Two things are true about AI in revenue operations right now, and the industry keeps insisting you pick one.
Salesforce just crossed $1.5 billion in Agentforce ARR. HubSpot has agents running in more than half its Pro and Enterprise base. Four separate GTM vendors shipped MCP integrations in a single month. The infrastructure is real and the spend is real.
Also: Salesforce partners told a sell-side survey that Agentforce is not driving bookings. HubSpot missed its own customer-add expectation and reset guidance downward. McKinsey’s global survey found the share of companies attributing EBIT impact to AI is flat year over year. Gartner predicts agents will outnumber sellers ten to one by 2028 and that fewer than 40% of sellers will say it helped.
Both sets of numbers are correct. The gap between them is the actual RevOps story of the fall, and September (Dreamforce on the 15th, UNBOUND on the 16th) is when every vendor in your stack will try to talk you out of noticing it.
Here is what happened, and what to do with it.
Salesforce Crossed $1.5B in Agentforce ARR, and Also Changed What Counts
On August 26, Salesforce reported Q2 FY27 (quarter ended July 31). Revenue $11.3 billion, up 11%. cRPO $33.5 billion, up 14%. Non-GAAP EPS $5.90, up 103%. FY27 guidance raised to $46.1 to $46.4 billion. Agentforce ARR passed $1.5 billion, up 240% year over year, with Agentforce and Data 360 together at roughly $3.9 billion, up 210%. Marc Benioff: “We’re seeing incredible demand for our AI and data products, with ARR about to cross $4 billion.” The release is here.
Read the footnote, though. The same release states: “Effective Q2 FY27, Agentforce ARR includes our AI offerings, Slackbot and Headless 360.” The definition got wider in the quarter the number got impressive. That does not make the growth fake. It does mean the 240% is not a clean comparable, and anyone benchmarking their own agent spend against “Salesforce did $1.5 billion” is benchmarking against a moving definition.
Salesforce also introduced a new disclosure: 7.0 billion “Agentic Work Units” delivered to date, 3.2 billion of them in Q2, up 97% quarter over quarter. A work unit is a consumption metric, not an outcome metric. It tells you how much the agents did. It does not tell you whether any of it was worth doing. Hold that thought.
Salesforce Investor Day falls on September 16, mid-Dreamforce, so expect the framing to get considerably more confident before it gets more specific. If you are heading into a renewal, the consumption-pricing question (what a work unit costs you at scale, and what happens when usage triples) is the one to walk in with. That is a CRM implementation and governance conversation more than a licensing one.
Meanwhile, Salesforce’s Own Partners Say It Is Not Producing Revenue Yet
On August 21, The Register reported on two sell-side channel checks that land differently than the earnings release. The piece is here.
A TD Cowen partner survey across the US, Europe and Asia found 11% of partners seeing little immediate Agentforce interest, 56% expecting interest but saying it needs more time, and 33% reporting strong interest with buying or trial activity. The line that matters: none reported Agentforce becoming a driver of bookings. TD Cowen’s own summary was “Agentforce adoption [is] still subdued.” Only a third of partners were meeting or beating their targets, down from 43% the prior quarter.
A separate KeyBanc CIO survey in July found partners and customers reporting that their enterprise data was not coherent enough to do meaningful AI work.
Standard caveats apply: this is The Register reporting on analyst notes that are not public, and implementation partners are a lagging indicator by design, since they sell services against products that customers have already bought. But the KeyBanc finding is the one that should stop you, because it is the same finding every honest RevOps leader has been making privately for two years. The agent is not the bottleneck. The data model underneath it is. You cannot deploy an agent onto four systems of record that disagree about what a customer is and expect the agent to resolve the disagreement. It will just make the disagreement faster.
HubSpot Grew 20% and Still Had the Most Instructive Quarter of the Season
HubSpot reported Q2 on August 5: revenue $911.7 million, up 20% reported and 17% in constant currency. 306,446 customers, up 14%. Non-GAAP operating margin 20.3%, up from 17.0%. A new $1 billion buyback. The release is here.
