60 — Assist layer

• Add one classification AI step with low-confidence human path.
• Add one draft path for support or sales with mandatory human send.
• Measure edit time; if humans rewrite 80%, fix prompts or kill the step.
• Start department catalogue.

Days 61–90 — Scale and fork

• Migrate or rebuild the worst cost offender (see Zapier vs n8n).
• Retire vanity automations.
• Train a second owner per critical flow.
• Present ROI and incident log to leadership — numbers, not vibes.

Free and paid stack honesty

Deep free-tier walls: what free AI plans include, free tools for small business.

Do not stack five paid “AI employee” subscriptions in month one. One solid glue tier + one chat tool you already use beats a graveyard of trials.

Anti-patterns (kill these)

1. Automating a broken process — you get faster garbage.
2. No owner — orphan Zaps at ex-employee emails.
3. Auto-send charm — LLMs invent policies under pressure.
4. Dashboard theatre — daily AI summaries nobody reads.
5. Replatforming for Twitter — migrate for measured pain only.
6. Agent washing — multi-step glue renamed “agentic” in a deck.
7. Ignoring model bills — workflow invoice ≠ OpenAI/Anthropic invoice.

RACI for automation (keep it light)

If “accountable” is a committee, nothing ships and nothing dies. One name per row.

Executive one-pager (what leadership needs)

Leaders do not need node graphs. They need:

1. Hours returned per month on the top five flows (measured, not aspirational).
2. Incident count (silent failures caught vs customer-visible misses).
3. Spend — automation tools + model APIs + maintainer hours.
4. Risk posture — how many R2/R3 flows exist and whether human gates are real.
5. Concentration risk — single owner, single vendor, single integration user.
6. Ask — budget for platform tier, training, or a migration pilot with explicit success criteria.

If the one-pager cannot be filled, you do not have a programme — you have hobbyist glue. That can still be valuable at small scale; call it what it is.

Change management (the part tool vendors skip)

People resist automation when it threatens expertise, exposes process chaos, or dumps review work on them without removing other work.

Practical moves:

• Automate the steps people hate, not the steps that define their craft.
• Publish before/after time samples so wins are visible.
• Train the editors, not only the first builder.
• Keep a human exception path for weird cases — total automation of edge cases creates shadow processes.
• Never surprise a customer-facing team with a bot that emails their book of business.

Meeting capture deserves a special callout: assistants that join calls without culture work create political incidents faster than they create productivity. Pair any meeting assistant rollout with consent scripts and default-off sensitive calendars.

Vendor evaluation scorecard (copy/paste)

Score 1–5 each; weight by your context:

Fill with evidence from a pilot, not from memory of a demo. Cross-check with Zapier alternatives and Zapier vs n8n.

Sample quarterly OKRs (illustrative)

Objective: Make automation a measured ops capability, not a side quest.

• KR1: 80% of production flows have a named owner and failure alert (catalogue audit).
• KR2: Three R0/R1 flows each show ≥4 hours/month measured savings net of maintenance.
• KR3: Zero R3 flows with unsupervised external side effects (audit).
• KR4: Model + automation spend forecast within 15% of actuals for the quarter.

Adjust numbers to company size. The point is measurability, not theatre OKRs.

Verdict checklist

Before any new automation ships, every item below must be true in writing:

1. Risk class labelled R0–R4.
2. Owner named (a human, not a channel).
3. Failure alert configured and tested.
4. ROI hypothesis written, even roughly.
5. Human gate decided for anything R2 or higher.
6. Dual-run plan for flows that touch customers or money.
7. Docs — five bullets in the catalogue.

If any item is pretend, the automation is not production — it is a demo with credentials.

Further reading

• AI automation for small business — first 30 days
• Zapier AI review
• Zapier vs n8n for AI automation
• Best Zapier alternatives
• Make vs n8n
• Complete AI productivity guide
• Complete AI agents guide
• Best AI for email
• Best AI meeting assistants
• Best productivity AI tools
• What free AI plans actually include