Kommo Salesbot is the automation builder that ships inside Kommo CRM: it runs pre-defined, template-based scenarios — sending messages, creating tasks, changing deal fields, reassigning responsible managers — triggered by events in your pipeline or by customer messages. It is not artificial intelligence: it never invents an answer, it follows the script you drew. That single fact is both its superpower and its ceiling, and it is the key to choosing correctly between Salesbot, a simple AI agent, and a full orchestrated AI system.
This choice now confronts every team on Kommo, and most comparisons you will find are written by someone selling one of the options. We build all three tiers for clients, so we have no horse in this race — our interest is that you buy the cheapest thing that actually solves your problem. Sometimes that is the built-in bot. Sometimes it genuinely is the $20,000 system. Here is how to tell.
What does Kommo Salesbot actually do — beyond chat?
Here is the part almost every comparison misses: in real Kommo accounts, Salesbot spends much of its life nowhere near a customer conversation. Because it can react to pipeline events and edit anything in the deal card, operators use it as a general-purpose automation glue — dozens of small, unglamorous, money-saving routines:
- Reassigning deals between shifts. The evening manager logs off, the bot quietly moves their open conversations to whoever works nights.
- Creating follow-up tasks. Deal enters «Quote sent» — the bot schedules a call task for two days later, inside working hours.
- Updating fields. Customer clicks a button in chat — the bot writes the answer into a deal field your reports can actually filter by.
- Routing by source. Lead came from Instagram — tag it, assign it to the rep who handles social, fire the right greeting.
These template jobs are where Salesbot is at its best: the logic never changes, the bot never gets tired, and the cost is zero on top of your plan. If your «automation problem» looks like the list above, you do not need AI at all — you need an afternoon with the Salesbot builder.
Where does Salesbot fall short in conversations?
Now the uncomfortable part, which the vendor's own materials will not tell you directly: in live customer chat, Salesbot is often clumsy. It matches incoming messages against the patterns you predicted in advance. Customers, unhelpfully, do not read your script:
- It misreads questions. «Do you deliver to Austin and how much would two units cost?» contains two intents. A template bot picks one — or neither — and answers something adjacent. The customer repeats themselves, gets the same answer, and asks for a human with declining patience.
- It cannot rephrase. Every answer is verbatim from your script. Ask the same thing twice and you get the identical paragraph twice — which reads as exactly what it is, a machine wall.
- Typos and slang break matching. Real customers write «hw mcuh» and «cuanto sale». Keyword triggers miss, the fallback branch fires, and the dialog turns into «I didn't understand that» loops.
There are also hard platform limits worth knowing before you architect anything serious on it: on the Base plan you can build a bot but not launch it; a bot is capped at 100 actions; only one bot runs in a conversation at a time, and it stops when the conversation closes; and on WhatsApp, Meta's 24-hour window means the bot cannot initiate messages a day after the customer's last reply.
None of this makes Salesbot bad. It makes it a script player — superb when the conversation has rails, weak when it does not. The question is what to do when it does not.
What is a simple AI agent — and when is one prompt enough?
A simple AI agent replaces the rigid script with a language model and one carefully written instruction — a single prompt that describes your business, your products, your tone, and what the bot is allowed to do. The customer writes in their own words; the model actually understands «hw mcuh» and two-part questions; the answer is composed fresh each time.
What surprises most executives: this tier does not require any infrastructure of your own. It runs inside Kommo — configured correctly, with a knowledge base and a clear handoff rule for when a human takes over. Setup is measured in days, not months.
Strengths and weaknesses, honestly:
| Simple AI agent — pros | — cons |
|---|---|
| Understands natural language, typos, mixed questions | One prompt covers every situation — quality floats with question complexity |
| Set up in about a week, no servers | Can confidently say wrong things (hallucinate) when asked outside its knowledge |
| Cheap to run — roughly a few cents per dialog | Cannot perform complex operations: no database lookups, no document generation |
| Easy to update — edit the prompt, not a flowchart | Needs a well-written knowledge base, or it improvises |
The single-prompt design is the defining trade-off. One instruction handles the greeting, the pricing question, the complaint and the weird edge case — which works remarkably well right up until your dialogs get genuinely diverse, and then quality starts to wobble exactly on the conversations that matter most.
What is an orchestrated AI agent?
The orchestrated tier solves that wobble with division of labor. Instead of one prompt trying to be everything, there is an orchestrator — a coordinating model that reads each incoming message, decides what kind of question it is, and routes it to a specialist agent: one expert in pricing, one in delivery terms, one in complaints, one in documents. Each specialist gets a prompt generated on the fly for that exact situation, answers within its narrow expertise, and the orchestrator assembles the reply.
Two things change qualitatively at this tier:
- Accuracy stops floating. A narrow expert with a situation-specific prompt is dramatically harder to confuse than one generalist prompt. For industries where a wrong answer costs real money — banking, insurance, anything regulated — this is the difference between «nice toy» and «production system».
