AI for Franchise Networks: Central Standards, Local Unit Rules

AI for Franchise Networks: Central Standards, Local Unit Rules

AI for franchise networks works when the network separates two layers: a central standard for tone, policies and limits, and a data sheet for each unit with hours, prices and local rules. The agent answers in the brand's voice, checks the unit's sheet for everything that changes from store to store, and hands the conversation to a person when the data does not exist. Without that separation, the network ends up with inconsistent answers or wrong information about the unit.

Key takeaways

• The central standard defines how the agent speaks, what it promises, what it never does and when it passes the conversation to a person.
• Hours, prices, promotions and local rules live in a structured sheet for each unit, not inside the instruction text.
• Missing or expired data should trigger a handoff to the unit, never an invented answer or one copied from another store.
• Each sheet needs an owner, a review date and a change history.
• The safe path is to map, set limits, pilot in a few units and only then expand to the network.

Why a franchise is a different case for AI customer service

In a single-location company, the AI agent answers from one knowledge base. In a franchise network, the same question has different answers depending on where the customer is. What are the Saturday hours? Is the monthly promotion running? Can I reschedule at no cost? The correct answer changes by unit, and sometimes by day.

There is also a business tension. The franchisor wants consistency: the same tone, the same warranty policies, the same way of handling complaints. The franchisee wants autonomy over what is theirs: the prices they charge, staff schedules, local campaigns. A poorly designed AI agent breaks one side or the other. If everything is central, the customer gets wrong information about the unit. If everything is local, the brand loses its voice.

So the core problem is not choosing an AI model. It is deciding where each piece of information lives and who is accountable for it. Anyone who has built a support agent knows that the quality of the instructions for an AI support agent matters a great deal, but in a network there is a second half to the work: the data for each point of sale.

What stays central and what belongs to the unit

Here is a simple guide for classifying each type of information. For every item, ask: if this answer varied from unit to unit, would the brand suffer?

Stays in the central standard:

• Tone of voice, form of address and how the agent introduces itself.
• Complaint, cancellation and refund policy when it is contractual across the network.
• A list of topics the agent never resolves alone, such as billing disputes, legal questions or health matters.
• Privacy rules, the notice that the customer is talking to an AI and how to ask for a human.
• Handoff criteria: when, to whom and with what summary.

Belongs to the unit:

• Address, phone number, opening hours and local holidays.
• Prices actually charged, when the network allows free pricing, and local promotions.
• Services or products the unit does or does not offer.
• Its own operating rules, such as lateness tolerance, delivery fee by area or appointment requirements.
• The name and channel of the person who receives handoffs.

There is a gray zone: rules the network sets but the unit may adjust within a limit. One example is the maximum discount a manager can grant. In these cases, the central standard holds the limit and the unit's sheet holds the chosen value. The agent negotiates nothing beyond what is written.

How to structure each unit's sheet

The sheet is the heart of the setup. Instead of putting hours and prices inside the agent's instruction text, keep a structured database with the same fields for every unit. The agent receives the unit's identifier, looks up the matching sheet and only then answers. This is the same idea as retrieval at answer time, described in the AWS documentation on retrieval-augmented generation: the model does not have to memorize the data, it looks it up.

A checklist of fields that usually pays off:

• Identification: unit name, address, official channels and service area.
• Hours: regular opening hours, exceptions by date and special service.
• Local catalog: services or products available and what is not offered.
• Prices and terms: prices charged, accepted payment methods and the validity of each promotion.
• Local rules: scheduling, lateness, cancellation, pickup and delivery.
• Handoff contact: who receives the customer and during which hours.
• Governance: the sheet owner, the date of the last review and the date of the next one.

Two practical cautions. First, every temporary piece of information needs an expiry date. A promotion with no defined end becomes an error within weeks. Second, the sheet should be easy to update for the people working in the unit, through a simple form, not only for people who understand AI. If an update depends on a ticket to the technical team, the sheet goes stale.

When the channel is WhatsApp, there is also an architecture decision between one central number and one number per unit. Both can work with the same sheet, but they change how the unit is identified. The guide on how to integrate AI into your company's WhatsApp helps you understand these choices before you decide.

What the agent does when data is missing or conflicts

This is where most networks go wrong. The agent has a question, the sheet has no answer, and the model, by nature, tends to fill the gap with something plausible. The golden rule must be written into the central standard: if the information is not in the unit's sheet, the agent does not answer by inference, does not use data from another unit and does not quote values from memory. It says it will confirm with the unit and hands off the conversation with the context already summarized.

A few situations to handle explicitly:

• Empty field: the agent says it will confirm and alerts the unit's owner.
• Expired information: if the promotion's validity has passed, the agent treats it as nonexistent and hands off if there is doubt.
• Conflict between what the customer saw in an ad and the sheet: the agent follows the sheet and routes the case to a person to assess.
• Unit not identified: the agent asks for the city or neighborhood before giving any local data.
• Out-of-scope request: a serious complaint, a legal matter or a threat of public exposure goes straight to a human.

