AI agent for CRM in 2026: 20 solutions, prices, and why agents lie about your pipeline

Three markets under one query

The first thing to do before choosing is to work out which of the three markets your task belongs to. Solutions from different markets are not interchangeable, and conversations with vendors go nowhere precisely because the two sides are discussing different classes.

AI inside the CRM. A button in the deal card and in chat. Call transcription, conversation summaries, a draft email, auto-filled fields, a suggested next step. The user is the sales rep who already spends the day in that system. Nothing to install: it turns on with the plan.

A bot in the conversation. A separate participant that handles the chat or the call instead of a person: qualifies the lead, answers routine questions, books a meeting, creates a deal with the fields filled in. The user is your customer, not your employee. Sold as a widget for amoCRM or Bitrix24, or as its own platform with a CRM connector.

An external agent over the data. It does not write to customers and does not sit in the record. It answers the manager's questions: where the pipeline stalled, which deals are stuck, how much is uncollected on overdue invoices, whether payments in the CRM match the bank. It reads the CRM through the API from the outside — and almost always reads more than the CRM.

We call them markets A, B and C below. Every row of the table states its class.

Comparison table: 20 solutions for CRM

Prices are collected from public price lists and reviews as of 20 August 2026. Where a vendor publishes only "on request", we say so — filling that in with "roughly from" would be dishonest: in the enterprise tier the spread is a multiple, not a percentage.

How we selected: six criteria

The solution has to work in Russia and be sold in roubles. Zoho, HubSpot and Salesforce with their AI are not on this list because of payment availability and where your customers' data is stored, not because of quality.

It has to have declared AI, not automation. A trigger that says "create a task after three days of silence" is not AI, it is a rule, and CRMs have had it for a decade.

It has to be clear who the user is. A sales rep, a customer or a manager are three different products, and we do not mix them in one ranking.

There has to be a public price or an honest "on request". "From 0 ₽" for a product that does not work without the AI package is not a price.

There has to be a connection method you can describe to your own admin. If after a call with the vendor it is unclear what exactly lands in your portal, that is a risk, not flexibility.

The solution has to be alive in 2026. Half of the "top 10 AI for CRM" lists in search results name products whose latest news is from the year before last.

Market A: AI inside the CRM — and its meter

The main thing to understand about 2026: AI inside a CRM is no longer "included in the plan". It is included in a volume sufficient for a demo, and beyond that it is billed separately — differently in every system.

Bitrix24 counts requests. At the Cosmos release in April 2026 CoPilot was renamed BitrixGPT, the Martha AI agent appeared alongside it with connections to external systems over MCP, and since 10 August 2026 Martha's capabilities are available directly in the chat with the BitrixGPT agent. For the self-hosted edition, a combined "BitrixGPT + Marketplace" subscription has been sold since 11 March 2026, lifting the cap on the number of requests in the Russian zone; the minimum platform version is 24.700.0.

The official documentation is careful about the free volume: it says "basic limit". Integrator write-ups give numbers: 100 requests a month on the free plan (and no more than five a day per user), 100 on Basic and Standard, 200 on Professional, 400 on Enterprise. The key part is not the number but the words per company: the limit is shared, not per employee. And consumption is uneven — an ordinary operation such as an email or a task costs one request, processing a call (transcript plus summary plus field filling) costs three, speech analytics two, generating a site between five and twenty-five. A team of ten reps that starts transcribing calls will burn the monthly Basic allowance in under a working day. After that come boosts: 1,000 requests a month for 9,900 ₽/mo or 100,001 ₽ a year; while a boost is active, free requests are not accrued.

The Basic plan itself costs 2,490 ₽/mo billed monthly and 1,990 ₽/mo billed annually, and that is a workspace for up to five employees. So the honest cost of "Bitrix24 with AI for a sales team" is the plan plus the boosts, and the second part easily outweighs the first.

amoCRM counts tokens. Amma is the built-in AI agent presented at AMOCONF on 18 April 2026. It plugs into conversations, suggests replies, reminds you about customers left unanswered, finds problem areas in deals, assembles reports without hand-built dashboards, and can configure an account from scratch through dialogue. Billing is by tokens: 100,000 tokens for 999 ₽, some included in the standard plans, consumption depending on the complexity of the request. The base subscription has not changed: 599, 1,199 and 1,699 ₽ per user per month.