- The bot gains hands, not just a mouth. An orchestrated system can perform operations: check stock in your database, verify a client against your records, generate and send an invoice, book a slot in an external calendar. It does not just talk about your business — it operates inside it.
The price of this is paid twice. Once in development — this is real software engineering, not configuration. And once per dialog: where a simple agent makes one model call per answer, an orchestrated system makes three to ten (orchestrator, specialists, final assembly), so each conversation costs proportionally more to run. For a bank, that cost is noise compared to one mis-sold product. For a flower shop, it is the whole margin.
What tech stack does each option need?
This is the question that separates a weekend project from a six-month one, so let us make it concrete:
- Salesbot: Kommo alone. Everything lives in the visual builder inside your account. No servers, no code, no subscriptions beyond your Kommo plan.
- Simple AI agent: still Kommo alone — it just has to be configured correctly: the prompt, the knowledge base, the handoff-to-human rules. No infrastructure of yours is involved; the model runs on the provider's side.
- Orchestrated agent: your own server plus real programming. And there is a reason it cannot be otherwise. The orchestrator has to hold the state of every conversation across multiple agents; it has to connect to your databases, billing and ERP — credentials that should never live inside a chat platform; it runs custom code, queues, retries and logging that no CRM can host for you. In this architecture Kommo remains the window where conversations happen, while the brain lives on your server and talks to Kommo through its interface.
Translation for the budget meeting: tiers one and two are configuration projects; tier three is a software product you will own — with the costs and the control that ownership implies.
What does each option cost — and how long does it take?
Real numbers from our own practice — what we charge and how long the work actually takes:
| Tier | Build cost | Timeline | Running cost |
|---|---|---|---|
| Salesbot scenario | $100–500 | from 1 day | included in your Kommo plan |
| Simple AI agent (single prompt) | $300–600 | about 7 days | token costs — typically cents per dialog |
| Orchestrated agent (multi-agent + database/billing integrations) | $3,000–25,000 | 1 to 6 months | 3–10× the model calls per answer, plus your server |
The spread inside tier three is honest, not evasive: an orchestrated agent that checks stock and emails invoices sits near the bottom of the range; one that operates inside a regulated industry with audit trails and multiple system integrations sits near the top.
Side-by-side: Salesbot vs simple agent vs orchestrated agent
| Salesbot | Simple AI agent | Orchestrated agent | |
|---|---|---|---|
| How it answers | plays your script verbatim | composes from one prompt | routes to specialist agents, prompts built per situation |
| Free-form questions | weak — misreads intent | good | excellent, accuracy stays stable |
| Complex operations (databases, invoices) | no | no | yes |
| Beyond-chat automations (tasks, fields, reassignment) | excellent | limited | yes, via integrations |
| Tech stack | Kommo only | Kommo only, configured right | own server + custom code |
| Build cost / time | $100–500 / from 1 day | $300–600 / ~7 days | $3,000–25,000 / 1–6 months |
| Per-dialog running cost | none extra | cents | 3–10× a simple agent |
| Best for | fixed funnels, internal automations | diverse questions, modest stakes | regulated industries, operations in chat |
Which option fits your business? Six real scenarios
Match yourself to the closest row — this is the shortcut version of the consultation we run with clients:
- Window or furniture manufacturing with site visits. The funnel is fixed: qualify, book a measurement, confirm. Questions repeat. Salesbot — and spend the savings on response-time discipline instead.
- Retail shop where 80% of chats are «price, delivery, returns». Predictable menu of intents. Salesbot, with one good fallback branch to a human.
- Internal pipeline housekeeping — tasks, field updates, reassignments between shifts. Salesbot, always: this is what templates are for.
- International team qualifying leads in several languages, varied questions. Templates break here. Simple AI agent with a solid knowledge base and a human-handoff rule.
- Online school: schedules, payments, lesson rescheduling. Diverse questions, moderate stakes. Simple AI agent; revisit the orchestrated tier only if you add payment operations in chat.
- Bank, insurer, or B2B distributor that needs the bot to check a client in the database, reserve stock and issue an invoice. Accuracy is money and the bot needs hands. Orchestrated agent — the expensive option is the cheap one here, because one mis-handled operation costs more than the build.
What about WhatsApp compliance?
One constraint sits above all three tiers: since January 15, 2026, Meta restricts open-ended AI chatbots on WhatsApp Business API. Whichever tier you choose, the way the bot is registered, scoped and handed off to humans on WhatsApp must follow the new rules — we wrote a separate step-by-step guide on setting up a compliant AI bot in Kommo under Meta's 2026 policy. And once a bot hands a hot conversation to a person, the clock starts ticking on the human side: our SLA First Touch Control widget measures exactly that gap.