Hours also need attention. An agent that replies outside business hours needs to know what it can resolve right away and what it should leave for the unit the next day. The article on after-hours customer service with AI details this split, and in a network it depends on each location's schedule.

Limits on actions matter as much as limits on answers. If the agent can change an appointment or register an order, access should be restricted to the customer's own unit, with minimum permissions. The OWASP authorization cheat sheet reinforces the principle of least privilege and of checking access on every request, which applies directly here.

Illustrative scenario: a pet shop chain with different services per store

Imagine, purely as an example, a pet shop chain. The brand defines that the agent addresses the customer by name, uses warm language, never gives veterinary advice and always hands off when an animal seems to have a health problem. That is the central standard.

A customer asks on WhatsApp whether she can bring her dog for a bath on Saturday afternoon. The agent identifies the unit from the contact number used and checks the sheet. At this store, Saturday service is mornings only, and baths require an appointment. The agent replies that the unit is open on Saturdays only in the morning, offers the free slots the calendar shows and asks whether she wants to book.

At another store in the same chain, the answer would be different, because the sheet is different, but the tone and the safety rules would be the same. If the customer mentions that the dog has irritated skin, the agent offers no opinion, says it will pass the case to the store team and sends a summary to the owner. If the sheet does not list the price of a specific service, the agent does not guess: it says it will confirm.

What this scenario shows is that the customer perceives a single brand, while the factual information always comes from the local source.

Who is accountable for the sheet: governance between franchisor and franchisee

Technology solves half the problem. The other half is knowing who owns what. In a network, it is worth recording in writing:

• Who maintains the central standard and approves changes to tone and policies.
• Who updates each unit's sheet and how often.
• Who answers when the agent communicates something wrong because of an outdated sheet.
• How the franchisee can request an adjustment or switch the agent off in a specific case.

There is also the personal data layer. Customer conversations involve names, phone numbers and sometimes more sensitive information. The network needs to define, with legal support, which roles the franchisor and each franchisee take in processing that data under LGPD (Brazil's data protection law), how long conversations are stored and who can access the records. The ANPD, Brazil's data protection authority, is the official source for following guidance on the subject. If the AI model or platform sits outside the country, also read the piece on international data transfer under LGPD when using AI providers.

Another governance point is access to records. The franchisor may want to see the agent's performance across the whole network, while each franchisee needs to see only what happens in their own unit. Design these views from the start instead of opening everything to everyone.

How to start: map, limit, pilot, hand over

The method we apply in AI customer service projects follows four steps, and in a network each one gets its own detail.

Map. Survey how each unit handles customers today: channels, hours, most frequent questions, what each person answers and where the answers diverge. Record the starting point before changing anything, so you can compare later. This stage also reveals what the network already has standardized and what exists only in each manager's head.

Limit. Define what the agent may answer, which data it can access and which actions it can perform. In a network, this means choosing a few question types to begin with, such as hours, address and scheduling, and leaving out negotiated prices, complaints and any sensitive subject.

Pilot. Choose one real flow and a few units that differ from each other, to test the variation in rules. Keep records of the conversations and of the agent's decisions, and agree in advance which indicator defines success. The article on customer support KPIs with AI helps you choose between handling time, resolution and satisfaction. Also test with deliberately hard questions: wrong unit, empty field, expired promotion.

Hand over. Deliver the operation with the sheet, the records and the documented rules, and decide together whether to expand, adjust or stop. Expanding to the network only makes sense when the pilot units can keep their sheets up to date without outside help.

If the network also plans to send outbound messages, such as appointment reminders or campaigns, treat that as a second stage with its own consent rules. The guide on WhatsApp outbound messages covers templates and opt-in.

Frequently asked questions

Can a single AI agent serve every unit in the network? Yes, as long as it identifies the unit before answering and checks that unit's own sheet. What should not happen is mixing data from different stores in the same instruction text. In networks that differ a lot from one another, it is worth assessing whether one agent is enough or whether it makes sense to split by region or by unit type.

Can the franchisee change what the agent says? They can change the data that belongs to the unit, such as hours, services and local terms, within the sheet. The tone, the brand policies and the agent's limits stay in the central standard and change only by the franchisor's decision. This split should be described in the contract or in an operating document for the network.

What happens if a unit's sheet is out of date? The agent may state something wrong, and that is the main risk of the model. That is why every field needs a review date, an expiry for temporary information and someone responsible. If the data expires or is missing, the rule should be to hand off to the unit instead of answering.

Do I need a new system to store the sheets? Not always. A well-structured spreadsheet or the management system the network already uses can work, as long as the fields are the same for every unit, access is controlled and there is a change history. What matters is having a single source of truth, not several copies scattered around.

Sources

• AWS: What is retrieval-augmented generation
• OWASP: Authorization Cheat Sheet
• ANPD: Autoridade Nacional de Proteção de Dados
• Anthropic: Effective context engineering for AI agents