A token is a unit you cannot plan for in advance: not even the vendor knows how many tokens "analyse the month by pipeline" will consume. So the only honest way to budget is to enable it for one team and watch the burn for two weeks.

RetailCRM and OkoCRM bill it as a separate line. RetailCRM starts at 9,360 ₽/mo for three users billed annually, with agent usage on top; functionally that is transcription, scoring rep conversations against criteria, auto-tagging dialogues and chat summaries. OkoCRM is 572–712 ₽ per user (price lists differ across reviews) plus AI from 5,000 ₽ per 1,500 requests. PlanFix works the same way: 366 ₽ per user plus from 915 ₽ for AI credits — but its agents are at least named after the work they do: Dataminer pulls data out of a customer message into deal fields, Summarizer updates the record from the dialogue, Interviewer returns the summary.

The enterprise tier runs on different logic. BPMSoft's release 1.9 (February 2026) added a "Work with LLM" element to the process builder — the model is embedded straight into the BPMN diagram and an agent is assembled without code; next to it sits classical ML: predictive lead scoring, churn prediction, auto-classification and routing of requests. ELMA365 got ELMA Cortex, a separate platform of corporate agents embedded into processes. SimpleOne is developing its own GenAI architecture. Prices there are project-based, and comparing them with 999 ₽ for a token pack is meaningless: these are different purchases.

What built-in AI does not do in any of these systems is answer a question involving data outside the CRM. "Do the payments on deals match the bank statement", "why is revenue in 1C lower than the sum of won deals", "which CRM customers stopped ordering on the marketplace" — a built-in assistant does not cross the border of its own system. That border is the edge of market A.

Market B: bots in the conversation — where the meter counts dialogues, not requests

Here the billing unit is different, and more convenient: a dialogue or a minute.

Widgets for amoCRM are the densest segment. ChatAI: 9,990 ₽/mo plus 7–14 ₽ per dialogue, runs the conversation and qualifies the lead. NOVA builds a knowledge base out of your own reps' conversations — that is, it learns from how your best people answer rather than from a general corpus. Nextbot works with two CRMs at once, amoCRM and Bitrix24, which matters if you historically run both. Salebot wins on the cost of starting — a free tier and paid plans from 2,999 ₽/mo — and covers messengers and Avito.

Voice is a separate story. Tomoru is billed at 1.85 ₽ per dialogue plus 4.40 ₽ per minute of conversation; TWIN bills per second and claims seven CRM connectors — the widest matrix in the class — plus emotion recognition. One thing to understand about voice agents: their metric is not "answer quality" but the share of connected calls carried through to the target action, and you cannot judge that from a demo recording. Ask for a pilot on your own base and your own script.

The shared limitation of market B: the bot answers your customer, but it does not answer you. It will not tell you where the pipeline stalled or why demo-to-payment conversion dropped nine points. That is the neighbouring job.

Market C: an external agent over CRM data

Here the user is the manager, not the customer and not the rep. The questions sound like this: "which deals at the contract stage have been sitting for more than two weeks", "how much is uncollected on overdue invoices", "which reps have below-average meeting-to-invoice conversion and by how much", "does the sum of won deals for July match revenue in 1C".

There are three ways to buy that work. An agent builder — GigaChat Business (launched by Sber in March 2026, no-code, on-prem, compliant with Russian data law) or Just AI for enterprise: you get an environment in which the agent is assembled by you or your contractor. iPaaS — Albato with its thousand-plus connectors: that is transport, not thinking; it moves data and makes no decisions. And a ready-made external agent that already knows how to reach into the CRM and the systems around it — that is our class, and we cover it below.

How an AI agent actually gets into a CRM

There are five ways, and they determine almost everything else: how fast it connects, how it depends on updates, and who is accountable for the data.