Good quarter. Now the part that moved the stock. HubSpot added roughly 7,000 net new customers against an internal expectation of 9,000 to 10,000, and CFO Kate Bueker guided to just 5,000 to 6,000 net adds per quarter for the rest of 2026. Net revenue retention was 102%, down a point year over year.
The AI adoption numbers are strong: Data Agent at more than 16,000 activated customers (up 80% sequentially), Prospecting Agent around 17,000 (up 28%), Customer Agent above 10,000, more than 55% of Pro and Enterprise customers using agents or Breeze Assistant, and monthly agentic actions up more than 3x since the start of the year.
So adoption is up and customer growth slowed. CEO Yamini Rangan’s explanation on the call is the most useful sentence any GTM vendor said in August: “Customers adopting AI want proof of value before they commit and predictability in what it costs.”
That is the whole market in one line. Buyers are not resisting AI. They are resisting consumption pricing they cannot forecast, attached to outcomes nobody has proven. Every vendor in your stack is currently trying to move you from seats to usage, and the finance team you report to has noticed.
HubSpot shipped Agent Hub and Agent Builder into public beta for all Pro and Enterprise customers on July 23, with more than 2,700 customers activating during the beta quarter, and an Agent CLI that runs inside Codex, Claude Code and Claude Cowork. Chief Product and Technology Officer Duncan Lennox framed the problem precisely: “The problem isn’t managing a single agent in isolation. It’s that once you have multiple agents, they become fragmented, all working from different pictures of the customer, or even worse, no picture at all.” The announcement is here.
UNBOUND runs September 16-18 in Boston and is sold out. Worth stating plainly: no product announcement list has been published, so every “what HubSpot will launch at UNBOUND” post currently circulating is a partner blog guessing. Wait for the Spotlight keynote on the 16th.
If your team is standing up agents on top of marketing automation that was configured for a different business three years ago, Lennox just described your next twelve months.
Microsoft Killed Dynamics Release Waves, and Your Governance Calendar Just Lost a Checkpoint
This is the least discussed and most operationally annoying item of the month.
On August 25, Microsoft announced that starting in September 2026 it is retiring the twice-yearly release wave model for Dynamics 365, Power Platform and Dataverse, moving to continuous publishing on the unified “AI at Work” roadmap. The announcement is here, authored by Richard Riley, GM of Agents and Low Code.
The FAQ leaves no wiggle room. “Will there be a September 2026 release wave 2 announcement or release wave 2 release plan? A. No.” New release plans stop publishing to Microsoft Learn. Release Planner retires by November 15, 2026, and your personalized saved views do not carry over. Microsoft did ship a Release Communications MCP Server so AI clients can query roadmap data conversationally, which is a very modern answer to a problem it just created.
Here is why this matters more than it sounds. Plenty of ops teams built their entire change-management rhythm around release waves. Twice a year, someone read the plan, flagged what would break, scheduled sandbox testing, and briefed the business. It was a free governance checkpoint that Microsoft maintained on your behalf. Continuous publishing means features now arrive whenever they are ready, and the burden of noticing shifts entirely to you.
If nobody on your team owns “watch the roadmap and tell us what changes,” that role now needs an owner, a cadence and an RSS feed, before November. Continuous delivery is only an improvement for organizations with continuous attention. For everyone else it is an operations gap wearing a modernization label.
Gartner: Ten Agents Per Seller by 2028, and Most Sellers Will Shrug
On July 28, Gartner published a prediction with unusual bite: by 2028, AI agents will outnumber sellers ten to one, and fewer than 40% of sellers will say agents improved their productivity. The release is here.
It draws on a survey of 210 CSOs and senior sales executives fielded January to February 2026. Sixty percent of CSOs say their revenue number is largely driven by factors outside their control, which is the kind of admission that only shows up in anonymous surveys. Gartner predicts that CSOs who overhaul data, automation and user experience will be five times more likely to see ROI than those choosing quick fixes.
