61 agents and automations · Running daily at Rejuve Salon Spa
What runs every day
Everything below runs in production at Rejuve Salon Spa, the founder's own business — the proving ground before anything reaches a client. Results come from run logs; where we haven't measured something yet, we don't claim it.
Intake Form Reminder Workflow
New guests arrive with their paperwork done — without the front desk chasing it.
The problem
New guests are supposed to complete an intake and consent form before their first visit. About half did. The rest filled it out at the desk, which pushed appointments late and left gaps in the guest record.
What it does
Tracks which upcoming guests still have an incomplete form. Sends a text 24 hours before the appointment and again the morning of, and alerts the front desk when a guest arrives with the form still open. No AI involved — every step is known in advance, so a plain automation is the right tool.
Form completion averaged 53% in the six weeks before reminders started and 80% in the weeks after (77–83% weekly, September 2026). Before-and-after, not a controlled test.
INTAKE FORMS — Tomorrow New guests booked: 6 Forms complete: 4 Reminder sent (24h): 2 ✓ Morning-of reminder: queued 8:00 AM Front desk alert: on arrival if still open
Review Reply Drafting
Every new Google review gets a reply drafted in your voice — and a person approves it before it posts.
The problem
Reviews went unanswered because writing a thoughtful reply takes time nobody had. Generic replies are worse than none, and a careless reply to an unhappy guest can make things worse in public.
What it does
Checks Google Business Profile three times a day. For each new review it matches the provider and the service the guest had, then drafts a reply in the business's voice. The draft arrives as an approval card in Teams — Approve, Edit, or Deny. Nothing posts until a person approves it, whatever the star rating.
88 of 92 new reviews answered since June 2026 (96%), every reply approved by a person before posting.
NEW REVIEW · ★★★★★ "Best color I've had in years." Provider matched · service: color + gloss DRAFT REPLY "Thank you! We're so glad you love it — see you at your next visit." [ Approve ] [ Edit ] [ Deny ]
Fleet Watchdog
When something breaks at 2 a.m., a named person hears about it — once, with the fix attached.
The problem
With dozens of scheduled workflows, the dangerous failure is the quiet one: a login expires, an API changes, and a workflow keeps "running" while doing nothing. Nobody notices until a customer does.
What it does
Watches every scheduled workflow's runs and logs. When a new kind of error appears it alerts once, with context and the remediation runbook, instead of flooding a channel with repeats. A separate health check restarts always-on services if they stop.
27,974 workflow runs from July to October 2026, 0.63% failed — and every new kind of failure raised an alert instead of failing silently.
WATCHDOG · 06:15 New error type: review-sync "401 Unauthorized" — 3 runs since 05:00 Likely cause: expired Google login Runbook: re-authorize, then re-run Other workflows: all normal
Recruiting Ads with Earned Autonomy
It runs the campaign day to day — inside a budget and a time window the owner approved.
The problem
Recruiting ads need daily attention to stay efficient, and nobody had time for it. Handing an AI an ad account with no limits is not an option.
What it does
Drafts recruiting campaigns and sends the flight dates and budget to the owner as an approval card. Once approved, it adjusts and prunes ads on its own — but only inside that approved budget and window. Every live change is logged, spend is tracked against a ledger, and an emergency stop halts everything.
Spend cannot exceed the budget the owner approved; every live change is logged. Autonomy here was granted in stages, not on day one.
CAMPAIGN PROPOSAL · Stylist recruiting Flight: 14 days Budget: [APPROVED AMOUNT] Audience: licensed stylists, 25 mi [ Approve flight + budget ] [ Edit ] After approval: daily pruning within budget Spend ledger · change log · emergency stop
And more
Before the team arrives: today's bookings, projected revenue, utilization, unconfirmed and new guests. Runs on a local model; names are added after the AI step, never sent to it.
Forecast vs. actual at close of business, with a short "why we missed" section. Code calculates every number; the model only writes the narrative.
Finds guests who cancelled and never rebooked, and posts a follow-up card with their history. A person makes the call — the workflow just makes sure nobody falls through.
Five rule-based signals per guest — visit cadence, loyalty, spend — feed one daily "what matters" digest. Guest identities never reach the AI.
One form submission triggers 15 steps: employee record, Microsoft 365 invite, SharePoint folders, Teams channels, 30/60/90-day check-ins, and a welcome email.
Emailed receipts and vendor invoices are read, categorized, and filed to SharePoint and QuickBooks — no more shoebox at month-end.
One vendor promotion email becomes a marketing-calendar card, a newsletter draft, and signage copy — drafts for a person to review.
Pre-shift clock-in reminders for staff who opt in, plus late, no-show, and missed-punch tracking for the manager.
Sorts a departed employee's inbox: what matters is forwarded, the rest is archived. Classification runs on a local model.
Thirty minutes before each meeting: who's attending, the agenda, open items, and recent email with them. Read-only, no AI needed. Runs for the founder today.
Finds emails you sent that got no reply after three days, and drafts you forgot to send. It flags; you decide. Runs for the founder today.
Reviews overdue tasks, does the research and drafting, and writes up what it found. It never closes a task itself — the owner approves. Runs for the founder today.
The orchestration layer
The layer that keeps 61 agents coordinated.
Individual workflows are the easy part. This is what lets dozens of them share context, hand work to each other, and fail loudly. Every agent is inventoried and risk-classified, writes to outside systems go through an audit log, new agents ship with a kill switch, and nothing new can be added without an approved build record.
| Agent | Lane | What it produces |
|---|---|---|
| Brain | Information routing | Scans inbound Microsoft 365 information and routes summaries into the owner's operating memory, so every other agent works from current context. |
| Finance, Marketing & SEO snapshots | Shared context | Scheduled snapshots of financial, marketing, and search data that the rest of the system reads instead of each agent pulling its own. |
| Executive personas (CEO, COO, CMO, People) | Cross-agent handoffs | Hourly processing of handoffs addressed to each role, so an issue one agent finds reaches the lane that owns it. |
| CFO Watch | Financial anomalies | A weekly check for numbers that moved when they shouldn't have. |
| Salon Intelligence digest | Signal correlation | Collects findings from individual agents into one daily "what matters" summary instead of a dozen separate alerts. |
| Fleet Watchdog | Reliability | Watches every scheduled run; alerts once per new error type with a remediation runbook. |
| Model Router | AI cost and privacy | Chooses a model per task — local first, cloud when needed — and logs cost and tokens. 97% of calls run locally. |
Stop executing. Start governing.
The worst case: you do the mapping session and leave with a clearer picture of what's costing you — before spending anything on a build.
Start with an operations audit →