Two conferences said the same thing, and neither one said what it costs

Last week's issue flagged that Dreamforce and HubSpot UNBOUND were landing inside four days. Both landed, and both gave the same answer to the same question. Salesforce launched AIforce on September 16 and Salesforce Ben summarized the pitch in four words: AI replaces the UI (Salesforce Ben). HubSpot, on the same day in Boston, named the thing that feeds the agent Growth Context and shipped a product that scores how complete yours is (HubSpot release). One vendor moved the surface. The other moved the substrate. Neither published a cost-to-run figure.

This is the same pattern the September 14 issue tracked with Klaviyo and Box, now arriving at the two platforms most B2B revenue teams actually sit on. The difference this week is that the argument got specific enough to test. Salesforce is saying your CRM data, permissions, business logic and governance should travel into Claude, Slack and Amazon rather than sit behind a Salesforce login. HubSpot is saying an agent is only as good as the written description of your business, and here is a page that grades it.

Both claims are testable, and both vendors published numbers that do not survive a careful read. HubSpot says Professional and Enterprise customers using AI with high-quality context create 3.6x more marketing qualified leads, win 3.2x more deals and close over 2x more tickets. The footnote compares those customers against Professional and Enterprise customers not using AI at all, with no sample size, no time period and no control for how the two groups were performing before (The Agile Brand Guide). Salesforce says its new Koa model matches or exceeds leading general-purpose models on CRM actions with three times fewer errors, on a benchmark Salesforce designed and administered (Tech Times). A team that already keeps clean records starts ahead of one that does not, and a vendor grading its own exam is not a benchmark.

Then, at 12:50 AM Pacific on the second morning of its own conference, Salesforce went down across hundreds of instances in more than a dozen countries, with an internal login service named as the cause (Salesforce Ben, The Register). That is not a gotcha. It is the argument. If the plan is that agents in Claude, Slack and Amazon all call back through one platform for permissions and business logic, then that platform's login service is now a single point of failure for work happening in three other places.

So the useful conversation this week is not which keynote was better. It is three questions a buyer can ask before a renewal: what does the context layer actually contain and who maintains it, what happens to the agents when the system of record is unreachable, and what does a run cost per unit of work. Nobody on stage answered the third one.

Every item below carries a source link. Where a figure comes from a vendor describing its own product, its own customers, or its own benchmark, we label it a claim and attribute it.

The big picture

Three shifts to take into your next pipeline call

The interface layer now has a vendor name and a price attached. Salesforce packaged it as AIforce, with Claudeforce, Slackforce and Agentforce Coworker as the first three products, carrying Salesforce data, permissions and governance into tools people already use (Salesforce Ben, Salesforce Break). Two weeks ago this was an architectural trend. It is now a line item, and if you are on Salesforce you will be asked to buy it.
Context is being sold as a product, which means it can be scoped as a deliverable. HubSpot's Context Home scores business, customer and team context and shows where the gaps are (diginomica). Whatever you think of the score, the framing is useful: the ideal customer profile, the competitor list, the sales methodology and the product descriptions are now inputs a vendor will grade. Most teams cannot fill that page in without help.
Every outcome number published this week compares adopters to non-adopters. HubSpot's 3.6x, Google's 5% feed lift, Pindrop's 94.3% bot reduction at one customer, StoreClaw's 133% connection growth. None of them published a holdout. The buyer who asks for the control group is going to be the only one in the room who did, and that question is what protects your budget.

Dreamforce and the interface layer

Salesforce says the UI is the thing being replaced

Platform

AIforce lands, and Claudeforce is the part with 37 sales skills in it

Announced September 16 at Dreamforce, September 15 to 17, Moscone Center, San Francisco. Coverage: Salesforce Ben, Salesforce Break, ITPro, Salesforce newsroom.