VP Analyst Dan Gottlieb named the failure mode in eight words: “If those systems are fragmented, the agents will scale the fragmentation.” His other line deserves a plaque in every sales ops war room: “AI agents should not be viewed as a shortcut to sales productivity.”
Gartner calls the risk “agent sprawl,” and it is the same disease as tool sprawl with a faster metabolism. Every team spins up its own agents against its own view of the data. Nobody owns the layer underneath. Six months later you have forty agents, four definitions of a qualified lead, and a pipeline number nobody trusts. The difference from tool sprawl is that tools sit there when unused, and agents keep acting.
The practical implication is boring and correct: agent governance is a RevOps function, not an IT function, and it needs to exist before the agent count gets away from you. Somebody has to own the definitions, the permissions, and the question of what an agent is allowed to do without a human. That is sales enablement infrastructure, not a Slack channel.
McKinsey: Conviction Is Growing Faster Than Returns
On August 25, McKinsey published its State of AI survey, fielded May 4 to June 8 across 1,719 respondents in 97 nations, 36% from organizations above $1 billion in revenue, weighted by national GDP contribution. The report is here.
The headline finding is a flat line. 37% attribute at least some EBIT impact to AI, essentially unchanged from 2025. “AI high performers” remain 6% of respondents, also flat. Meanwhile 80% say AI improved their individual productivity, and agent scaling at large organizations jumped from 27% to 40%.
McKinsey’s own summary: “Organizations’ conviction in AI is growing faster than the immediate financial returns they can attribute to it.”
Two details are worth carrying into your planning. First, revenue gains are most often attributed to AI in marketing and sales, which means the function you work in is where the returns are actually showing up, such as they are. Second, on the jobs question that dominated 2025: only 14% report AI actually contributed to a workforce decline last year, against the 32% who predicted one a year earlier. Prediction ran about 2.3x ahead of reality.
The gap between “80% feel more productive” and “37% can point to EBIT” is not a measurement failure, or not only one. Individual productivity gains get absorbed into slack unless somebody redesigns the process around them. A rep who saves six hours a week on research does not automatically produce more pipeline. They produce six hours. Whether that becomes revenue depends on whether anyone changed the coverage model, the quota, or the territory design to use it.
That redesign work is unglamorous, it is not a software purchase, and it is the entire difference between the 37% and the 6%. It is also why AI integration that stops at tool rollout reliably produces enthusiasm and no EBIT.
Agent Interoperability Consolidated Fast, and Your Stack Is Already Voting
Two structural moves in three weeks, and then a pattern.
On July 28, the Model Context Protocol shipped its 2026-07-28 spec, the biggest breaking change since remote MCP. It retires initialize/initialized and Mcp-Session-Id in favor of a stateless core, adds multi round-trip requests and header-based routing, deprecates dynamic client registration in favor of CIMD, and introduces a formal extensions framework with a 12-month deprecation policy. Named enterprise support includes AWS Bedrock AgentCore, Google Cloud, Microsoft Foundry and Cloudflare. The spec post is here.
On August 17, Google’s A2A protocol became a hosted project of the Linux Foundation’s Agentic AI Foundation, which has grown from under 40 members at its December 2025 launch to more than 250. Axios covered it. AAIF executive director Mazin Gilbert: “Companies don’t want just one protocol; they want the whole stack to be open.”
Now the pattern, which is more interesting than either announcement. In the same window: Pipedrive’s MCP connector landed in Claude’s official marketplace on August 18. 6sense shipped an MCP server into Claude, ChatGPT, Writer and Agentforce on August 10. Microsoft published “Extending agentic Dynamics 365 Sales with MCP” on July 21. HubSpot’s Agent CLI runs inside Claude Code. That is four GTM vendors shipping the same integration surface inside five weeks. That is not a trend claim, it is a count.
