Reference build · Fitness franchises · Reporting

A review-pack agent for a 50-studio fitness franchise

An agent that pulls every studio’s numbers from member, billing and scheduling systems, applies one written KPI dictionary, runs the two integrity checks that catch most scorecard errors, and writes a review pack that leads with what moved.

Reference build — a working agent and a worked example, not a client deployment

01 · The problem

Two area managers pull "churn" for the same studio and get two numbers. The review becomes an argument about the spreadsheet.

In a franchise with fifty studios and five area managers, each manager spends about four hours assembling data for each of five weekly reviews: exports from the CRM, the member system and billing, pasted into a template, reconciled by hand when they disagree, which they do. That is half of every manager’s week before any coaching happens, and about 4,800 hours a year across the network.

The hours are the visible cost. The invisible one is that conversion is counted period-based in one report and cohort-based in another, so a studio can show trial-to-member conversion above 100%; that members on open-ended freezes sit in the active base and understate churn; and that two studios with near-identical numbers get different priorities depending on which manager had which export open.

Constraints that shaped the design

  • Automating assembly on top of disagreeing definitions produces wrong answers faster. The KPI dictionary comes first and is owned by people.
  • The agent reads systems of record and writes only the pack and the commitments log.
  • Studios under six months are measured against a ramp curve, not mature benchmarks.
  • Nothing reaches a franchisee without an area manager approving it.
02 · Agent design

A scheduled agent with a dictionary, two integrity checks and a memory of what was agreed.

  1. 1The night before a review cycle the agent runs for each studio in the week’s schedule.
  2. 2It pulls leads from the CRM, joins, freezes and attendance from the member system, revenue from billing, and class utilisation from scheduling, for the review window and the comparison window.
  3. 3It applies the KPI dictionary: one formula per metric, one source of truth, the cohort window for conversion, the freeze rule for active members, the ramp curve for new studios.
  4. 4It runs the integrity checks: any conversion above 100%, any freeze with no end date, revenue in the member system more than 1% from revenue in accounts. Failures are flagged on the pack, not silently corrected.
  5. 5It reads the last review’s commitments and checks each one: done, overdue, and whether the metric it was meant to move actually moved.
  6. 6It writes the pack: exceptions first (what moved more than the benchmark allows), then the commitments, then the full table with every number linked to its source rows. The area manager reads it in the morning.

Why an agent, and not a chatbot or a Zap: A BI dashboard shows numbers once someone has decided what they mean and loaded them. The work here is everything before that (reconciling sources that disagree, applying the ramp curve, catching the 104%) and everything after (remembering what was agreed). A dashboard does neither, and a scheduled SQL job cannot write "revenue disagrees with finance by 4%, here are the rows" in a sentence a manager will read.

03 · Tools

The 6 tools it may call, and nothing else

The tool list is the permission list. If a capability is not here, the agent does not have it, however it is asked.

ToolWhat it doesWhat it may touch
fetch_crm_leadsLeads and trials by studio and windowReads the CRM
fetch_member_metricsJoins, cancellations, freezes, attendance by studioReads the member management system
fetch_billing_revenueRevenue by studio and month from billing and from accountsReads billing and the accounting system
apply_kpi_dictionaryComputes each KPI by its written definition and ramp stageReads the versioned dictionary
run_integrity_checksFlags conversion above 100%, open-ended freezes, revenue mismatchesWrites flags to the pack
write_review_packWrites the pack and updates the commitments logWrites to the pack store and commitments table only

Model and orchestration

  • Claude (Anthropic API) with tool use
  • Supabase Edge Functions (Deno) for the agent loop
  • Postgres for memory, logs and the approval queue

Sources

  • Mindbody, Glofox, ABC Fitness or Perfect Gym
  • HubSpot or Salesforce
  • Stripe, GoCardless, Xero, QuickBooks or Tally

Outputs

  • A pack per studio (PDF and web), a commitments table, a drift ticket queue
04 · Guardrails

What keeps it safe to run on real data

Guardrails

  • Read-only connections to every source system. The only writes are the pack and the commitments log.
  • Every figure in a pack links to the rows it was computed from.
  • Integrity failures are shown, never auto-corrected: the pack says "this number cannot be trusted" rather than guessing a better one.
  • Draft follow-ups to franchisees wait in an approval queue for the area manager.
  • The dictionary is versioned; a pack records which version it used, so last quarter’s red and this quarter’s green are comparable.
05 · Result

The illustrative 50-studio arithmetic, which is the same one the Franchise KPI Trust Check runs on your inputs.

What the build does

  • Assembles a pack per studio from four source systems on a schedule, on one set of definitions
  • Flags the two most common scorecard errors before anyone reads the number
  • Measures new studios against a ramp curve so month three is not red on everything
  • Carries last cycle’s commitments into this one with their status
  • Leads with exceptions, so a manager with ten studios reads ten pages, not ten spreadsheets
Assembly hours across the network, per year4,800Baseline
Fully loaded cost of that timeUSD 216,000 a year (INR 24,00,000 at INR 500/hour)Baseline
Prep time reduction60% or more against the measured baselineTarget
Studios per area manager without added headcount10 → 15+Target

Assumptions

  • 50 studios, 10 per area manager, 5 reviews a week, 4 hours of assembly each, 48 weeks, USD 45 (INR 500) an hour
  • Targets are pilot targets from the Data Trust Diagnostic, not guarantees

What a pilot measures that this page cannot

  • Hours per review before and after, by manager
  • Count of integrity flags per cycle, and how many were real errors
  • Whether two managers given the same pack name the same top priority
  • Commitments closed on time, by studio

Want this agent for your fitness franchise?

A pilot runs it on your data in two to three weeks for a fixed fee, measured against your own baseline. Thirty minutes to scope it, free.

Other case studies