Vendor claimSalesforce describes AIforce as a live interface layer above Agentforce, Data 360 and Customer 360 that carries Salesforce data, workflows, business logic, permissions, security and governance into the AI tools people already work in. The first three products are Claudeforce, which puts a prebuilt Model Context Protocol server with 37 ready-made sales skills inside Anthropic's Claude, Slackforce, which brings Salesforce context into Slack conversations, and Agentforce Coworker, described as available now. Salesforce said 30,000 customers are live on the Agentforce platform.

The August 31 issue flagged Claudeforce when it was a partnership announcement with an expected September beta. It shipped roughly on schedule and is now the flagship of a named product family rather than a side deal.

Why it matters: If you are on Salesforce, this turns an architecture question into a purchase decision you have to make this quarter. The real work is not buying AIforce. It is the permission and audit work underneath it: which of the 37 skills should be enabled, for which roles, with what write access, and what record proves what an agent did inside Claude. That work does not exist in the box.

Model

Koa is a CRM-specific model, graded on a test Salesforce wrote

Announced September 16. Coverage: Salesforce Ben, Inside AI News, Tech Times.

Vendor claimKoa was built with NVIDIA by post-training the open Nemotron 3 Super model on a synthetic dataset modeled on 27 years of CRM deployments across more than 14 industries. Salesforce says Koa matches or exceeds leading general-purpose models on CRM actions with three times fewer errors, that it runs entirely inside the Salesforce trust boundary, and that it is in pilot now with general availability expected in US regions this winter. Salesforce also announced Gemini inside the Agentforce Reasoning Engine and seven named agents: Casey for service across voice, SMS, WhatsApp and chat, Paige for IT and HR, Carter for ecommerce, Marshall for supply chain, Piper for inbound sales qualification, Fin for complex experience workflows, and Hunter for outbound sales.

The benchmark is Salesforce's own, which Tech Times named directly. That does not make the model bad. It makes the number unusable as evidence until somebody outside Salesforce runs it.

Why it matters: The trust-boundary claim is the part worth holding onto if you work in a regulated industry, because it is the first concrete answer to "where does my CRM data go when the agent reasons over it." The error-rate claim is not evidence yet. Use the first, park the second, and ask your Salesforce rep who else has run the benchmark.

Reliability

Salesforce went down across hundreds of instances on day two of its own conference

September 16, beginning 12:50 AM Pacific per Salesforce Trust. Coverage: Salesforce Ben, The Register, Quartz, Channel Insider.

Salesforce acknowledged the disruption at 12:50 AM Pacific and confirmed shortly before 2 AM that multiple instances across all regions were affected. Reporting put the footprint at hundreds of instances spanning the United States, Japan, India, the United Kingdom, France and Germany, with further reports from Australia, Brazil, Canada, Switzerland, Indonesia, Italy, South Korea, Singapore and Sweden. Salesforce said requests were stalling while waiting on an internal login service, which consumed available server resources. A fix was identified, tested and deployed across regions by mid-morning Eastern.

Why it matters: Use this carefully and without glee, because it is a real argument rather than a cheap shot. The entire AIforce thesis routes agent actions back through Salesforce for permissions and governance. That is the right architecture and it concentrates risk. If you are putting agents in Claude or Slack against Salesforce data, you need a written answer to what those agents do when the system of record is unreachable: fail closed, queue, or proceed on stale context. Most agent deployments have no defined behavior for that state at all.

Return on investment

The agent count went up, the return figures did not appear

Coverage: MarketScale, diginomica, Salesforce Ben on the TD Cowen partner survey. Context piece: Salesforce Ben, September 18.

Third-party surveyThe TD Cowen partner survey that framed the run-up to Dreamforce, published in late August and outside this coverage window, found customer interest in Agentforce rising, with about a third of partners reporting buying and trial activity, and no partner saying Agentforce had influenced any bookings. MarketScale reported that Dreamforce pre-event materials carried no customer return figures and no cost-to-run numbers, leaving procurement and operations teams to ask the measurement questions themselves. Separately, Salesforce said on September 18 that 20,000 of its own employees now sit in roles that did not exist a year ago, clustered around what it calls the builder archetype.

