Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Superintelligence for Franchises: One Standard, Many Locations

How franchise systems define shared prompts, memory and approval rules so each location's AI agent improves from one company-level standard.

Be Human typographic cover with the title: Superintelligence for Franchises: One Standard, Many Locations
Conceptual illustration created with AI · Be Human

Superintelligence for franchises means one company-level standard for how AI agents are configured, what they remember, which tools they can use and when a person must approve an action, with each location's agent running inside that standard. The franchisor owns the shared prompts, memory structure and approval rules. Each location supplies its own customers, staff and local facts, and an improvement made at one site is reviewed once before it reaches the rest of the network.

Key takeaways

  • Franchise AI works when the franchisor defines the standard once and locations adapt only the parts the standard allows.
  • Split every agent asset into three groups: what travels to all locations, what stays local, and what nobody changes without review.
  • Shared memory needs a named owner, a schema and a correction path, or one site's error spreads faster than its good ideas.
  • Approval rules should mirror the franchise agreement, not the default settings of a tool or vendor.
  • Measure each location against the same success criterion, so a result at one site means the same thing at another.

What a franchise standard actually contains

A standard is not a long manual that nobody opens. For an AI agent, it is a small set of versioned artifacts: the instructions the agent follows, the definitions of the facts it may remember, the list of tools it can call, the permissions attached to each tool, and the criteria used to judge whether it did its job. Keep these in one place the franchisor controls, with a version number and a date of last review.

A practical test: if a new location opened next week, could its agent be set up from these artifacts plus that location's own records? If the answer depends on a manager remembering how things were done, the standard is not written yet.

The standard should also state what each agent is for. A role such as answering first-line booking questions and drafting follow-ups for staff approval is specific enough to test. Helping customers is not. For a broader view of how this layer fits with the rest of a company's systems, the portfolio companies operating layer guide describes the same idea of standardizing once and owning the local details.

Separate what travels from what stays local

The most common mistake is treating every location as either identical or fully independent. Neither works. Divide the agent assets into three layers before anyone configures a single location.

  • Travels everywhere: brand voice rules, the definition of a qualified lead, escalation categories, cancellation and refund policy as the franchise agreement states it, and the tool catalog.
  • Stays local: opening hours, staff names, local supplier contacts, regional language variants, neighborhood promotions and the history of local customers.
  • Locked centrally: anything that touches money, customer data exports, legal wording or prices. Locations may request a change, but they do not edit these items directly.

Write the three lists down and attach them to the standard. When a location asks for a change, the first question is which layer the change belongs to. Most disputes disappear once that question has an answer. The same split helps when a business has two sites with separate permissions, as described in the two-location guide.

Shared memory needs one owner and one correction path

Memory is where a franchise agent either compounds value or spreads mistakes. If one location teaches its agent that a product is discontinued and nobody checks the note, every other location inherits the error. Treat memory as a shared record with an owner, not as a notebook each agent keeps alone.

Define a short schema. Each memory entry should state what it contains, who may write it, where its source is recorded and when it should be reviewed. Entries that describe policy should point to the current policy document rather than restating it. Entries about a local fact should carry the location identifier so they never appear in another site's answers.

Retrieval design matters as well. When an agent pulls the current document into its context at answer time, the company updates one source and every agent that reads it sees the change. That pattern is safer than letting each location copy policy text into its own prompts, where copies drift apart over months. Give one named person the job of reviewing flagged entries each week. A location can flag a wrong entry, but it cannot overwrite a shared one.

Approval rules follow the franchise agreement

Approval thresholds are where a standard meets legal and financial reality. Start from the franchise agreement and the operations manual. If the agreement requires franchisor consent for a discount above a stated level, the agent must not offer that discount without a human approval step, whatever the vendor's defaults say.

Write approval rules in plain language inside the standard, with three things stated for each action: the action itself, the condition that triggers approval, and the person or role who approves it. Then translate those rules into permissions the tools actually enforce. Each agent should hold only the access its role needs. The OWASP authorization guidance explains why this check belongs on the server side, where the tool decides, rather than in the agent's own instructions.

Protect the instructions too. Text inside a customer message or a retrieved document can try to override what the agent was told to do. The OWASP prompt injection cheat sheet describes why tools should validate requests independently of the model's output. A franchise agent that can issue refunds must check the refund rule in code. A line in the prompt is not enough. Changes to any of these rules should follow the same path described in how to request and approve changes to a managed AI agent.

Measure every location against the same criteria

A standard is only as good as its test. Choose one success criterion per role and define it precisely enough that two managers reading the same transcript would score it the same way. For a booking assistant, that might be: the request is resolved without staff involvement, or it is handed to a named staff queue with the details the staff member needs.

Build a short evaluation set before go-live. Include routine requests, edge cases and a few deliberately awkward messages. Add to the set every time a real case goes wrong. Anthropic's guidance on evals for AI agents recommends building test tasks from real failures, which is a sound habit to apply across a whole network.

Illustrative scenario (not a real example): a franchisor with a handful of locations notices that one site's agent answers cancellation questions faster than the others. Before copying its wording to the network, the operations lead reads the transcripts and finds that the faster answers skip a confirmation step the policy requires. That version is not adopted as the template. The confirmation step is added to the standard, and every location is retested against the same evaluation set.

Use one dashboard for all locations. Useful views include escalations to staff, approval wait times, corrections made to memory entries and cases that failed the evaluation set. When the numbers sit in the same place, differences between sites start a useful conversation instead of an argument about whose site is right.

A rollout plan for the first standard

Do not attempt the whole network at once. A workable sequence looks like this:

  • Name one role and one success criterion, such as first-line booking questions.
  • Write the three-layer split: what travels, what stays local, and what is locked.
  • Draft the shared instructions, memory schema and tool list in one versioned document.
  • Set approval thresholds from the franchise agreement and confirm them with whoever signs and enforces it.
  • Pilot at one or two locations with a named manager who reviews transcripts every week.
  • Build the evaluation set from the pilot results, then freeze version one of the standard.
  • Roll out to the remaining locations by loading local facts through the memory schema, not by editing shared prompts.
  • Review permissions and memory quarterly, and route every change through the same request and approval path.

Be Human designs and operates this layer for companies that want it under their own control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review and controlled improvements, so the franchisor keeps ownership of the standard while each location receives updates through one governed process.

If you want to see which task in your network is the right first candidate for an agent, take the free Be Human AI diagnostic to find it.

Frequently asked questions

Can every franchise location run its own AI agent? Yes, and that is usually the right design, but each location's agent should be an instance of the same standard rather than a separate build. The shared instructions, memory schema, tool catalog and approval rules stay the same, while local facts are loaded into each agent's own memory. This keeps the network consistent without forcing every site to share the same local details.

Who decides when a franchisee disagrees with a standard? The franchisor owns the standard, so disagreements go through the change request path rather than local edits. A franchisee can submit a proposed change with evidence from its own transcripts, and the franchisor decides whether it belongs in the shared layer, the local layer or nowhere. Recording the reason for each decision prevents the same debate from returning every quarter.

How often should the franchisor review the standard? Review the shared layer on a fixed schedule, typically quarterly, and review it sooner when the franchise agreement, pricing policy or a legal requirement changes. Permissions and memory entries deserve the same rhythm, because stale access and outdated facts cause more trouble than new features do. Each review should end with a dated version of the standard and a short note on what changed.

Does this replace the franchise operations manual? No, the standard complements the manual rather than replacing it. The manual describes how people run the business, while the standard describes what the AI agents are allowed to say, remember, touch and escalate. Where the two disagree, the agreement and the manual are the source of truth, and the standard should be corrected to match them.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →