AI Models — model selection
Which model is responsible for the dialog logic (Summary / Router), and which one is responsible for the response text (Reply). Strategies (“all strong,” “strong router + lightweight reply,” “all ligh
Which model is responsible for the dialog logic (Summary / Router), and which one is responsible for the response text (Reply). Strategies (“all strong,” “strong router + lightweight reply,” “all lightweight”) and the Following instructions / creativity slider.
Settings screen — Automation & AI (
CRM → Automation & AI) → model selection block ( AI Models). This article is about selection strategy, not a list of providers: names and pricing change. The current catalog and prices are in the interface selectors.Tip
When to come here: after How to connect AI to a bot, when the basic
AI Reply is already working. If the bot “gets lost” in the flow or branches incorrectly, the cause is not always the prompt — check the router model.Two different tasks
Role in the UI (typically) | What it does | Symptoms of a weak model |
AI Summary and Router | Logic: transitions, Requirements, Routings triggering, consistency with the Goal | Wrong branch, looping, “forgets” the step context |
AI Reply | Final text for the lead | Dry, templated, not in the right tone, wrong length |
Tip
Reasoning (branches, conditions) and writing (wording) “fit” different models differently. Separate selection gives you an independent balance of quality and cost.
What the interface shows
AI Summary and Router selector — the model for logic.
AI Reply selector — the model for the response text.Tip
The Following instructions or creativity slider (0–1) affects Reply (wording variability while preserving the instruction).
Below, the strategies are described in terms of light / medium / strong — choose the specific names from the catalog.
Three strategies
All strong
Strong models for both Summary/Router and Reply.When: high average order value, a premium audience, complex flows with many AI Routings — semantic routing (CRM→ Flow → Steps → AI Routings — Deep Dive), B2B and sales, where every wording choice matters.Downside: usually the highest price per message and conversation.Strong Router + light (or medium) Reply
A strong model for logic, a cheaper one for text.When: standard replies with complex routing, a large lead flow, and the need to preserve branch accuracy without a “premium” literary layer in every reply.Downside: Reply may sound simpler. For a subtle premium tone, it may sometimes not be enough.Note: savings compared with the “all strong” strategy depend heavily on pricing and volume. Check estimates like “several times cheaper” against Billing — bot expenses, not against general articles.All light
Cost-effective models for both roles.When: tests, secondary bots, simple flows (few steps and routes), a tight budget.Downside: a higher risk of logic errors in complex funnels and more “flat” text. Be sure to review the first few dozen conversations after enabling this in Messenger.
Following instructions / creativity
The 0–1 slider applies to the Reply model (it does not replace Response Style — How the Bot Speaks).
Closer to 0 — more predictable and closer to the instruction template. Suitable for FAQ, support, and regulated responses.
Tip
0.5–0.7 — a compromise for most scenarios.
Closer to 1 — higher variation in phrasing and a greater risk of deviating from the instruction. Suitable for entertainment or a very lively tone with mild Restrictions.
Tip
If responses feel “copy-pasted” — raise the slider slightly. If the model “drifts” away from the policy — lower it and strengthen Restrictions / Response Style.
Tone and persona are set primarily in AI Settings — Bot Personality. The slider controls the variety of phrasing, not “intelligence”.
Where to view cost
Tip
Charges usually depend on the number of model calls and token volume (long Global Legend — who is the bot, conversation history, Reply Addition Info, a long response).
Where to check:
Billing — bot expenses — bot-level trends, average conversation cost.
Tip
A single lead message at the
AI Reply step typically involves at least one Reply pass and transition-checking logic (Requirements / Routings). With Recognition and additional calls, the total increases — see AI Recognition — media processing.Anti-patterns
Tip
Weak router in a complex funnel. Many steps and Routings with inexpensive logic create unstable branches and lower conversion — see Steps Conversion — funnel by flow steps.
Tip
Premium Reply for a fixed FAQ. If answers are short and standard, a lighter Reply model is often enough. Confirm the difference through billing.
Changing both models at once. If there is a regression, it is unclear which of the two is at fault. Change one at a time, observe 30–50 conversations, then change the second.
Adjusting only creativity for the sake of “liveliness.” Start with Response Style and examples in Restrictions. The slider is about variation in phrasing, not compliance.
How to test a model change
Tip
5–10 scenarios: a typical question, an objection, off-topic, media if available.
Tip
After 1–2 weeks — before/after metrics: step conversion, average cost, share of escalations to an operator — see How to improve AI quality.
A full-fledged A/B test of two bots with different models is an expensive operation. It makes sense for critical decisions.
Frequently asked questions
What’s next
Related
Was this article helpful?