Why it matters: Two useful things here. First, the gap between trial activity and booked revenue now has a named survey describing it, which makes it easier to raise inside your own company. Second, the 20,000-roles figure is a hiring signal worth quoting: the vendor reorganized its own workforce around building and evaluating agents rather than administering software. That is the same shift the GTM engineer postings have been showing across the market.

HubSpot UNBOUND and the context layer

HubSpot named the input, then built a page that grades it

Platform

Growth Context, Context Home, and a CRM that updates itself

Announced September 16 at UNBOUND 2026, September 16 to 18, Boston. Coverage: diginomica, The Agile Brand Guide, HubSpot release, TechTarget.

Vendor claimHubSpot introduced Growth Context as the layer of business, customer and team data that feeds its agents, and Context Home as a single place to see and manage it, with a completeness score and named gaps. Inside Context Home a customer fills in business profile, brand kit, technology stack, competitors, products and services, ideal customer profile and personas, sales methodology in private beta, custom context, and uploaded knowledge. HubSpot also shipped a self-updating Smart CRM that captures and syncs calls, emails and meetings, a rebuilt Breeze Assistant that assigns work to specialized agents, Marketing Studio in public beta with an answer engine optimization visibility score, an updated Prospecting Agent monitoring more than 40 buying signals, a Mobile Notetaker, revamped Deal Progression, and automated quoting in a new Revenue Hub.

HubSpot's outcome figures: 3.6x more marketing qualified leads, 3.2x more deals won, over 2x more tickets closed, plus 81% more campaigns created by early Marketing Studio customers and 2.2x more leads for Breeze Assistant customers. The first three compare AI users with high-quality context against customers not using AI. The last two carry no baseline or measurement period. diginomica's reviewer also noted that Breeze Assistant asks a lot of questions as it works, which is good practice and means it still needs directing.

Why it matters: This is the most useful thing in either keynote. HubSpot has now published, as a product, the exact list of written inputs an agent needs to be worth anything: ICP, personas, competitors, methodology, products, brand. That list is a to-do list. If your Context Home is mostly empty, the positioning and qualification work has not been written down yet, and HubSpot's own footnote is the budget argument for doing it, because the figure rewards context quality rather than license count.

Paid media

HubSpot puts ChatGPT Ads inside the CRM, and bundles the seats

Announced September 16. Coverage: CMSWire, diginomica, Innovation Visual.

Vendor claimHubSpot says it is the first CRM to integrate with ChatGPT Ads, letting marketers build, manage and measure those campaigns from HubSpot with contacts flowing into landing pages and workflows. Microsoft Advertising and Reddit Ads now connect the same way Google, LinkedIn and Facebook already did. HubSpot and OpenAI also announced an AI Growth Bundle for small and mid-sized businesses: HubSpot Starter and credits at up to 65% off for one year, buy one get one free on ChatGPT Business seats for a year, and a 750 dollar match on ChatGPT Ads spend for new accounts opened by September 30.

The September 14 issue noted that Amazon opened its demand side platform inside ChatGPT without publishing an impression count, deduplication method, attribution model or pricing. That is still true, and this integration does not change it. What it changes is that the buying is now one click from a CRM you already pay for.

Why it matters: The September 30 deadline is a real, dated reason to decide this week if you are on HubSpot Starter. The discipline point stands: putting the buy inside the CRM makes the spend easy and does not make the measurement exist. Write the baseline, the test window and the definition of a result before anyone matches the 750 dollars.

ABM

Demandbase ships Mojo, an agent that runs campaigns across other people's tools

Announced September 15. Coverage: MarTech Series, CMSWire, Yahoo Finance.

