Agentic CRM in 2026: When to Buy and When to Build

Agentic CRM stopped being a roadmap slide this month. Between July 21 and July 22, 2026, three separate vendors announced agent products that write to, read from, or sit on top of a customer record — Affinity’s Ascend, SeoSamba’s ActionEko, and TrueDialog’s TrueConnect and TrueOptOut. What none of them has, as of this writing, is a single independent evaluation, published benchmark, or audited customer outcome.

That gap matters more than any feature list. A CRM agent is not a chat window — it is a piece of software you are proposing to give write access to the system your revenue forecast is built on. The usual buying reflex, which is to compare feature grids and pick the longest one, is exactly the wrong reflex here, because every feature grid on the market right now is written by the company selling the feature.

So this guide does something different. It does not rank the three launches — they serve different markets and are not substitutable. Instead it gives you five questions that expose a thin wrapper before you sign, a seven-axis decision matrix for buy versus build, the first-party pricing you need to model either path honestly, and a 90-day sequence for teams that have to decide this quarter.

The trade press covered these three announcements as a roundup, which is fair — they landed within a day of each other. But treating them as competitors would be a category error. One targets private capital firms, one targets B2B and SMB marketing operations, and one targets enterprise SMS compliance. They are not substitutes. What they share is a shape: an agent layer bolted onto or beside a customer record, announced with capability claims and no external verification.

Everything in the cards below is what the vendor says. We have not tested any of it, and neither, as far as we can find, has anyone outside the companies themselves.

Affinity’s chief executive, Ken Fine, framed the launch in the GlobeNewswire release relayed by Yahoo Finance as “We’re building a future where every private capital professional can work alongside AI agents.” The company also published a customer testimonial from Lukas Huber at Speedinvest, who said meeting prep went from 10 to 15 minutes down to “a two-minute skim.” That is a vendor-supplied case example from a named individual, not an audited outcome — useful as colour, not as a benchmark for what your firm should expect.

SeoSamba’s ActionEko was covered by MarTech Series, which relays the vendor’s claim that the product activates in “minutes” against “months” for traditional CRM setup. There is no independent benchmark behind that comparison, and it is the kind of claim a pilot exists to test.

TrueDialog’s claims, published via PR Newswire, are the most specific of the three, and therefore the easiest to test later: the vendor says TrueOptOut processes inbound messages in under 10 milliseconds and targets a 90% recall rate on opt-out intent, logging each decision as an immutable compliance record. Those are targets the company has published about its own product. No independent accuracy test exists that we could find. Treat them as a specification to hold the vendor to during a pilot, not as a measured result.

The argument for caution here is not that these products are bad. We do not know whether they are good, and neither does anyone outside the vendors. The argument for caution is that the two categories this sits between — CRM implementations and enterprise generative-AI pilots — both have unusually poor completion records, and an agentic CRM project inherits the failure modes of both.

On the CRM side, the most concrete public figure we could source comes from Johnny Grow, a CRM consultancy that self-publishes research on implementation outcomes. Its CRM Failure Report puts the overall failure rate — defined as deployments that did not achieve their planned objectives — at 55%, with 70% of failed or struggling deployments overrunning their timeline by 30% or more, and typical budget overruns landing in the 30% to 49% range. Read that as one consultancy’s self-reported research: the public page does not disclose a sample size or a survey date, so it does not carry the same weight as the first-party pricing figures later in this guide.

The NANDA finding that actually matters for a buy-versus-build call is not the headline 95%. It is the delivery-model split: pilots run by a blend of internal specialists and external partners succeeded at roughly three times the rate of IT-only builds. That is a statement about capability and ownership, not about model quality — and it maps directly onto the fifth axis of the matrix below. If you have no internal owner for an agent, buying does not fix that. It just moves the failure from your build to your rollout.

Worth noting alongside this: the wider agentic-AI market is already producing cancellation forecasts of its own, which we covered separately in our analysis of projected agentic AI project cancellations. The pattern across both datasets is the same: the technology rarely fails on its own terms, and the projects fail on integration, ownership, and data.

A thin wrapper is a product whose entire differentiation is a prompt template and a connector, sitting on a general-purpose model you could call yourself. Thin wrappers are not automatically bad buys — sometimes the connector is genuinely the hard part. They are bad buys at platform prices, on multi-year terms, with your workflow logic locked inside them.

These five questions are our own procurement framework, built from the governance patterns that recur across enterprise AI deployments. They are deliberately answerable in a first call. A vendor with real engineering behind the agent answers all five without escalating. A wrapper stalls on at least two.

The standard a serious answer sets is a vendor that can tell you which direction its product errs in, and why. TrueDialog’s CTO put that on the record when the company announced its opt-out layer — unverified as a performance claim, but a different class of answer than a feature bullet.

Every buy-versus-build framework we found in the market is either generic AI-agent cost maths or a vendor feature-comparison grid. Both miss what actually decides a CRM agent project, because the deciding variables are properties of the buyer, not the seller. So the matrix below is built the other way round: the rows are conditions you can self-assess without a single sales call, and the verdicts fall out of your own answers.

Read it as three states, not a score. Where both columns feel half true, the fourth column tells you what to resolve first — which is almost always more useful than a forced verdict.

