Salesforce's own partners are not seeing Agentforce ROI: what that means before you buy a packaged agent platform. Nucleo

Q: Is Agentforce delivering ROI for enterprises?

The evidence is split by who is measuring. Salesforce reports Agentforce annual recurring revenue growing at a triple-digit percentage rate year over year, and a Salesforce-commissioned Forrester Total Economic Impact study models a large return for a composite customer. A TD Cowen survey of Salesforce's own global implementation partners, reported in August 2026, found no partner naming Agentforce as a driver of recent bookings. Platform revenue and field-level return are answering different questions.

Q: Does a packaged agent platform lock you in?

It locks in the part that matters most, which is not the model. The agent's reasoning is bounded by the platform's data model, its permissions, its integration surface and its roadmap. You can usually change the underlying model. You cannot easily take the accumulated context, the evaluation history or the workflow logic somewhere else, because they were never assembled as an asset you hold.

Q: When is buying a packaged agent platform the right decision?

When the workflow is commodity, sits inside one system of record, carries no proprietary data advantage, and you have no intention of owning its maintenance. Case deflection inside a CRM you already run is a reasonable buy. A workflow that crosses the CRM, the ERP and two legacy systems, and that encodes how your company specifically competes, is not.

Q: How should a CIO test a vendor ROI claim?

Ask who paid for the study, what the baseline was, whether the organisation described is real or composite, and what the same number looks like from the people doing the implementations. A vendor-commissioned model and a survey of the vendor's own delivery channel are both evidence. When they disagree, the disagreement is the finding.

The short answer

Salesforce reports Agentforce as its fastest-growing product line while a survey of its own implementation partners finds none of them crediting it with revenue, and both statements are true because they measure different things. A TD Cowen survey of Salesforce partners across the United States, Europe and Asia, reported on 21 August 2026, found that not one respondent named Agentforce as a driver behind recent bookings activity, with a third reporting buying and trial interest and the rest waiting for the product to mature. (The Register)

For a CIO being pitched a packaged agent platform as the fast, low-risk path, the useful reading is not that Salesforce is failing. It is that platform adoption and delivered business value have come apart, and the gap between them is the risk you inherit when you sign.

Key takeaways

Salesforce’s own delivery channel is the group reporting no Agentforce-driven bookings. That is the least motivated source available, which is what makes it worth reading. (source)

The analyst downgrades on 9 July 2026 were about customer feedback, not about financials. Both KeyBanc and Bernstein cited the same two things: customer data not in shape for agents, and the product not being ready for broad deployment. (source)

Salesforce’s second-quarter fiscal 2027 results, reported 26 August 2026, show Agentforce annual recurring revenue growing at a triple-digit percentage rate year over year. Revenue growth is real and it is not the same claim as return. (source)

A Salesforce-commissioned Forrester Total Economic Impact study models a large return for a composite customer. A composite customer is a construct, not a reference. (source)

The recurring blocker named by analysts is not the agent. It is that the company’s data was never assembled into something an agent can reason over, which is work no packaged platform does for you.

What the partner survey actually found

TD Cowen surveyed Salesforce implementation partners globally and asked them what they were seeing on Agentforce. The distribution matters more than the headline. A minority said they had seen little immediate interest. A clear majority expected interest to arrive later, once the product had matured further. Roughly a third said interest had risen enough that buying and trial activity had started. And none said it was driving bookings today. (Salesforce Ben)

Read that as a delivery signal rather than a market verdict. Implementation partners are the people who install the thing, bill for installing it, and would very much like it to sell. They are not sceptics with a short position. When that group reports interest without conversion, the constraint is usually somewhere between the pitch and the production system, and the partners are the ones standing in that gap.

The analyst work published six weeks earlier said the same thing from a different vantage point. On 9 July 2026 KeyBanc moved Salesforce from Overweight to Sector Weight and removed its price target, and Bernstein downgraded on the same day. Neither cited a weak quarter. Both cited customer feedback: enterprises do not have their data in a state where agents can do meaningful work, and the product, in KeyBanc’s phrasing, “just isn’t there” for broad deployment. Both firms also reported surveys in which more CIOs planned to deprioritise Salesforce spend over the coming year than planned to increase it. (Investing.com)

The vendor’s numbers say something else, and both can hold

Six weeks after those downgrades, Salesforce reported second-quarter fiscal 2027 results and called it a record quarter. Agentforce annual recurring revenue grew at a triple-digit percentage rate year over year, and the broader artificial intelligence and data portfolio grew faster still. (Salesforce)

Nothing about that contradicts the partner survey, and a buyer who lets one cancel the other will draw the wrong conclusion from both. Annual recurring revenue measures contracts signed. It rises on new logos, on bundling, on renewals repriced to include a product, and on enterprises buying capacity ahead of a deployment they have not yet built. Bookings attributable to Agentforce, as reported by the people delivering it, measure work that reached a production workflow and produced something a customer would pay again for. A product line can grow quickly on the first measure while the second stays flat, and that is precisely what a two-year adoption curve for enterprise software usually looks like from the inside.