Vendor claimDemandbase describes Mojo as a business-to-business marketing agent that learns a company's audiences, campaign structures and revenue process, then defines audiences, develops briefs, and launches and manages campaigns across connected tools including Google Ads, LinkedIn Ads, Meta, Marketo, Salesforce and Slack. Demandbase says Mojo detects tracking errors and adjusts execution using performance data from previous campaigns. No customer outcome figure was published with the launch.

Why it matters: Note what Mojo is not: a chat window inside Demandbase. It is an agent with write access to ad platforms, a marketing automation tool and a CRM. That is the same permission question as Claudeforce, arriving from the account-based marketing side. Before you turn this on, somebody on your team has to decide which campaigns it can launch without approval and what the spend ceiling is. Nobody ships that decision with the product.

Advertising, AEO, and measurement

Google graded the feed, and the grade was partly conditional

Agentic commerce

Google opens AI performance insights in Merchant Center, and publishes a number worth reading twice

Announced September 16 by Ashish Gupta, VP and GM of Merchant Shopping. Source: Google. Analysis: The Agile Brand Guide.

Vendor claimGoogle made AI performance insights generally available in Merchant Center for Australia, Canada, India, New Zealand and the United States, comparing a retailer's share of voice with other brands across AI Mode and AI Overviews, and opened a United States beta embedding Business Agent inside YouTube ads. Google added cart transfer to a merchant site and checkout flow testing to its Universal Commerce Protocol integration hub, rolling out in the United States first with Australia and Canada early next year, and said the protocol already enables direct checkout for hundreds of thousands of brands and retailers. Google stated that merchants adopting core Merchant Center feed best practices see a 5% average increase in conversions the following month, and that conversational attributes submitted by lululemon appeared in relevant AI Mode product recommendations 50% of the time during testing.

Read those two figures together. The 5% is an average across merchants with no sample or baseline disclosed. The 50% comes from testing with a single brand, and it means the retailer writes the full attribute set and gets used on roughly half of it.

Why it matters: Share of voice inside AI Mode and AI Overviews is now reported by Google itself, inside a tool retailers already have. If you sell online, that is a free measurement baseline that did not exist before, and it should be pulled before the next answer engine optimization tool gets bought. The feed work is also the cheapest AI visibility work available, and it is owned by your own team.

AEO

Azoma publishes the buyer's test for an AI visibility tool, which is a competitor's test too

Published September 14. Source: GlobeNewswire, Yahoo Finance. Related, September 10: MarTech Series.

Vendor claimAzoma, which sells in the category, proposes a practical test for any answer engine optimization platform: can the tool show you, prompt by prompt, where your brand appears and where it does not across ChatGPT, Gemini, Amazon Rufus, Alexa for Shopping and Walmart Sparky at the same time, and can it tell you which sources each agent drew on. Azoma describes its own coverage across those surfaces plus AI Overviews and Perplexity, with citation-source analysis, competitive benchmarking and product-data optimization workflows.

This is marketing material with a usable checklist inside it, which is the same pattern the August 17 issue noted with Glean and Fin. Take the criteria, ignore the conclusion.

Why it matters: Prompt-level coverage and citation-source attribution are the two questions that separate a real AI visibility tool from a dashboard that reports a single invented score. Both belong in any evaluation you run, and neither requires trusting the vendor who wrote them down first.

Ad infrastructure

Magnite extends its orchestration layer, and whoever describes the inventory sets what agents can find

Announced September 16. Source: GlobeNewswire. Analysis: The Agile Brand Guide.

Vendor claimMagnite says Magnite Orchestration now describes inventory for buyer and seller agents across local linear TV through ITN, addressable TV through partners including DIRECTV Advertising, streaming home screen and pause units from Samsung Ads, live audio and podcast inventory from iHeartMedia, and custom audience packaging across web and mobile display and video. Magnite describes it as a neutral layer where agency holding companies and software vendors such as Newton Research bring their own optimization models, with interoperability with Databricks. Magnite published no performance figures.