The pattern the matrix surfaces is that only two of the seven axes are really about the agent. The other five are about your data, your people, your obligations, and your contracts — all of which you can assess this week, at no cost, without a demo. That is the point. A team that scores itself honestly across these seven rows usually finds the answer is not uniform: buy for two workflows, build for one, and defer the rest until the hygiene pass is done.

Row one deserves emphasis because it is the one teams skip. If your CRM has duplicate accounts, orphaned contacts, and fields that three departments interpret differently, no agent — bought or built — will behave predictably on top of it. Our CRM data-hygiene agent pipeline walks through the dedupe and enrichment sequence we run at the start of every CRM automation engagement before a write-enabled agent goes live, and the context-layer failure mode that undermines RAG over messy CRM data is the same failure mode in a different costume.

Agent pricing has split away from seat pricing, and the two now move independently. Every figure below came off the vendor’s own pricing page on July 27, 2026 — Salesforce Agentforce, Sales Cloud, HubSpot, and Zoho — no aggregator, no resale estimate. The annualised totals are our own arithmetic from those unit prices, labelled as such, and modelled on an illustrative 20-person revenue team.

Two things fall out of that table. First, the base-seat spread is wider than most buyers assume: $175 against $40 per user per month is roughly a 4.4× difference before a single agent feature is switched on, which at 20 seats is $32,400 a year of pure platform premium. Whether that premium is worth paying is a real question with a real answer for some organisations — but it should be answered deliberately, not inherited.

Second, the three vendors are no longer priced on the same axis at all. Salesforce sells agent capacity as credits or per conversation, HubSpot sells outcomes, and Zoho bundles its assistant into the seat. You cannot compare them on a single per-user number, and any comparison that produces one has flattened away the thing that matters. Model your own expected action volume against all three shapes.

The outcome-priced model is the one worth watching, because it moves execution risk from the buyer to the vendor. Announcing the change on April 14, 2026, HubSpot’s chief customer officer Jon Dick put it as “Outcome-based pricing removes that risk. You pay when it works, full stop.” Whether that framing survives contact with edge cases — what counts as a resolved conversation, who adjudicates a disputed resolution — is exactly the sort of definition to pin down in a contract rather than a launch post.

Building is not the brave choice or the cheap choice. It is the correct choice in a narrow, identifiable set of circumstances, and an expensive mistake outside them. The circumstances are these: the logic is proprietary, the volume is high, the compliance requirement demands custody of the reasoning, or your incumbent platform has priced the native agent above what the workflow is worth to you.

Note what is not on that list. “We want control” is not a reason on its own — control is a cost you pay for in maintenance, and it only earns out when one of the four conditions above is genuinely present. Neither is “vendor pricing seems high”, unless you have modelled the build-side total honestly, including the operations and observability layer that never appears in the initial estimate.

If you want a worked example of the build path rather than the argument for it, our walkthrough of how to build an AI lead-scoring agent for your CRM covers signal design, the scoring rubric, CRM sync, and the human review loop end to end. And the general TCO question — at what conversation volume does a build stop being more expensive than a platform — is worked through in our enterprise AI agent build-versus-buy analysis, which puts the general crossover in the region of a million agent conversations a year. That figure is not CRM-specific and should not be transplanted directly, but the shape of the curve holds: buy-side cost is linear, build-side cost is concave, and the crossing point moves every time a vendor reprices.

This same tension has already played out one category over. The buy-versus-build argument in agentic ABM tools versus custom account systems reached the same conclusion by a different route: the tools win on the standard motions, the custom systems win where the account model itself is the differentiator, and the expensive failures are the teams that bought a platform to solve a data problem.

The moment an agent can modify a CRM record, it enters your control environment. If those records feed revenue recognition, credit terms, or pricing, security and audit commentary through 2026 has increasingly framed over-privileged agent write access as an emerging internal-controls concern — the sort of thing an auditor will ask about even before regulators formalise anything. That framing is directional trade commentary rather than settled guidance, but it is the direction the conversation is moving, and it is cheaper to architect for now than to retrofit.

The six controls below are our own standard for enabling agent writes, bought or built. None of them are exotic. All of them are easier to specify before signature than to negotiate afterwards.

The sequence below is the one we run inside AI transformation programmes, and it exists because the failure pattern in this category is almost never a bad product choice made in month one. It is a good product choice made before the data was ready, the owner was named, or the rollback path existed — so the agent gets switched off in month four and the licence runs to renewal unused. The sequence below front-loads the things that cannot be bought.

Our own view on where this goes over the next year is that the evidence gap closes from the compliance end first, not the capability end. Products with narrow, testable claims — opt-out recall rates, deduplication precision, routing accuracy — are the ones buyers can actually verify in a pilot, so they will accumulate independent evidence fastest. Broad next-best-action agents, whose value is diffuse and whose counterfactual is unmeasurable, will stay unverifiable for longer, which means the burden of proof stays with the buyer. Plan your pilots accordingly: pick the workflow whose success you can measure, not the one with the most impressive demo.

The second projection is about pricing shape. HubSpot moving its agents to per-outcome billing in April 2026 is the more interesting signal of the year so far, because it shifts execution risk from buyer to vendor. If that model spreads, the build-versus-buy maths changes materially — a vendor that only bills on success is competing against your build on a different footing than one billing per attempt. Watch whether competitors follow, and re-run your crossover model when they do.