Built in by the vendor. Nothing to connect, the AI is already inside. The upside is obvious. The downside: you do not choose the model, you do not know where the text of the conversation goes, and you run into the vendor's meter.

A marketplace app or a widget. Code lives in your portal or in the record's window, and permissions are granted at install time. Upside — it works out of the box. Downside — an app's permission set is usually broader than the task requires, and after installation it is nearly impossible to see what it actually reads.

An inbound webhook and REST from the outside. An external system reaches the CRM with ordinary HTTP requests. Nothing is installed into the portal, permissions are limited to what you granted when creating the webhook, and revoking access is one click. That is how we work. Downside — a webhook runs on behalf of the user who created it, and that user's permissions silently determine what the agent will see.

MCP. A layer that publishes a set of tools for a language model. In 2026 this became the fashionable word in decks, including Bitrix24's own — Martha AI reaches external systems through an MCP Hub. Remember: into the CRM itself, that layer still goes by one of the methods above, usually REST. MCP decides how the model calls a tool, not how the tool reads your pipeline.

Direct access to the self-hosted database. Fast, complete and very bad: bypassing the CRM's business logic you get tables, not deals, and any update breaks your integration. Fine for a one-off export, wrong for regular work.

Why an AI agent lies about your pipeline

This is the central section of the article, and it is missing from every vendor's material. The problem with the "agent over CRM data" class is not that it fails with an error. The problem is that it answers successfully and wrongly: the API returns 200, the number looks plausible, and you cannot check it without a second export. Here are eight mechanisms, each of which produces a silently wrong answer. This is not a list of fears — it is the list of questions worth asking any vendor before you buy.

There are several pipelines, and the question assumes one. Bitrix24 splits deals across directions, amoCRM across pipelines, and a single account can hold up to fifty. The question "how many deals did we close in July" without a pipeline filter will add sales, service, complaints and internal processes into one number. There will be no error — there will be a sum of incomparable things. The agent must either ask which pipeline or state in the answer that it counted all of them.

Leads are on or off — and the denominator of conversion changes. A Bitrix24 portal runs either in classic mode with leads or in simple mode where an enquiry becomes a deal immediately. In amoCRM, the Incoming Leads area sits apart from the pipeline. "Enquiry-to-payment conversion" computed on a portal without leads and on a portal with leads are two different metrics with the same name. They cannot be compared across teams or against last year — and they look identical.

Paging without a pinned sort order. Bitrix24 list methods return 50 records at a time, amoCRM up to 250. A selection is read page by page, and while the agent walks the pages, reps keep moving deals. If the order between requests is not pinned explicitly, one deal lands in the export twice and another never lands at all. The length of the result still looks plausible, and the only way to spot the substitution is to reconcile against the CRM itself.

Rate limits cut the export short, and a cut looks like the end of the data. Bitrix24's official limits: two requests per second on all plans except Enterprise (five there), a bucket of 50 accumulated requests, and on overflow QUERY_LIMIT_EXCEEDED with status 503. Resource intensity is metered separately: exceeding it returns OPERATION_TIME_LIMIT with status 429. A single REST request in the cloud must complete within 60 seconds. amoCRM allows no more than seven requests per second and up to 250 entities at a time. Bitrix24's batch method packs up to 50 calls into one HTTP request, but it does not lift the resource-intensity ceiling. What follows from all this: an agent that received a 503 halfway through the export and failed to tell it apart from "there is no more data" will report your July from three weeks.

Permissions trim the selection silently. A webhook runs on behalf of the user who created it; an app runs within the permissions granted at install. CRM access settings will remove other people's deals from the response with no error and no warning. A report built on trimmed data looks exactly like a complete one, and "the head of sales sees less than they think" is the most common way to get a beautiful wrong number.