Why it matters: One sentence to bring to your agency: a buying agent can only act on inventory somebody has described in a format the agent reads, so the party writing the description sets the menu. That question, who writes the inventory description your agency's agent buys from, is not on anybody's quarterly business review agenda yet.

RevOps and governance

Three releases about telling machines apart from people

Measurement hygiene

Fingerprint publishes a public bot directory, and it lands on two budgets at once

Announced September 18. Source: Business Wire via FinancialContent. Analysis: The Agile Brand Guide.

Vendor claimFingerprint launched Bot Directory as a continuously updated, searchable public repository of AI tools, bots and crawlers, listing each by identity, provider and purpose, with a way for operators to test their own agents for Web Bot Auth compliance and submit them. Fingerprint reports identifying more than 1 billion unique devices a month across more than 6,000 customers. Its release states that bots account for the majority of web traffic and that nearly half of technology leaders face AI-driven fraud attacks, and cites no source for either figure.

The Agile Brand Guide made the operator point better than the release did. One allow-and-deny list decides two marketing outcomes: a crawler you block never reads the page you optimized for answer engines, and an automated session your analytics counts inflates the denominator under every conversion rate you report. In most companies security owns that list and marketing owns the distorted number.

Why it matters: This is a small, concrete diagnostic. Pull the share of sessions your analytics classifies as automated, recompute the conversion rates you have been reporting to your board, and name the owner of the allow-and-deny list. It takes a day, it produces an uncomfortable number, and it is the cleanest starting point for a broader data-quality cleanup.

Procurement

A free, ungated agent procurement handbook, with the three questions marketing should already be asking

Published September 18. Source: GlobeNewswire.

Handvantage, a consultancy in British Columbia, published a 43-page vendor-neutral guide to evaluating AI platforms and agent governance before signing, with no email form. The framing is that procurement teams now field questions that used to belong to security architects: where company data goes, who approves what an agent does, and what record exists after the agent acts. Handvantage sells advisory services in the category it is writing about, which makes this marketing material containing useful questions.

Why it matters: Those three questions are the agent register the September 14 issue recommended building, arriving from the procurement side. Marketing signs more software contracts than most functions, so this belongs in the marketing contract review rather than waiting for a governance committee. It is also free and ungated, which makes it easy to forward to your procurement team.

Service operations

Pindrop ships detection for AI agents calling your contact center

Announced September 16. Source: GlobeNewswire.

Vendor claimPindrop launched BotStopper as a standalone product detecting AI voice agents and automated callers in real time, with an AI Voice Consortium registry of more than 5,000 AI voices and detection in roughly two seconds. Pindrop reported that bot calls at one healthcare organization peaked at 6,668 in a single month with approximately 15.1 million dollars in health savings account balances exposed, and that bot activity fell 94.3% within four months of deployment. Pindrop also reported AI-driven attacks rising nearly 1,680% since late 2024 across more than 700 million customer interactions. The 94.3% figure comes from one named customer.

Chief executive Vijay Balasubramaniyan framed the problem correctly: legitimate agents will call enterprises on behalf of real customers, so the task is knowing which agent is calling and what it is allowed to do.

Why it matters: This is the inbound mirror of everything else in this issue. If you expect buyers to send agents to research, return or reorder, then your contact center and your web forms need a written policy for agent callers before you need a detection vendor. The policy is the deliverable. The software is downstream of it.