The third number in circulation is the Salesforce-commissioned Forrester Total Economic Impact study, which models a large three-year return for a composite organisation built from interviewed customers. (Forrester TEI) Vendor-commissioned modelling is not worthless, but it is a model. The composite organisation does not exist, its baseline is constructed, and the study was paid for by the party whose product is being evaluated. Set it beside a survey of the vendor’s own delivery channel and you have two pieces of evidence pointing in opposite directions. The disagreement is the finding, and it is the thing to put in front of your board, not either number on its own.

Why the packaged platform is not the low-risk option

The pitch for a packaged agent platform is that it sits inside a system you already run, so the integration work is done and the risk is small. The first half is true. The second does not follow.

An agent is only as capable as the context it can reach and reason over. Inside a packaged platform, that context is bounded four ways, and none of the four is the model:

The data model. The agent sees what the platform’s objects and fields describe. Everything your business knows that lives in the ERP, the pricing system, the contract archive or the spreadsheets your commercial team actually runs on is outside that boundary until someone does the work to bring it in, and bringing it in usually means copying it into the vendor’s shape.
The permission surface. Agents inherit the platform’s identity and sharing model. That is convenient until the workflow crosses a boundary the platform was never designed to express, at which point you are building an exception rather than a capability.
The roadmap. What the agent will be able to do next quarter is decided by the vendor’s product organisation, weighed against every other customer. Your third-most-important workflow is nobody else’s priority.
The accumulated context. The evaluations, the corrections, the retrieval logic, the domain rules your team encodes over eighteen months of tuning: that is the asset. Inside a packaged platform it accrues as configuration in someone else’s system, and it does not travel.

Model portability gets discussed because it is visible and easy to promise. Those four layers are where the actual dependency accumulates, and they are the reason the KeyBanc feedback keeps returning to data readiness. The blocker enterprises hit is not that the agent is not clever enough. It is that nobody built the layer underneath it, and no packaged platform builds that layer for you, because doing so would mean reconciling systems the vendor does not sell.

This is the same argument our build vs buy vs embed post makes in the abstract. Agentforce is what it looks like when a real market runs the experiment in public.

When buying is the right call

Buying the packaged agent is correct more often than a company in our position usually admits. The test is narrow and worth applying honestly.

Buy when the workflow lives inside one system of record, is close to commodity, encodes no advantage specific to your company, and you have no appetite to own its maintenance. Case deflection and routing inside a CRM you already run, service-request triage, a first-draft assistant sitting on tickets: these are reasonable purchases, and building them yourself is an expensive way to arrive at a worse version of something you could have configured.

Do not buy when the workflow crosses systems, when the value sits in the handoffs between departments rather than inside any one of them, or when the context that makes the workflow correct is knowledge only your company holds. That is the case where the packaged platform gets you a demo quickly and then stalls at the same wall the partners are describing.

What to establish before you sign

Four questions, asked before the contract rather than after the pilot:

Where does the context live when this is running? If the answer is “in the platform,” you have rented the intelligence and you are paying to keep it.
Who owns the evaluation set? The record of what the agent got wrong and how it was corrected is the most valuable artefact the project produces. Establish in writing that it is yours and that it is exportable in a usable form.
What is the number, and when is it measured? A business number agreed before the build, against a measured baseline, in a defined window. Without that, the ROI conversation happens after the spend, which is the wrong order and the reason so many of these programmes cannot answer for themselves. Our ROI benchmarks post covers what a defensible baseline looks like.
Who is accountable when it does not work? The partner survey is a description of what happens when nobody is: interest, trials, no bookings, and a product everyone is waiting on someone else to mature.

The pattern here is not specific to Salesforce. It is the same repricing of a delivery assumption that showed up in Accenture’s worst trading day in June. In both cases the market moved on a gap between what a delivery model promises and what it has demonstrated. In both cases the buyer’s protection is the same: own the layer the agents run on, keep it model-agnostic, and tie the engagement to a number somebody is accountable for.

FAQ

Is Agentforce delivering ROI for enterprises? The evidence splits by who is measuring. Salesforce reports triple-digit percentage growth in Agentforce annual recurring revenue, and a Salesforce-commissioned Forrester study models a large return for a composite customer. A TD Cowen survey of Salesforce’s own global implementation partners found no partner naming Agentforce as a driver of recent bookings. (The Register)

Why did analysts downgrade Salesforce over Agentforce? KeyBanc moved Salesforce to Sector Weight on 9 July 2026 and Bernstein downgraded the same day, both citing customer feedback rather than financial results: customer data not ready for agents, the product not ready for broad deployment, and more surveyed CIOs planning to cut Salesforce spend than to raise it. (Seeking Alpha)

Does a packaged agent platform lock you in? It locks in the part that matters, which is not the model. The data model, the permission surface, the vendor roadmap and the accumulated context all sit inside the platform. Swapping the model underneath does not move any of them.

When is buying a packaged agent platform the right decision? When the workflow is commodity, lives in one system of record, carries no proprietary advantage and you do not want to own its maintenance. A workflow crossing the CRM, the ERP and two legacy systems, encoding how your company specifically competes, fails all four tests.

How should a CIO test a vendor ROI claim? Ask who paid for the study, whether the organisation described is real or composite, what the baseline was, and what the same claim looks like from the people running the implementations. When a vendor’s model and a vendor’s delivery channel disagree, that disagreement belongs in the board paper.

Nucleo builds the context engine your company owns, model-agnostic and in production, with engineers inside your stack and a number agreed before the build starts. Talk to us.