Flow charts break the week your prices change
Keyword trees are cheap to build and expensive to maintain. What replaces them, and what you lose in control.
Decision-tree bots are genuinely appealing on day one. You can see the whole thing. You can point at a box and say “that is what it says when someone asks about parking.” Nothing is mysterious.
The problem is not day one. It is the Tuesday six weeks later when you change your prices.
The maintenance shape nobody budgets for
A flow chart encodes your business as structure. Every price, hour and policy is baked into a node, and every node has edges. Change one fact and you are not editing one place — you are hunting for every branch that mentioned it.
This produces a characteristic failure: the bot is not wrong everywhere. It is wrong in the two branches you forgot, which are the ones you look at least often, which is why nobody notices for a month. By then it has quoted the old price to a few hundred people.
The cost curve is the opposite of what you expect. Cheap to build, cheap for a month, then permanently expensive — and the expense is not build time, it is the risk of being confidently out of date.
What replaces it
The alternative is to stop encoding facts as structure and start treating them as knowledge the agent reads at answer time. Your prices live in one place. Your hours live in one place. When they change, they change once.
The agent then does the part the flow chart was badly approximating: reading what was actually asked, and composing an answer from current facts.
What you genuinely give up
This is the part most write-ups skip, so let us be direct. Moving off a decision tree costs you two real things:
1. Exhaustive predictability. With a tree you can enumerate every possible output. You cannot do that any more. What you get instead is bounded behaviour — sourced answers, explicit refusals, and escalation — which is a different and weaker guarantee. For most businesses it is the right trade. For some regulated ones it is not, and they should know that going in.
2. The visual artefact. The chart was also documentation. Losing it means you need somewhere else that plainly states how your business answers things, or the knowledge quietly disperses into people’s heads again.
The control you keep
What you do not give up, and should insist on:
- Hard rules that always win. Thresholds, refund limits, anything you never want handled automatically. These stay deterministic. They are not suggestions to the model.
- Escalation paths. Defined, not emergent.
- A refusal that is real. “I don’t know, I’m passing this to someone” has to be an available and frequently-taken action, or you have traded a stale answer for an invented one, which is worse.
The practical test
Ask what happens when you raise a price by ten percent.
If the answer involves opening a builder and finding every node that mentioned the old number, you have bought a maintenance liability. If the answer is “change it in one place and it is live,” you have bought a system.
The week your prices change is the week you find out which one you have.