Plumbing worth knowing about

Two deals that change whose data proves your results

Infillion agreed to acquire Foursquare on September 18. The combination puts Foursquare visitation data, covering more than 100 million points of interest across 200 countries, 16 billion check-ins and aggregated coverage of 250 million United States devices, alongside Catalina purchase data (ACCESS Newswire). Founder Rob Emrich told Axios the deal lifts Infillion revenue and headcount by roughly 20% to 25%. The release published coverage figures and no lift figure, no holdout design and no price. Store-visit measurement counts device presence after an exposure, which is a correlation between two logs. Ask for the test design at renewal, not the panel size.
Taboola made a Rule 2.7 offer for Dianomi on September 18. Dianomi works with more than 600 advertisers including Charles Schwab, Invesco and Bank of America, placing with publishers including Reuters, CNN Business, The Times and The Wall Street Journal, and would run through Taboola's Realize platform (GlobeNewswire). BusinessCloud reported the deal at up to 27 million pounds with Dianomi first-half 2026 revenue of 13.4 million pounds; Taboola's own release carries no price. Completion is expected before the end of 2026. If you run finance-vertical media through Dianomi, confirm publisher lists, rate cards and reporting continuity during the transition.
Quantum Metric moved Voice of Customer to general availability on September 16, connecting surveys, chat, tickets and contact center interactions to behavioral session data, and said beta partners used it to increase feedback on review sites commonly cited by AI systems (GlobeNewswire). That last detail moves review generation from a reputation line into the answer engine optimization budget, which is a reframe worth using.

The week ahead

One deadline, one event, one release to read

The HubSpot and OpenAI AI Growth Bundle closes September 30. The 750 dollar ChatGPT Ads spend match applies to new accounts opened by that date, alongside up to 65% off HubSpot Starter for a year and buy one get one free on ChatGPT Business seats (diginomica). If you are going to test ChatGPT Ads at all this year, the test design needs writing this week, not the purchase.
Amazon Accelerate, September 22 to 24, Seattle. Expect agentic commerce and listing announcements aimed at sellers heading into the holiday peak. StoreClaw reported store connections up 133% and Amazon Ads account connections up 194% after the August 5 Prime Big Deal Days early-bird deadline, which describes its own user base and includes no absolute counts or sales outcome (GlobeNewswire).
The Salesforce Winter '27 release notes are the real Dreamforce document. Keynote naming and release scoping rarely match, and the gap between AIforce as announced and AIforce as generally available is where expectations get set wrong (Salesforce Ben). Read the release notes before anyone on your team promises anything.

Put it to work

What to do with this, this week

Turn Context Home into a scoped deliverable. HubSpot just published the exact list an agent needs: ICP, personas, competitors, sales methodology, products, brand, custom knowledge. Fill it properly, with the positioning work done rather than guessed. It is a two-week project, not a quarter. HubSpot's own footnote, which rewards context quality rather than license count, is the budget argument.
Decide what your agents do when the system of record is down. Salesforce was unreachable across hundreds of instances for part of September 16. Fail closed, queue, or proceed on stale context are the three options, and almost nobody has picked one in writing. This is a short document, and most teams have not written it.
Run the automated-traffic check this week. Pull the share of sessions analytics tags as automated, recompute the reported conversion rate, and name the owner of the bot allow-and-deny list across marketing and security. Fingerprint's directory is the public reference. The number will be uncomfortable, which is the point.
Pull Merchant Center AI performance insights before buying any AI visibility tool. If you sell online, Google now reports share of voice across AI Mode and AI Overviews inside a tool you already have. Establish that baseline first, then evaluate paid platforms against Azoma's two criteria: prompt-level coverage and citation-source attribution.
Put the cost-to-run question to every agent vendor. Not one vendor this week published what it costs to operate an agent per unit of work, and MarketScale named the omission at Dreamforce directly. Asking for cost per resolved ticket, per qualified lead, per campaign launched is the fastest way to separate a product from a demo.

Every claim above carries a source link. Figures attributed to vendors, to companies describing their own customers, or to benchmarks a vendor designed and administered are their claims, not independently verified facts, and are labeled as such. Items sitting slightly outside the coverage window are included only where they frame something inside it and are dated in the text: the TD Cowen partner survey on Agentforce bookings was published in late August, and the Azoma generative engine optimization comparison referenced alongside its September 14 piece was published September 10. Dreamforce ran September 15 to 17 and UNBOUND ran September 16 to 18, so some product detail below the keynote level may still change before general availability. Coverage window: September 14 to September 21, 2026. Compiled September 21, 2026.