Most CRM discussions about agents concentrate on the software and AI working for the sellers. It’s time to start paying more attention to the agents working for the buyer.
Gartner predicts that 90% of B2B buying will be agent‑intermediated by 20281 and Forrester expects agents to play a role in 75% of B2B purchases already in 2027.2 The interesting question from the CRM perspective is what happens when one of those agents needs something from our CRM.
CRM has changed... or has it?
I’ve been implementing CRM systems since 2003. We’ve changed the software, moved to cloud, added and removed social CRM, went through Agile hype and changed implementation models more times than I can count. Yet most CRM data models still describe customer relationships in much the same terms.
In B2B CRMs we have Accounts, Contacts, Leads, Opportunities and Activities. We can track who works for whom, who attended a meeting, who should receive an offer and who is in the buying committee. We can control which employee can access a customer record and audit the changes they make.
Now consider a request that comes from a customer’s AI agent rather than the customer directly. It might be asking a simple question about our product or it might want to check contractual terms on behalf of a procurement team. It might be asking to renew a service, amend an order or even cancel a contract.
Most CRM discussions about agents still concentrate on the software working for the seller. It’s time to start paying more attention to the changes to software working for the buyer.
What are the predictions?
Gartner predicts that by 2028, 90% of B2B buying will be intermediated by AI agents, representing more than $15 trillion in spend passing through agent exchanges. Forrester expects agentic buying to play a role in 75% of B2B purchases in 2027.
These are forecasts, and neither says that agents will independently negotiate and sign most enterprise contracts. That will not happen soon. An agent that compares suppliers or drafts an RFP is already part of the buying process, even when a person makes the final decision. Forrester explicitly expects fully autonomous purchasing to remain concentrated in routine, lower-risk transactions.
In a Gartner survey published in May 2026, 45% of B2B buyers said they had used generative AI during a recent purchase. In the same research, 69% preferred to validate AI-generated insights with a sales representative.3
Buyers are using AI to do more work before speaking to a seller, but they still want someone accountable when the decision matters. A procurement agent might narrow a list of ten providers to three but the buyer still needs to understand whether the shortlist makes sense.
That creates two separate challenges for CRM and the systems around it: what an agent can learn about us, and what it can do on customer’s behalf in our systems.
Being discoverable is not the same as being credible
We focus our agentic efforts on disjoint matters. On one side we use them to generate deman, score leads and run marketing acitvities / generate content. On the other side buyers us AI agents to compare sellers. In our case, a B2B highly skilled boutique, such an agent looks for industry experience, supported countries, project references, certifications and evidence of delivery quality.
Most companies already publish some of this information and generate a lot of marketing content, even actively pursuing GEO, not only SEO. The trouble is that it is scattered across websites, sales presentations, case studies and PDFs. The website says a provider has projects in 47 countries; a deck says 53 (our own case). The sales deck is newer than the website, but only the account team knows that.
A human buyer may call and ask for clarification. An agent may rank another provider higher because its claims are easier to check. Forrester predicts that two-thirds of B2B providers will lose deals when buyers cannot verify their claims.
However, that doesn’t mean that we should publish all data about us for it to be “AI discoverable”. Not everything needs a public API. Enterprise pricing is often negotiated, references may be confidential and some documents require an NDA. But we should know which facts we are prepared to publish, which need authentication, and which require a conversation. The same rules should apply whether the request comes from a website visitor, a salesperson or a buyer’s agent.
The harder problem begins when an AI agent wants to act
Consider a less hypothetical enterprise scenario.
A procurement manager asks an AI agent to review the company’s software agreements and identify contracts renewing within 90 days. The agent contacts a supplier and requests the current license count, renewal date and available options.
So far, the request is mostly about access to information. But suppose the agent then asks the supplier to reduce the license count by 20%, accept a new term or cancel an auto-renewal.
At what point does an inquiry become an instruction?
Who authorized that instruction?
Does the procurement manager have the right to make that change, or is finance approval required?
What evidence does the supplier have that the agent is faithfully carrying out the instruction it received?
CRM systems can already handle API calls, service accounts, customer portals and integrations for years. I don’t think the arrival of AI agents means we need to replace the Contact entity or invent a new identity system from scratch. The missing piece is more specific - an authenticated technical caller is not necessarily an authorized representative for the business action it requests.
