I have been building custom AI workflows into multi-unit operations for about two years now. Some of them saved thousands of hours across a region. Some of them broke in interesting ways and taught me what not to build. This post is the honest map: what worked, what did not, what the pipe looks like, and what it actually costs.
The context. I ran regional operations across 21 franchise units for a $36M P&L, and now I do the same fractionally for a $30M Bay Area group and a handful of smaller operators. What follows is not theory. It is what is running in production right now, and what I would build first if I walked into a new operation tomorrow.
The first rule: automate documentation before judgment
Every operator I talk to about AI wants to automate the hard thing first. Labor forecasting. Menu engineering. Vendor negotiation. Those workflows are seductive because they promise to remove real cognitive load from a general manager.
They also fail hard. When AI gets forecasting wrong you overspend on labor for a week before you notice. When AI writes a bad end-of-shift note, nothing breaks. Start where the failure cost is low.
The workflows that hold in production, in my experience, are the ones that replace documentation and communication time. The manager still makes the decision. The AI drafts the artifact. That framing keeps the risk contained and the value obvious.
Assist, do not replace. The AI workflows that survive give managers back their time. The ones that try to replace their judgment fail in a way you do not notice until the P&L catches up.
The stack, minimally
You do not need a data team. You need three layers, and one operator who is willing to prototype and rip things out.
Fig. 1 · Three layers. Nothing exotic.
Triggers
Something happens in the operation and you want to react. Toast fires a webhook when a check closes. 7shifts fires an event when a shift ends. An Airtable form gets submitted. A scheduled cron runs at 5am. An email lands in a shared inbox. Every automation starts here.
Pipes
Zapier for simple flows. Make (formerly Integromat) for anything with branching logic. n8n if you want to self-host. Native webhooks and scheduled queries when the pipe should not depend on a third-party tool. Pick one primary pipe tool per operation and stick with it. Splitting across three tools is where debugging goes to die.
AI
Custom GPTs on ChatGPT Team accounts for structured drafts that non-technical managers will run themselves. Claude Projects for longer-form analytical tasks. The GPT or Claude API for anything triggered by a pipe (because Custom GPTs and Projects are user-triggered, not pipe-triggered). A small vector store for context if you have SOPs, menus, or historical documents you want the AI to reference.
Nine automations that hold
These are the workflows that survived past six months in real operations. Not every operator needs all nine. Most operators should start with two or three.
1. The daily flash draft
Every morning at 6am, a Make scenario pulls yesterday's sales, labor, food cost variance, and top three items from Toast and 7shifts. Feeds the numbers into the GPT API with a fixed prompt. Out comes a two-paragraph flash report in the operator's voice. Drops into the group Slack at 6:15am.
Time saved: roughly 20 minutes per unit per day. For a 21-unit region, about 7 hours a day of area-director drafting time.
Where it failed early: Version one wrote too much. Managers stopped reading it. Version two capped the output at 90 words. Adoption jumped.
2. Catering intake summarization
A catering client emails in an order in prose. Sometimes with attachments. A Zapier flow forwards the email body and any attachment text to a Custom GPT with a schema: date, delivery time, contact, headcount, dietary flags, address, special notes. Output goes to Airtable as a structured record. Catering manager reviews and confirms.
Time saved: roughly 4 minutes per order. For an operation running 60 catering orders a week, that is 4 hours.
Where it failed early: Dietary restrictions in prose ("no shellfish for two guests, one is celiac") got flattened into a single flag. Version two extracted a per-guest dietary list and flagged anything ambiguous for human review.
3. Shift-note rollup
Each closing manager types a two-line shift note into 7shifts at end of shift. At 5am the next morning, a scheduled scenario pulls all the notes across the region, feeds them into Claude with a structured prompt, and returns a five-line regional summary: what themes appeared, which units flagged issues, which incidents need followup.
Time saved: the area director used to skim 21 shift notes every morning. Now they read a summary in one minute and dive into the ones that matter.
4. Comp and void anomaly alerts
Every hour during operating hours, a scenario pulls comp and void volume by location. Compares to the trailing 30-day median for that day and hour. Anything above two standard deviations triggers a Slack alert to the general manager and area director with the number, the trend, and the top three items involved.
Time saved: catches theft patterns and service-recovery events in hours instead of at end of month.
Where it failed early: During a promotional day the alert fired every hour. Version two added a suppression rule for known promo days pulled from a marketing calendar.
5. Weekly P&L narrative draft
Every Sunday night, a scenario pulls the week's P&L by unit from QuickBooks Enterprise. Feeds into Claude with the prior four weeks for context. Returns a draft weekly narrative: what moved, why (based on the data available), what to ask about in Monday's meeting. Area director edits it in 10 minutes instead of writing it in an hour.
Fig. 2 · Draft the weekly narrative before the operator wakes up.
