This is Week 21 of the AI Ops Playbook, and week five of Round 3. Week 17 was the money. Week 18 was what customers are saying. Week 19 was what you say back. Week 20 was the year you are trying to have. This week is the automation week — and unlike the automation drops in earlier rounds, all four of these have an AI model doing the judgment in the middle, not just moving data from one app to another. That distinction is the whole point of this drop, so it is worth being precise about it. Classic automation is plumbing: when this happens, copy that field over there. It works brilliantly for the small number of tasks in a business that are genuinely mechanical. But most of the admin that eats an owner's week is not mechanical — it needs a decision first. Is this email a commitment or just noise? Which expense category does this receipt belong to? Is this review a nuisance or a legal problem? Did that call actually move the deal forward, or did it just feel nice? Plumbing cannot answer those, which is why so much small business admin has stayed manual long after the tools to automate it existed. What changed is that the judgment step now fits inside the pipe. ## What's New This Week **Zapier: Inbox Triage to Task List** (Internal Ops, Intermediate) — The lead template, and the one that addresses the most widely shared bad habit in small business: using the inbox as a to-do list. Every new email is read and classified into one of three buckets — something you now owe someone, something someone owes you, or noise — and only the first two produce anything. What survives becomes a properly worded task in your task manager: verb first, named person, real due date, and one line of context so it still makes sense in six months. The prompt is deliberately biased toward ignoring things, because a task list polluted with non-tasks stops being read within a week and that failure costs far more than one missed follow-up. Roughly two thirds of inbound email produces no task at all, which is the correct result. The WAITING classification tends to be the sleeper: chasing what you are owed is the discipline almost every small business quietly lacks. **Make.com: Receipt to Bookkeeping Entry** (Internal Ops, Intermediate) — A receipt arrives, by email from a supplier or as a phone photo synced from a folder, and within a minute there is a bookkeeping row: vendor, date, description, a category from your own chart of accounts, amounts split out, and a link to the archived image. The critical design decision is the confidence score. Anything the model is not sure about goes to a review queue with a specific reason — "unsure whether this is Materials or Equipment" — instead of quietly entering a wrong number into your books. We shipped a receipt template in Week 16 that reads a stack of photos in one batch; this is the version where the stack never accumulates in the first place. Bring your real chart of accounts to the setup, not a generic list off the internet. That single input determines whether the output is something your accountant accepts or something you spend an evening recategorizing. **Zapier: Review Monitor to Response Drafts** (Customer Ops, Intermediate) — We already had a template that writes a good reply to a review. It had one flaw: it assumed you knew the review existed. This one solves the noticing. Every platform gets watched, every new review triggers a draft, and the routing is by severity — five-star reviews wait quietly for a weekly batch, three-star reviews land in your email because they usually contain a specific fixable complaint wrapped in politeness, and anything at one or two stars reaches your phone within minutes with the draft plus a one-line read of what the customer is actually angry about. Every draft carries a flag: normal, legal, fake, or names-a-staff-member. Nothing publishes automatically, and that is deliberate rather than a limitation — a review alleging injury needs a human before a word goes public, a coordinated fake review wants an argument, and an employee named in a review should hear about it from you first. **Make.com: Meeting Recording to CRM Update** (Marketing and Sales, Advanced) — The call ends, and before you have walked back to your desk the deal record has been updated: stage, amount, timeline, decision makers, objections raised, competitor mentioned, next step with a date, and a plain-language summary attached as a note. Twelve extracted fields, with the extraction instructed to distinguish what the buyer committed to from what your rep hoped for — because enthusiasm is not a buying signal and a friendly conversation with no next step is not progress. The field most owners end up valuing is RISK: one honest sentence on the biggest threat to the deal, written by something with no ego invested in the outcome. Read a month of those in a row and the pattern in your losses becomes visible in a way your own notes never quite manage. ## The Guardrails Are the Interesting Part Any of these four can be built badly in an afternoon. What separates a version you keep from one you switch off in a fortnight is almost always a constraint rather than a capability, and this week's templates are built around three. The first is bias toward doing nothing. The inbox triage prompt is explicitly told that when it is torn between creating a task and ignoring the email, it should ignore it. Most automations are tuned to catch everything, which produces a system that is technically comprehensive and practically abandoned. The second is routing uncertainty to a human instead of hiding it. The receipt scenario scores its own confidence and sends the doubtful ones to a review queue with the reason stated. The meeting scenario writes UNKNOWN rather than inferring a number that was not said. A tool that admits what it does not know is one you can leave running; a tool that is confidently wrong 10 percent of the time has to be checked 100 percent of the time, which is worse than doing it yourself. The third is the stage-regression lock in the CRM scenario, and it is the most specific lesson in the drop. Models read hesitation as regression. A buyer saying "we may need to revisit the scope" during a negotiation is normal, but an extraction pass reading that call in isolation can score it as a fall back to Discovery — and left unguarded it will quietly corrupt your forecast, with the first person to notice being whoever is staring at a pipeline report wondering why it moved backwards overnight. So the automation may advance a stage or hold it, never regress it. Regressions get flagged for a person. The general rule underneath all three: let the AI do the reading and the drafting, and keep the publishing, the posting, and the irreversible writes with a human. That is not caution for its own sake. It is what makes the automation survive contact with the messy week it was built for. ## Round 3 Continues — The Industry Slate Locks Next Week Round 3 is "Run the Business": eight weeks from the owner's seat. Money, voice of the customer, the content engine, the owner's year, and now AI-native automation are behind us. Next week is the industry week — home services, salon and spa, auto repair, and nonprofit are penciled in for Week 22, and this is the last call to change it. In Round 2, healthcare, legal, fitness, and hospitality all earned their templates because members replied and asked. If your industry's most painful weekly workflow should be on that list, reply to this week's Quick Win email. After this week the slate is locked and we build what is on it. ## Get Started All four templates are available now for Pro members: - **Zapier: Inbox Triage to Task List** — [open the template →](/templates/zapier-inbox-triage-to-task-list) - **Make.com: Receipt to Bookkeeping Entry** — [open the template →](/templates/make-receipt-to-bookkeeping-entry) - **Zapier: Review Monitor to Response Drafts** — [open the template →](/templates/zapier-review-monitor-to-response-drafts) - **Make.com: Meeting Recording to CRM Update** — [open the template →](/templates/make-meeting-recording-to-crm-update) Build one, not four. That is the honest advice for automation week, and it is the opposite of how most people approach a drop like this. Each of these takes 20 to 30 minutes to set up and then needs a few days of running before you can tell whether it is tuned correctly — the inbox triage in particular is calibrated by reading three days of output and adjusting, not by getting the prompt right on the first attempt. Pick whichever of the four maps to the admin you resent most, build it today, and leave it alone until Monday. If you want the fastest visible payoff, it is the review monitor: 20 minutes, and you stop finding out about one-star reviews six weeks late. If you want the biggest one, it is the inbox triage, because the thing it replaces is not typing — it is the low background hum of wondering what you have forgotten.