A permission for an AI agent is not the same as a mandate
Knowing that a user or an AI agent connected to our API is useful. Knowing that it is acting for Company ABC is even better. However, neither fact establishes that the agent can terminate ABC’c agreement or commit ABC to another three years of service.
The distinction exists today with humans too. A purchasing assistant can ask for a quotation without being authorized to sign it. An agent makes that distinction more important because the request may move directly from an API response to a system action, without the normal conversation in which a salesperson would notice the difference.
We’ve spent years designing permissions around what employees and applications can read or write. A customer’s agent introduces a slightly different question: what has the customer authorized this particular agent to do, within which limits, and for how long?
How to solve it?
We need to take buyers’ AI agents into consideration when implementing enterprise CRMs. Based on our experience, I can suggest a design that distinguishes at least four things (my take):
The principal: the person or organization the agent represents.
The agent: the identifiable system making the request.
The mandate: the action it is allowed to request, including limits, applicable records and expiry.
The evidence: how the permission was established, how it was checked and what actually happened.
It’s harder than it seems because the customer is still a Contact in our CRM, but the agent representing them may exist only in the identity or integration layer, or as a temporary API caller.
The human handoff will matter more, not less
One of the more interesting parts of Forrester’s 2027 predictions is that buyers are expected to increase their interactions with people to verify recommendations generated by AI.4
That seems plausible for complex B2B purchases. One can use an agent to compare contract terms across a dozen suppliers. They would not expect the agent to determine, on its own, whether one trusts the team that will run a five-year transformation programme.
The supplier needs to know how to handle both situations. An agent should be able to retrieve approved, accurate information without unnecessary friction. When the buyer wants to challenge an assumption, discuss an exception or approve a material commitment, the process must make the right person reachable.
This is also why a conventional CRM sales stage is a poor substitute for an agreement about authority. Moving an Opportunity to Negotiation tells our team something about the forecast. It tells the buyer’s agent nothing about whether a proposed discount is binding, whether legal has approved the terms or whether the person making the offer can commit the company.
How should we prepare?
I wouldn’t start by adding an “Is AI Agent?” field to Contact. Nor would I start by building an agent that talks to other agents because the diagram looks good in a presentation.
We should rather pick one real customer process, preferably one that crosses from information retrieval into an authorized action. Then we should walk through the entire exchange from the customer’s side and try to map the customer journey:
Can a third-party agent find the correct product and contractual information?
How does it authenticate when it needs account-specific data?
How do we establish whom it represents?
Which requests can be completed automatically, and which require approval?
When the transaction is audited six months later, can we reconstruct what was requested and why we accepted it?
If we cannot answer those questions for one process, launching more agents on the seller’s side will not help.
For years, we have treated CRM automation as something the supplier deploys to serve its customers. But in 2027, customers will increasingly bring their own software into the relationship and some of it will be capable of more than asking questions.
Before deciding how our CRM should respond to those agents, we need to know what they are actually allowed to represent and do. That is a much more interesting design problem than deciding how many agents we can give a salesperson.
“AI’s Influence Runs Deeper Than You Think — 2026 Gartner Strategic Predictions Explain Why”, Gartner, 2025-11, https://www.gartner.com/en/articles/strategic-predictions-for-2026
“Forrester’s 2027 B2B Go-To-Market Predictions: AI Agents Will Influence 75% Of B2B Purchase Decisions”, Forrester, 2026-10, https://www.forrester.com/press-newsroom/forresters-2027-b2b-go-to-market-predictions-ai-agents-will-influence-75-of-b2b-purchase-decisions/
“Gartner Survey Finds 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights”, Garner, 2026-05, https://www.gartner.com/en/newsroom/press-releases/2026-05-20-gartner-survey-finds-sixty-nine-percent-of-b-two-b-buyers-turn-to-sales-reps-to-validate-ai-generated-insights
“Predictions 2027: AI Takes Over Business Buying, But Humans Retain Control”, Forrester, 2026-10, https://www.forrester.com/blogs/predictions-2027-ai-takes-over-business-buying-but-humans-retain-control/