6. Health inspection prep
A Custom GPT trained on the local health code plus the operation's own SOPs. General manager pastes in the top-of-mind concerns and the date of the pending inspection. Out comes a targeted prep checklist for that specific unit, prioritized by common inspection findings in the region.
Time saved: the general manager gets an inspection prep checklist in 90 seconds instead of an hour. Failure rate on repeat inspections dropped by about a third in the first six months of use.
7. Vendor followup automation
Every Monday, a scenario pulls open invoices, missing credits, and short deliveries from a shared Airtable log. Feeds them into a Custom GPT that drafts followup emails in the operator's voice, one per vendor, with the specifics attached. Purchasing manager reviews and sends.
Time saved: vendor followup used to slip constantly because it took 2 hours a week nobody had. Now it takes 20 minutes and it does not slip.
8. Guest complaint categorization
Every guest complaint email or Google review below three stars flows into a Custom GPT. Returns a category (service, food, environment, wait time, other), a severity, and a suggested next step. Feeds a weekly report by category, by unit. What used to be a stack of incident reports is now a heat map.
9. New-hire welcome sequence
When a new hire is added to 7shifts, a Zapier flow fires. Custom GPT drafts a personalized welcome message referencing their role, first shift, and the manager they will report to. Drops into their email and text. The general manager gets 45 seconds back per hire and the new hire feels seen.
What failed and why
Three workflows I built that did not survive. Sharing them because knowing what fails matters as much as knowing what works.
Labor forecasting from prior sales
I built a Custom GPT that took the last 12 weeks of sales by day-part and returned a proposed schedule. It worked well on units with stable demand. It failed on units with promo swings, weather sensitivity, or a nearby event calendar. The general managers stopped trusting it after two bad weeks, which is the right response. Custom labor forecasting needs a real time-series model, not a language model. See the separate post on custom AI for labor forecasting.
Menu engineering recommendations
I fed six months of item sales and margin data into Claude and asked what to cut. The recommendations were mathematically defensible and operationally wrong. Claude did not know that the low-margin salad was the item that got a party to book a table for eight. Menu decisions are a data-plus-context problem. AI helps with the data half.
Real-time customer chat responses
We tried a Custom GPT to draft responses to Instagram DMs. The tone was off in a way that Bay Area guests picked up on immediately. Pulled it after two weeks. The right shape for customer-facing AI is much narrower: intake forms and structured responses, not conversation.
The cost, honestly
For a 5 to 20 unit operation:
- ChatGPT Team seats for the operators who will run Custom GPTs: $30 per seat per month. Usually 4 to 8 seats.
- Claude Pro or API for structured analytical work: $20 per seat, or usage-based on the API.
- Make or Zapier: $50 to $250 per month depending on scenario volume. Make is cheaper per operation for anything past 5,000 monthly runs.
- POS API access: often included in the POS contract. Toast and Square include reasonable webhook access. Some legacy POS vendors charge extra or make it painful.
- Airtable or a lightweight database: $20 to $60 per seat per month.
Total software spend for a working stack: $200 to $800 per month. The real cost is the operator time to design, iterate, and maintain. Budget one operator half-time for the first 90 days, then quarter-time ongoing. If you cannot allocate that person, do not build the stack. Buy a vertical tool instead.
Build first, buy second
Every vertical AI tool for restaurants promises to solve the workflow you are trying to build. Sometimes they do. More often they solve a slightly different workflow, and the mismatch takes six months to notice.
The move I now use: prototype the workflow in Make or Zapier with a Custom GPT for two weeks. Learn what the workflow actually needs. Then decide whether to keep the prototype, upgrade it, or replace it with a purpose-built vendor tool.
You will keep about half the prototypes. That is a good outcome. The half you replace, you will replace with better specs than any operator who never prototyped can write.
Rollout sequence
The order I would build these in, if I were starting fresh in a multi-unit operation tomorrow:
- Week 1 to 4: Daily flash draft. Prove the pipe end to end. Get one automation running and adopted.
- Week 5 to 8: Shift-note rollup and comp anomaly alerts. Same pipeline pattern, different content. Compound the learning.
- Week 9 to 12: Catering intake and vendor followup. Documentation replacements. High value per hour.
- Month 4 to 6: Weekly P&L narrative and health inspection prep. Longer-form workflows that require some iteration on prompts.
- Month 7 onward: Custom GPTs for field leaders (see the separate post). Anything experimental.
The point
AI automation in multi-unit operations is not glamorous. It is dozens of small workflows that give managers back their time. Done well, a working stack returns 15 to 30 hours a week across a mid-sized regional operation. Done poorly, it turns into a graveyard of half-built scenarios nobody trusts.
Start with documentation. Prove the pipe. Iterate publicly. Rip out what does not hold. And keep the human in the decision seat, always.