How do you make an AI sales agent stick to your playbook?
Not by writing a longer prompt. Toppp turns your sales playbook into stages, each with its own goal, tools, and knowledge range. Every transition is gated: fail the condition and the conversation does not move on. You draw the boundary; anything beyond it goes back to a human.
Drift has three shapes. One fix won’t do
Tell them apart first — each has its own cause and its own fix. In regulated work like insurance or aesthetics, the second is a compliance problem.
Improvising
Asked something the playbook never covered, it invents an answer — and a fluent one.
Over-promising
To move the deal, it offers a discount or a deadline nobody authorised. You find out later.
Skipping steps
It pitches before qualifying — or keeps qualifying a customer who is ready to buy.
Same goal, three ways to build it
Everyone wants the agent to follow your process. How it is built decides what holds.
- 01
How it’s expressed
Write a longer prompt: A paragraph of prose; the model takes what it takes / Draw a fixed flowchart: Fixed branches; it breaks at a fork nobody drew / Stage machine + guard rules: A chain of stage nodes: goal, script, available tools, next stage — each defined
- 02
Skipping ahead
Write a longer prompt: It follows the customer and drifts / Draw a fixed flowchart: Pushed back, or stuck / Stage machine + guard rules: It answers inside the stage; whether to advance is the guard’s call
- 03
Over-promising
Write a longer prompt: One line: “never offer a discount”. Held or not, turn by turn. / Draw a fixed flowchart: Only the branches you drew are covered / Stage machine + guard rules: What it must not say and must not call become machine-checked cases — no pass, no launch
- 04
Changing a rule
Write a longer prompt: Edited in place, live on running conversations, with no version to fall back to / Draw a fixed flowchart: One edit ripples through the map / Stage machine + guard rules: A new version, applied explicitly; reviews follow it, so you can roll back
- 05
Proving it held
Write a longer prompt: A transcript only — you see what was said, never why / Draw a fixed flowchart: The outcome again, never the reasoning / Stage machine + guard rules: Every turn’s reasoning, tool calls and latency replay, each assertion’s verdict logged
One stage, four fields, all yours
A stage is the smallest unit here. It is not a script — it is a node with a goal, a limit and an exit.
What this stage is for
What has to be learned here — budget range, decision maker, visit time. No goal, no move.
Script and instructions
How it talks here, which knowledge it may cite, which tools it may call — bound to this role.
Entry conditions
Guard rules decide whether the next stage opens. Deal size, discount range, what to confirm first — your call.
Where it goes
Conditions met, it moves on. Hit a line you drew, and the exit is a human.
Custom does not mean built from scratch. The platform is shared; how you cut stages, set goals and write conditions is your business logic.
Four gates, four moments
Staying in bounds is not one switch. It is four things at four moments — drop one and the rest get walked around.
No pass, no launch
01 / Before launch: What it must not say and must not call become test cases, judged one by one. A red-line case with no rule attached cannot be approved at all.
No transition until the condition is met
02 / At run time: Entry conditions apply mid-conversation. Fall short and it stays in the stage rather than skipping the confirmation. Sensitive-topic limits apply at this layer too.
Back to a human, invisibly
03 / Out of bounds: Switch one contact at a time to a human. The takeover happens backstage; to the customer it still feels like one advisor.
Replayable turn by turn
04 / Afterwards: Every turn’s reasoning, tool calls and latency replay, each assertion’s verdict logged. You never take the AI’s word for it.
Three reasons not to buy this
Guard rules cost something — they ask you to state your limits first. In these three cases, that cost buys nothing.
You only want a FAQ bot
You only want a FAQ bot / You only do one-way outreach
- Answer the question, end of job. Stages and entry conditions have nothing to do. Cheaper options exist.
- One fixed script, then hand it to a person. No stages to advance through — a dialler will do.
You have no playbook yet
You have no playbook yet
- Stages and limits come from a method you have already proven. No method, nothing to build from.
Five questions to ask any vendor
Skip the adjectives. Any vendor should be able to demonstrate the answer to these five on the spot.
How a stage is defined
Have them open one: goal, available tools, knowledge scope, exit condition — show me each field.
Which layer holds the red line
Where “must not say” lives, who judges it, how it is verified before launch. If the answer is the prompt, it is only the prompt.
After you change one rule
Live immediately, or applied explicitly? Can you roll back? Do review results follow the version?
The moment it goes out of bounds
Trigger a takeover and watch that handoff from the customer’s side.
Replay one conversation
Ask for a real replay: what it reasoned each turn, which tool it called, which assertions passed.
What people ask about guard rules
Won’t all those rules make it rigid?
Toppp constrains the transitions, not the talking. Inside a stage it answers questions, handles objections and speaks the way you told it to. Advancing is the guard’s decision, not a reward for a chat that happens to be going well.
We change models and configs — do the rules still hold?
The rules live in a config version, not in the model. Changing config saves a new version; it takes effect only when you apply it, and conversations already in flight switch to the new version the moment you do. Reviews are tied to the version, so you can compare and roll back.
Can it really stop an unauthorised price promise?
In two layers. Before launch, “must not say” and “must not call” are hard rules a machine checks one by one — no pass, no launch. After launch, limits like deal size sit on the stage’s entry conditions and the takeover boundary; you draw the line, and crossing it sends the conversation to a human. Rather than trust a “never”, ask to see a replay.
Bring your limits, see them become process
A demo takes about 30 minutes. Bring the list of things your AI must never say; we turn it into stages and entry conditions on the spot, then replay a conversation.
Or reach us directly business@toppp.ai