What it means practically: the question “which AI vendor do we standardize on” is quietly becoming less important than “is our data accessible through an open protocol.” If your CRM, your intent data and your engagement platform all speak MCP, the assistant layer on top becomes a swappable choice rather than a five-year commitment. If they do not, you are going to be doing custom integration work in 2027 that your competitors will not be.
The stateless spec change also has a real deadline attached. Anything you built against session-based MCP has a 12-month runway. Put it on the roadmap now, while it is a small task, rather than in Q3 2027 when it is an incident. Protocol-level plumbing is exactly the sort of thing that belongs in an AI operations plan rather than in somebody’s browser tabs.
ZoomInfo Wrote Off $650 Million and Pivoted to Being Plumbing
On August 5, ZoomInfo reported Q2: revenue $310.4 million, up 1.2%. A $650.5 million goodwill impairment drove a GAAP operating loss of $622.0 million, taking goodwill from $1,692.7 million down to $1,042.2 million. Diluted EPS of negative $2.19. Adjusted operating income was still $110.0 million at a 35% margin, and full-year guidance went up to $1.207 to $1.217 billion. The release is here.
The number to sit with is net revenue retention of 89%. That is a company whose existing customers spend less each year than they did the year before, growing barely above flat, that just told its auditors a large chunk of what it paid for acquisitions is not worth what it thought.
ZoomInfo’s answer is a pivot to being infrastructure. It launched GTM.AI, a “headless” GTM context layer, with native MCP integrations. CEO Henry Schuck: “Our native MCP integrations with Anthropic’s Claude and OpenAI’s Codex ensure that AI agents are grounded in verified, real-time context rather than static, decaying data.”
That is a smart repositioning and it contains an unusually candid admission from a data vendor: static contact data decays, and selling access to a decaying database is not a durable business when an agent can go get the current answer. ZoomInfo is betting it can be the verification layer instead of the database. Whether that is worth what customers currently pay is exactly the renewal conversation the 89% NRR implies people are already having.
If a contact database is a line item on your budget, this quarter is the prompt to ask what you are actually buying: records, or verified context that your agents can act on. Those are different products at different prices, and only one of them still makes sense in a stack where demand generation runs on live signal rather than a quarterly list pull.
Our Take on the September RevOps News
Strip the vendor language out and every story above says the same thing: the agents are fine, and the substrate is not.
KeyBanc’s CIOs said their enterprise data was not coherent enough to do meaningful AI work. Gartner’s Dan Gottlieb said fragmented systems mean agents will scale the fragmentation. HubSpot’s Duncan Lennox said multiple agents end up working from different pictures of the customer, or no picture at all. ZoomInfo wrote off $650 million and repositioned around the claim that grounding beats storage. Four different companies, four different incentives, one diagnosis.
McKinsey’s flat 37% is what that diagnosis looks like in aggregate. Conviction is up, agent scaling is up, individual productivity is up, and the EBIT line has not moved, because the thing standing between individual productivity and enterprise profit is process design, and nobody sells process design as a SKU.
So here is the uncomfortable read on the fall. The companies that get ROI from agents in 2027 will not be the ones that bought the most agents in 2026. They will be the ones that spent this fall on the boring layer: a single definition of an account and a qualified opportunity, a data model the agents can actually read, an owner for agent governance, and a change-management cadence that survives Microsoft deleting the one Microsoft used to run for you. None of that is on a Dreamforce stage. All of it is what determines whether anything announced on a Dreamforce stage works in your building.
And the pricing question is going to force the issue faster than the technology does. Rangan’s line about customers wanting “proof of value before they commit and predictability in what it costs” is the buyer telling the vendor, out loud, that consumption pricing on unproven outcomes is a bad trade. Salesforce is now reporting agentic work units. That is a consumption meter with a growth chart attached. Before you sign anything this fall, model what your bill looks like if usage triples, because triple is roughly what every vendor is projecting, and figure out what outcome you would need to see for that to be worth it.
Two conferences, one week apart, both promising the agentic future. Go, take notes, enjoy the party. Then come home and fix the data model, because that is the part nobody is going to do for you.