CRM and AI agents in 2026: what a business needs first and why

“CRM or an AI agent?” is the wrong question. An agent works on top of whatever data and processes a company already has. If your customers are scattered across a spreadsheet, a few messengers and one salesperson’s memory, an agent will simply spread that mess faster. Get your customer records in order first, automate the routine next, and bring in AI only where it clearly pays for itself.

That is not just our view. Over the summer of 2026, Gartner and Forrester published findings that arrive at the same conclusion from different directions.

10 : 1

AI agents per seller by 2028

Gartner forecast

< 40%

of sellers will say agents improved their productivity

Gartner forecast for 2028

5×

growth in inference cost per agentic workflow by 2028

Gartner estimate, "more than"

What the analysts found

Gartner, July 2026. Gartner predicts that by 2028 sales organizations will run ten times as many AI agents as they have sellers, yet fewer than 40% of sellers will say those agents made them more productive. The forecast draws on a survey of 210 chief sales officers and senior sales leaders carried out in January and February 2026. Gartner analyst Dan Gottlieb puts the core idea plainly: agents are only as good as the systems around them, and fragmented systems simply get scaled. The same release adds a second prediction: sales leaders who rework their data, automation and user experience will be five times more likely to see a return on AI than those going for quick fixes.

Gartner, August 2026. A separate forecast is about money. Inference cost per agentic workflow is expected to rise more than fivefold through 2028. An agent does not answer in one message: it reasons, checks itself and takes several steps, and each step burns tokens. Gartner estimates that sending a task to an agentic reasoning model costs the provider at least five times more than a basic chatbot reply. Its advice is not to default to a general-purpose autonomous agent for every task, but to route each job to the cheapest model that can handle it.

Forrester, July 2026. Forrester VP and principal analyst Kate Leggett, drawing on interviews with consultancies and customers, describes where agents already deliver measurable results in CRM. According to the post, agents fully resolve 71% of customer inquiries at Vanta, take over more than half of customer cases at Engine, and help Siemens qualify inbound B2B leads. These are results reported by the companies themselves, and all of them are large organizations with mature data.

The common thread: an agent is a layer on top. Without a clean customer database and a defined sales process it is expensive and of little use.

When a business needs a CRM

In practice, the need for a CRM depends on symptoms, not company size. We suggest taking it seriously if at least two of these sound familiar:

• leads arrive through several channels (website, Telegram, Instagram, phone) and nobody sees them in one place;
• a salesperson forgets to call back and the customer goes to a competitor;
• when an employee leaves, the history with their customers leaves too;
• you cannot say which channel last month’s sales came from;
• the only way for a manager to know the state of deals is to ask everyone in person;
• repeat sales and customer reminders happen “when someone remembers”.

If you have one salesperson, a dozen customers a month and a single lead channel, a tidy spreadsheet may well be enough. A CRM is not a goal in itself; it is a way to stop losing money that has already come in.

Automation, bot, agent: what is the difference

“AI agent” is now stuck on almost anything, so let’s separate the terms. Anthropic distinguishes workflows, where a model and tools follow predefined code paths, from agents, which decide for themselves what to do next and which tools to use. Its guidance is to start with the simplest option and add complexity only when needed: agents trade latency and cost for flexibility, and their mistakes can compound.

Level	What it does	Example	When it fits
CRM rule	Acts when a condition is met	New lead → the manager gets a notification and a task	Always; this is the foundation
Scripted chatbot	Guides the customer with buttons	The bot takes a request and books a consultation	Standard questions, simple requests
AI assistant	Answers free-form questions from your knowledge base	The bot explains delivery terms and warranty	Many similar questions phrased in different ways
AI agent	Plans its own steps and acts across several systems	Reads a lead, checks the customer’s history in the CRM, drafts an offer and creates a task	Many variations, unstructured data, rules that no longer fit a script

Our take: most small companies get what they need from the first three levels. A full agent makes sense once a CRM is already running and there is a steady flow of repetitive but not formulaic work.

The order we recommend

1. 1
   Describe your sales process
   deal stages, who owns what, what counts as closed
   ↓ →
2. 2
   Move customers into a CRM
   one database, every lead channel, full contact history
   ↓ →
3. 3
   Simple automations
   notifications, reminders, assigning an owner
   ↓ →
4. 4
   AI on one narrow task
   FAQ answers, email drafts, call summaries
   ↓ →
5. 5
   An agent with a human in the loop
   actions only after review, limits and a decision log

Veb-Dev framework based on recommendations from Gartner and Anthropic

Each step should pay for itself. Even if you never get to an agent, a CRM with reminders already recovers leads that used to slip away, and that is the cheapest sales growth there is.

When an AI agent actually pays off

We recommend adding an agent when these conditions line up:

1. You have the data. Customers, deals and conversation history live in one system, not in people’s heads.
2. You have the volume. The task repeats dozens of times a day: sorting leads, answering inquiries, qualifying prospects.
3. The rules no longer fit a script. If the logic fits into a dozen if-then conditions, regular automation is cheaper and more reliable.
4. Mistakes are contained. The agent prepares, a person approves, especially where money, discounts or commitments to customers are involved.
5. You measure outcomes. Not just hours saved, but conversion, response time and resolved requests. Gartner specifically advises looking beyond time savings to the effect on sales results.

Common mistakes we see in the market:

• buying “AI for sales” before there is a CRM, so the agent works with incomplete data;
• automating a chaotic process instead of describing it first;
• letting the agent message customers without review;
• leaving ongoing model costs out of the budget, even though Gartner expects them to grow for agentic workflows.

Anthropic also advises testing agents in a sandbox, capping the number of steps and building in checkpoints where the agent stops and waits for a human decision. For a business, that boils down to one rule: the agent suggests, a person decides.

Off-the-shelf CRM or your own

Cloud CRMs are quick to start with, but they often charge per user, and your process has to bend to their logic. A custom system costs more up front, but it mirrors the way you actually sell and does not get more expensive as the team grows. For a small team with a non-standard pipeline we often suggest the second route: your own CRM system from us costs a fixed $950 and takes 2–3 weeks, including data migration from Excel and team training. An AI bot or agent can be added on top later without rebuilding the database.

What to do next

• Go through the symptom list: if two or more apply, you need a CRM now.
• Describe your sales process on a single page before choosing any system.
• Start with simple automations and try AI on one narrow task with a clear success metric.
• Give an agent the right to act only after a few weeks of “it suggests, a person approves”.

Not sure where to start in your case? Tell us how your sales work today and we will suggest whether you need a CRM, a bot or an agent, and in what order.

Author: Sergey Filatyev, founder of the Veb-Dev studio.