A customer asks the same question at two of your sites and gets two answers. One says the policy is X, the other says Y, and both agents are confident. Neither is lying: one site learned the rule a year ago, the other learned it from a colleague who had changed it slightly. The customer sees one company. You see eleven small ones.
This happens in hotels with several properties, regional retail and service chains, telecom stores, schools with more than one campus, and support teams split across offices and shifts. This guide covers how to decide what should be identical everywhere, what should legitimately differ, and how to find out when the two have drifted apart.
Separate three kinds of answer
Most inconsistency comes from treating every answer as the same kind of thing. Sort them first.
- Company-wide answers. Policies and facts that are true everywhere: how a complaint is escalated, what the company will and will not do, what an agent may promise. These must match to the word, and they need a single owner.
- Local facts. Hours, addresses, parking, which services a site offers, who to ask for. These differ by design. The failure here is not variation but staleness: the website says Saturday hours that the site stopped keeping in the spring.
- Local judgment within limits. Exceptions, goodwill gestures, scheduling around a local constraint. Here the company sets the range and the site picks a point inside it.
Write the three lists down for your top twenty questions. The exercise alone usually turns up a few answers that everyone assumed were company-wide and are actually decided site by site.
One source of truth, with site-level fields
The pattern that holds up is a central knowledge source that carries the company-wide answer once, and a small set of named local fields that each site fills in: hours, contact route, local services, local exceptions. An agent or an AI answering a question about delivery windows reads the shared policy and then the local field, rather than relying on memory of either.
Two practical rules keep it working:
- Every local field has a named owner at the site and a review date. A site manager who is on the hook for six fields keeps them current. A site manager who "can edit the wiki" does not.
- Local content may fill a field but may not rewrite the shared answer. If a site thinks the shared answer is wrong, there is a route to propose a change, and until it lands the shared answer stands. Without this, the central source slowly turns into a collection of local overrides.
This is the same principle as in keeping answers consistent across channels: one place wins when sources disagree, and everyone knows which.
Say who may decide what
A lot of what looks like inconsistency is really unclear authority. Two agents give different answers to the same exception request because neither knows whether they are allowed to say yes.
Write a short decision grid for the situations that come up most:
- What an agent may do alone.
- What needs a shift lead or site manager.
- What goes to the central team, and how fast they answer.
- What nobody may promise under any circumstances.
Keep it to one page and put it where agents already work. The aim is that two people, at two sites, in the same situation, either both say yes or both hand it to someone who can. The escalation side of this is covered in writing a customer support escalation policy.
Train from real conversations, not from the policy
A policy document tells people what the rule is. It does not show them what the conversation sounds like when a customer is annoyed and the rule is inconvenient. Sites diverge most at exactly that point.
- Pull five or six real exchanges per month that went well and five or six that did not. Remove names and details.
- Review them in a short standing session, with people from several sites in the same room or call, so they hear how colleagues handle the same moment.
- Turn what you learn into a line or two of example wording, not a longer policy.
Do not rely on a one-off training day when a policy changes. Changes need a short written note that goes to every site on the same day, and a way for a site to confirm it has read it. A tick in a tracker is enough; the point is that you can see who has not.
Check for drift on a schedule
Consistency decays quietly, so build the check into the calendar instead of waiting for a complaint.
- Run a mystery-question round every month. Pick five questions. Ask each site the same five, by phone, chat or email, as a customer would, and record what each answers.
- Compare the answers against the shared source. Mark each as correct, outdated, or invented.
- Share the results by question, not by site. "Three sites gave an outdated answer to the transfer question" is a content problem to fix. "Site B is wrong" starts an argument and teaches people to hide.
- Fix the source first. If an answer was wrong because the shared text was unclear, rewriting the text fixes every site at once.
A related habit: keep a "changed this week" note with the policies, hours or prices that moved. When a customer says "your other site gave a different answer," you can check whether it was a change that had not reached everyone yet.
What not to centralize
Do not try to make every site sound identical. Customers do not need the same greeting at a campus in one city and a store in another; they need the same facts and the same fairness. A script that every site must read verbatim tends to make agents stop listening to the customer, and it breaks the first time a conversation goes off script.
Also be careful about centralizing anything that depends on local knowledge nobody at head office has. Road closures, a lift out of order, a local event that doubles the foot traffic: these belong to the site, and the system should make it easy for the site to publish them quickly instead of waiting for approval.
Where RyteDesk fits
If your sites answer the same questions all day, some of those answers can come from one place instead of from memory. RyteDesk is an AI front desk in early access, built for mid-market support teams. It answers on chat, email and phone from the knowledge and policies you give it, tells callers it is an AI, and hands off to your team or takes a message when it cannot help. You test it in a preview that sends nothing to customers and can read every conversation, which makes it easy to check the same questions the way you would in a monthly mystery-question round. Join the waitlist if you want to try it; it is free to join.