The deal amount is not where you think it is. In Bitrix24 a deal has an amount in the deal's currency and an amount in the reporting currency; in a company that buys in yuan and sells in roubles, adding up the former at face value produces nonsense. On top of that, half of all companies keep the real amount not in the standard field but in a custom one — because there it is "including VAT" or "net of the bonus". An agent that reads the standard field will not compute what your finance director means. List-type custom fields, moreover, arrive as element identifiers rather than values: "Source = 4172" instead of "Source = trade show".

The time zone shifts the period boundaries. Dates come out of the API in the portal's or the user's time zone, and in amoCRM time in requests is a unix timestamp. "July" in the export and "July" on the rep's screen can differ by several hours, and every deal closed late on the 31st drifts into August. Again, no error is raised.

Duplicate contacts and deleted deals. One customer as three records is normal for a CRM that is three years old and has been fed leads from several sources. Revenue per customer is then understated, the customer count overstated, and repeat purchases invisible. The same trouble sits next door with lost and deleted deals: computing conversion without separating the failed stage from the active one, and without accounting for the fact that the API simply does not return deleted records, produces an optimistic picture.

Two structural limits of CRM REST interfaces are worth knowing in advance. There is no server-side aggregation — no grouping, no sums: any "how much in total" is computed after the export, which means it depends entirely on the export's completeness. And there is no batch atomicity: batch saves HTTP requests but does not turn five changes into one transaction — if it breaks off, some of the changes stay in the CRM.

In our agent all of the above is closed by mechanics rather than by advice in a prompt: the pipeline and the period are clarified before the calculation, not after; paged reading pins the order itself and marks a selection as truncated if it did not finish; a 503 and an "no more data" response are separated and lead to different actions; period boundaries are extended to the end of the day in the portal's time zone; a missing direction filter is stated out loud; and the completeness check sits before the arithmetic and stops the run instead of printing a footnote. You can verify this in five minutes: ask the agent a question whose answer you know exactly, and require it to name the pipeline, the period and the completeness of the selection.

Samreshuuu: the CRM together with telephony, 1C and the bank

In this table we are an external agent over the data, and there are three differences.

Six connections in the CRM class, not one. Bitrix24, amoCRM, RetailCRM, BPMSoft, Megaplan and YCLIENTS for the service industry. Built-in AI by definition lives in one system; a widget lives in one CRM; we read whichever one you already run, and we do not require a migration.

Nothing is installed into the portal. An inbound webhook with read permissions on the entities you need, and that is all. No app in the portal, no extension, no access to the self-hosted database. Revoking access is one click in the CRM settings — a principled difference from a marketplace app, whose permissions are no longer visible once installed.

The CRM is not the only source. Real business questions rarely live in one system: deals against the bank statement, won deals against shipments in 1C, enquiries against ad spend, calls from telephony against the rep's activity in the record. The same agent sees the CRM together with 1C, telephony, the bank, EDI and marketplace accounts — 103 connected services in total. What it looks like in practice: where deals leak from the funnel, triage of inbound leads in Bitrix24, reminders on stalled deals, follow-up after a proposal, quarterly pipeline summary, funnel bottlenecks, the weekly sales digest, checking customers in arrears.

The price is a free start, Pro at 2,000 ₽/mo, Max at 10,000 ₽/mo. There is no AI request meter on top of the subscription.

Honestly about the limits

We are not a bot in the conversation. If the job is to answer customers in chat instead of a rep, look at market B: ChatAI, NOVA, Nextbot, Salebot, and Tomoru or TWIN for voice. We do not go there and do not plan to.

We do not replace built-in AI where it is already convenient. Transcribing a call right inside the record, a one-click email draft, a chat summary — that is exactly what BitrixGPT and Amma sit inside the system for. If that is your entire scenario, there is no reason to buy an external agent.

We are not an enterprise platform. If you run BPMSoft or ELMA365 with processes hundreds of steps long and an on-premises requirement, your path is ELMA Cortex, BPMSoft AI, Just AI or GigaChat Business inside your own perimeter.

And we do not forecast. Lead scoring and churn prediction on historical data are classical ML, and they live in BPMSoft and its peers. We answer questions about what has already happened, and we answer them so that the number can be re-checked.

