Before you buy a brain, audit what it would be reading

Seven on a Saturday night. A shift lead is at the front desk and a member wants to freeze her membership mid-term. There’s a fee attached, maybe. The member is standing right there, waiting.

The answer exists somewhere. It’s in the operations manual, or it’s in a training video from the last convention, or it’s in an email a franchise business consultant (the field rep who supports a cluster of units) sent last spring. The shift lead has about forty seconds. So she guesses, or she texts the owner, or she waives the fee and moves on.

That scene is sitting underneath the franchise AI pitch I keep running into lately: franchises don’t have a technology problem, they have a memory problem. The answer to the 7pm question already lives in the network. In a binder nobody updated. On a recorded call nobody rewatched. There’s just no way to find it.

I think that’s a genuinely good frame. I’ve owned the units where that text comes in.

There’s a second observation that usually travels with it, and it’s the sharper of the two: the suppliers who struggle in this category almost always nail the sales pitch and miss the adoption curve. Eighteen months in, the franchisor is paying for licenses the units don’t open.

Both of those are right. Both of them also assume the same thing, and that assumption is what I’d want to check before writing a check.

They assume the knowledge exists.

Go look for it

Pick a question. Not a hard one. Membership freezes, refund policy on a damaged product, what the unit does when a customer shows up holding a promo from a campaign that ended two weeks ago.

Now go find every answer to that question that exists in your system today.

Here’s what I find, day in and day out, in brands between 50 and 300 units. There’s an answer in the operations manual, and it’s three years old. There’s a different answer in the learning management system, because the training module got rebuilt after the manual was written and nobody went back to reconcile the two. There’s a third answer in a field email, which is usually the most current and the least discoverable thing in the whole company. And then there’s what the unit actually does, which nobody wrote down anywhere.

And that last one is usually the right answer. The person doing it has hit that situation two hundred times. The person who wrote the manual hit it zero. So the most accurate knowledge in the system is also the one version living nowhere a machine can read it, and no amount of ingestion fixes that.

Four answers. Three of them written, in three systems, disagreeing.

That’s not a memory problem. Memory implies the thing was known once and then got lost. This is a different animal. A large share of the operating knowledge in a franchise system was never written down at all, and the parts that were written down have been drifting apart from each other for years, quietly, because nothing in the org is responsible for making them agree.

What a brain does with a corpus like that

It answers. Immediately, in a full sentence, with no hedge.

That’s the whole product. Nobody buys one of these to get back “I found four conflicting sources, please escalate.” They buy it because the shift lead gets one answer in forty seconds.

So the tool picks. And when the corpus disagrees with itself, that pick is arbitrary in a way the reader has no way to see. Maybe the manual wins because it’s longer. Maybe the training module wins because it’s more recent. Doesn’t much matter which. What matters is that the network’s internal disagreements now have a confident voice and an unlimited distribution list, and they’re being read by the newest people in the system, who have the least ability to know when the answer is wrong.

You didn’t surface the network’s collective experience. You industrialized its inconsistency.

This is also where the adoption-curve point stops being a separate observation and becomes the tell. Franchisees don’t ignore these tools because they can’t find them. Adoption dies because the tool hands the unit something the unit knows is wrong. An owner who gets one confidently incorrect answer about a fee will never trust the thing again, and he will say so to the owner two territories over at the next regional meeting. Same trust math that kills reporting rollouts, one layer up.

What I’d do first

This is a thought starter, not a project plan.

Pull the twenty questions your field actually gets. Your franchise business consultants know them cold, and so does whoever answers the support inbox. Don’t theorize the list. Ask.

Then, for each one, go count. How many answers exist in the system today: zero, one, or more than one? And where it’s more than one, do they agree?

Three buckets, twenty questions. A morning of somebody’s time, two at the outside.

That count is the grade on your corpus, and it’s the only honest input to the buying decision. If most of the twenty come back with zero answers, or with answers that fight each other, the shared brain isn’t the project. Getting to one answer is the project. It’s mostly editorial work rather than technical work, which is exactly why it never gets funded, because it doesn’t demo.

Buy the brain second. It’ll work better, and you’ll know why it works.

Run the twenty before you sit through another demo. If the buckets come back split, you’ve just found a cheaper project than the one you were about to buy, and you found it in a morning.