How the agents actually learn a brand.
Every BPO owner has been shown an “AI” that turned out to be a chatbot with a script. This is the answer to that scepticism: the seven stages between a franchisor’s rulebook and a fleet allowed to answer its franchisees.
Rules in code. Brand knowledge in an index. Judgement in a human.
The fleet does not run on a model's general sense of how franchising works. It runs on three separate layers, and the split is the whole point.
Rules — approval ceilings, what may never be quoted, escalation triggers, SLA windows — live in code, in the action handler. A model cannot be talked past them, because it is not the thing enforcing them.
Brand knowledge — the operations manual, the SOP library, the approved asset set, historical franchisee tickets, precedent decisions — lives in a retrieval index scoped to your workspace alone. Agents answer from that brand's own indexed documents, not from anything the model happens to recall about franchising in general.
Judgement — anything ambiguous, novel, or touching contract, territory, royalty or legal exposure — goes to your supervisor. Always, by design, not as a fallback.
Seven stages from rulebook to production.
Brand capture Day 1–4
The franchisor's operating system is written into a machine-readable spec: the SOP library, escalation policy, approval ceilings, supply and reorder rules, audit checklists and scoring, training curriculum, brand tone and the approved asset list. Every rule is traced to a client document or a signed-off decision. Anything the brand cannot answer is logged as an open item rather than guessed — and in franchise work there are always some, usually around territory.
Knowledge base build Day 4–9
Operations manual, outlet register, product and supply catalogue, historical franchisee tickets, the macro library, past audit findings and precedent decisions are indexed into a retrieval layer scoped to your workspace. This is what gives an agent brand-specific competence on day one instead of month six — it has already read every ticket the network ever raised.
Agent configuration Day 6–12
Each agent gets four things: a scope defining what it may touch, tool bindings defining which panel actions and client systems it may call, an autonomy level deciding whether it acts or proposes, and its encoded hard limits and escalation triggers. An agent physically cannot call a tool it was never bound to. Permissions are the security model, not instructions.
Sandbox evaluation Day 9–16
Every agent runs against replayed historical franchisee tickets with known correct outcomes, drawn from the networks it will serve. Accuracy is measured per query type — supply, policy, systems, training, escalation — not as one headline number. No agent leaves sandbox until it clears its threshold on its own evaluation set, and every failure is traced to a missing knowledge item or a badly stated rule and fixed at source.
Shadow mode Day 16–21
The fleet runs on live franchisee traffic and sends nothing. Every output queues for review. Your supervisor compares what the agent produced against what they would have said, and every disagreement goes back into the rules or the index. This is where your team learns the fleet's behaviour and the fleet learns the brand's voice — at zero risk to the franchisor.
Graduated release Day 21–24
Autonomy is released one query type at a time, lowest risk first: dispatch tracking, training reminders, audit scheduling. Draft-only never comes off anything touching contract, territory, royalty, brand publication or a legal position. Each release is a decision with a name against it, recorded in the workspace log.
Continuous retraining Ongoing
QC re-checks 200 random decisions every week. Every human overrule is logged with a reason and reviewed for pattern. When the franchisor changes an SOP, a price list, a supply rule or a campaign, it is a versioned update to the spec, pushed to the fleet and re-evaluated before it takes effect. Every decision and its inputs are retained 24 months.
Two classes, and the line between them never moves.
Supervised Acts, logs, gets sampled
Supply AI, Compliance AI, Training AI, Field AI, Insight AI. These execute inside encoded limits and record every decision with its inputs. Nothing they do is irreversible, and everything is visible in the task centre as it happens.
Draft only Proposes, a human releases
Helpdesk AI, Onboarding AI, Lead AI, Marketing AI, Billing AI. These produce a proposed reply or action and stop. Nothing reaches a franchisee until a named person in your workspace releases it. This class does not graduate, however good the accuracy gets.
Enforced in code, not by instruction
The distinction matters more than anything else on this page. An instruction in a prompt is a request. A check in the action handler is a refusal. These are the second kind.
- Never approve, renew or terminate a franchise agreement
- Never quote or vary franchise fees, royalty rates, territory boundaries or exclusivity
- Never issue a credit note or waive a royalty above ₹10,000 without named approval
- Never publish brand or marketing material outside the approved asset library
- Never disclose one franchisee's commercial data to another
- Never send a reply referencing legal action, notice or default
- Never delete or overwrite a decision log — the log is append-only, retained 24 months
What it does badly.
Any vendor who will not answer this is selling you something.
- Novel situations with no precedent in the network's history. It flags rather than invents — but it flags a lot in month one.
- Judgement calls where the brand's own manual is silent or self-contradictory. There will be some every week, and they land in your queue.
- An angry franchisee who needs a person and a phone call. Escalation is a design feature, not a failure.
- Anything commercial on the franchisor's behalf. It has no authority and is not given any.
- Regional and language nuance beyond the channels and languages configured at setup.
See it running rather than described.
Fifteen minutes on a live workspace: the fleet working, the review queue, and what happens when an agent is asked a territory question and refuses.