What the AI law changes

On 1 September 2026 the main provisions of Russia's first law on artificial intelligence (243-FZ) come into force. Two things matter for a company connecting AI to a CRM. First: the rules on sovereign and national models, developers' obligations, marking generated content and the use of works for training start later — on 1 March 2027. Second: marking content is an obligation of large platforms, not yours; a button, not a fine.

What actually concerns a CRM is not the new law but the old personal-data one. The records hold the names, phone numbers and correspondence of living people, and the question "where does the text of a conversation go when the AI writes a summary" must be asked before connecting, not after. For built-in AI the answer is usually "into the CRM vendor's perimeter", for a widget it is "into the widget developer's perimeter" — and those are two different contracts.

Checklist: how to choose in 20 minutes

1. Name the user. A rep, a customer or a manager — that decides the market, and after that you only compare within it.
2. Add up the full price. The CRM plan plus the AI meter. For Bitrix24, count the burn in requests: a call is three, speech analytics is two. For amoCRM, enable it for one team and watch the tokens for two weeks.
3. Check how many systems the question involves. If even one of your questions involves more than the CRM, built-in AI will not close it — and that is settled immediately.
4. Ask about permissions. On whose behalf the solution reads the CRM, and what it will see if that user's access to other people's deals is restricted.
5. Ask a control question. Take a metric you know exactly, ask the agent, and require it to name the pipeline, the period and the completeness of the selection. An answer without those three is a pretty number of unknown origin.
6. Check the way out. Revoking access with one click versus deleting an app with opaque permissions — the difference is only visible when you leave.

Frequently asked questions

How is an AI agent different from a chatbot in a CRM? A bot follows the scenario you drew and answers "I did not understand" to anything off-script. An agent decides for itself which data to pull and in what order to answer the question. More detail in what an AI agent is and the levels of autonomy.

Bitrix24 or amoCRM — whose AI is better? These are different billing models, not different quality. Bitrix24 counts requests and gives only 100 a month per company on the Basic plan, but the boost is bought predictably: 1,000 requests for 9,900 ₽. amoCRM counts tokens — 999 ₽ per 100,000 — and consumption depends on what you ask. For a team that transcribes many calls, Bitrix24 is easier to budget; for a team that lives in chat, amoCRM is.

Can AI be connected to a CRM without installing an app? Yes, with an inbound webhook: you grant read permissions on the entities you need, the external system reaches the CRM over HTTP, and nothing is installed into the portal. That is how we work — see how to connect AI to Bitrix24 and how to connect AI to amoCRM.

What does AI in a CRM cost for a team of ten? Count it in two lines. Bitrix24: the plan from 1,990 ₽/mo billed annually (up to five employees, so two licences) plus boosts — with active call transcription that is at least one boost, 9,900 ₽/mo. amoCRM: 5,990–16,990 ₽/mo for the subscription plus tokens. RetailCRM: from 9,360 ₽ for three users plus agent usage. An external agent with no meter: 2,000 ₽/mo for the whole team.

Can an agent explain why conversion dropped? It can, if it has access to every stage of the pipeline and honestly names the period and the completeness of the selection. Be wary of an answer that names neither the pipeline nor the period boundaries: it most likely added up different directions over different intervals — see the section on why an agent lies about your pipeline.

Do replies written by AI to a customer have to be labelled? 243-FZ does not introduce mandatory labelling of all AI content, and the marking rules take effect on 1 March 2027 and address large platforms. The practical risk today is a different one: personal-data law and whose perimeter the customer conversation ends up in.

We have two CRMs for historical reasons. What do we do? Built-in AI works in each of them separately and will never give you the combined picture. What you need here is either a bot that supports both (Nextbot) or an external agent that reads both systems and merges them into one answer — see how to merge sales from different systems into one table.

What to read next? A review of AI agents for 1C — if your next question is about accounting. Top Russian AI agents for business — if you are choosing an assistant for the whole company. Speech analytics services — if the task lives in your reps' conversations. AI agent or Make, Zapier, n8n — if you are thinking of assembling your own from connected services.