For roughly two years the agent conversation in marketing has been a conversation about recommendations, which is to say about software that proposes an audience, drafts a subject line or flags an underperforming campaign, and then waits for a person to decide whether any of that should happen. The demos were impressive and the category was genuinely useful, but the software remained advisory, and advisory software occupies a fundamentally different place in an organisation than software that acts. Something shifted over the past few weeks. A cluster of marketing platforms shipped agents that take actions carrying real consequences: committing budget, entering commercial arrangements with third parties, writing to systems of record, and putting campaigns live without a person in the loop at each step. The distance between recommending and transacting is the entire distance between a tool and a colleague, and several vendors crossed it within about a month of each other. Every figure that follows is company-reported and none has been independently audited, which is normal for product launches and worth establishing once at the outset.
Upfluence's Agent Negotiates Fees And Releases Payments
The clearest instance arrived on the eleventh of September, when Upfluence launched a fully agentic creator marketing platform built on an engine it calls Jaice, and the company's own description of the workflow is more revealing than any summary of it. The agent understands the brand, discovers creators, conducts outreach, negotiates fees, manages approvals and payments, launches the programme, and reports on performance. Those middle verbs deserve attention, because negotiating fees and managing payments means an agent entering into commercial terms with independent third parties and then settling them, on a brand's behalf, which no marketing software has previously claimed at this level of autonomy. Upfluence puts the resulting timeline at as little as five minutes from campaign brief to live programme, against a process the company describes as typically consuming six weeks and up to twenty hours of manual work, though that is a vendor describing its own best case with no customer named, so it should be read as a ceiling rather than an average.
The supporting data is more useful than the headline speed, because it speaks to whether the output is any good rather than how quickly it arrives. Upfluence reports that brands are twenty-eight percent more likely to end up working with a creator the system selected than with one identified manually, and that outreach written by the system is producing response rates more than twenty percent higher than the human-written equivalent. The base underneath those claims is substantial enough to take seriously: 160,000 creators hired through the previous and less autonomous version of the platform during 2025 alone, across 3,600 brands and a database exceeding fourteen million creators. Co-founder and co-chief executive Kevin Creusy framed the change as one of operating scale rather than speed, which is the right reading of it. A team that could personally manage relationships with forty creators is not being asked to manage those same forty faster.
Mojo Has Live Access To Google Ads And LinkedIn Ads
Demandbase launched Mojo on the fifteenth of September, and the verbs follow an identical pattern. The agent defines audiences, writes campaign briefs, then builds and executes across email, social, digital, webinar and field channels, operating through Google Ads, LinkedIn Ads, Meta, Marketo, Salesforce and Slack, monitoring campaigns while they run and flagging problems such as broken tracking before they damage results. An agent holding live credentials into Google Ads and LinkedIn Ads is an agent spending money, which is the threshold being crossed here and the reason the approval architecture around it matters considerably more than the feature list does.
Demandbase also published a detail most launch announcements omit, and it reveals where the demand actually sits rather than where the marketing department imagines it sits. Across a seventy-two hour sample covering 252 users at 176 companies using the company's AI chat, list building accounted for nearly twenty percent of all prompts, with campaign return and attribution analysis taking a further seventeen percent. The two jobs marketers most want to hand over, on this evidence, are finding the right accounts and working out what worked, which are precisely the two jobs that have historically consumed the most junior headcount. Chief executive Gabe Rogol described Mojo as the first expression of a generation of agentic marketing in which agents carry strategy through to execution, and that phrase, carrying strategy through to execution, describes the whole shift in six words.
Barracuda's 15% Came From Zoom's Existing Product
Zoom announced Forecast and Engage on the fifteenth of September. Engage builds structured multichannel sales sequences across email, phone and other tasks, prioritises seller follow-up, and uses automated triggers to reach prospects at scale, while Forecast turns deal-level signals into a live forecast with drill-down from company commitments to individual opportunities, and writes updates back into the CRM. Writing to the CRM sounds like a smaller act than releasing a payment to a creator, and in practice it is arguably a larger one, because the forecast is the number a sales leader carries into a board meeting. Software that edits it is software operating inside the trust chain of the business rather than alongside it.
In the same release, Zoom cited Barracuda Networks reporting a fifteen percent increase in deals closed, five percent faster deal velocity, and sales representatives reaching between one hundred and one hundred and forty percent of quota. Two clarifications belong alongside those figures, because most coverage has skipped both. The results are attributed to Zoom Revenue Accelerator, the company's existing product, rather than to Forecast and Engage, which launched the same day, so reading them as evidence for the new products would be a mistake. And the release names no Barracuda executive behind the numbers and provides no time period or baseline, which makes this a customer-reported outcome relayed by a vendor. That is worth something, and it is not the same thing as an audited result.
Five Platforms Opened To Outside Assistants In Nine Days
The second half of this shift concerns where the agent is permitted to stand, and the answer has changed quickly. Constant Contact launched a connector inside Claude on the seventeenth of September, allowing its customers to draft campaigns and build emails conversationally without ever opening Constant Contact. Box made its content reachable inside ChatGPT on the tenth, with folder browsing, file preview, document mentions and governed actions, permissions left intact. Klaviyo opened more than 260 tools and 490 APIs to outside assistants on the ninth. Salesforce and Anthropic shipped Claudeforce with thirty-seven prebuilt sales skills routed through Salesforce's own governance layer. Five platforms inside nine days accepted both that the assistant window is a legitimate place from which to run marketing software, and that an agent sitting outside the product should be permitted to act inside it.
Every One Of These Releases Kept An Approval Gate
The consistency on this point across otherwise competing vendors is worth noting, because it suggests the companies building these systems understand what they are asking their customers to accept. Demandbase maintains human approval at critical decision points. Box preserved its permissions model rather than routing around it. Salesforce routes Claude's actions through its existing governance layer rather than granting direct access. Upfluence's workflow places an approvals step between the negotiation and the payment. Godard Abel, co-founder of G2 and quoted in Demandbase's launch materials, drew the line where most practitioners would draw it, with operational work shifting to agents while strategy, creativity and judgement stay with people. That division is holding at the moment, and how long it holds is the more interesting question for the next several quarters.
JetBlue Compressed Report Building From Weeks To Summaries
Alongside the agents that act, the instrumentation to observe them has started arriving. Sprout Social's agent Trellis has been generally available since July 2025, but the company reached a different kind of milestone on the nineteenth of August when it published what customers actually do with it rather than what it is capable of doing. JetBlue's social media lead described report building compressing from weeks into executive summaries, while Ipsy uses it to monitor member feedback on new launches and route sentiment analysis to product and care teams. Named people at named companies describing the work they have stopped doing is a more durable form of evidence than most of what this category produces, precisely because no percentage is attached to it.
Two further September releases point at where the next budgets go. Comscore launched AI Intelligence on the third, measuring sponsored chat-based advertising and connecting it to traffic and behaviour, and its data shows sponsored placements within hotel-related ChatGPT prompts rising from six percent in March to fourteen percent in April and twenty-four percent in May. Emberos launched on the same day with a brand knowledge graph running five agents and a $5.5 million seed round, claiming eighty percent directional accuracy across more than three hundred forecasts of how brands will be represented inside AI systems. That is early work from an early-stage company and should be treated as such, but forecasting AI representation is a discipline that did not exist eighteen months ago and now has funded competitors in it.
Three Categories Of Work Moved From Assisted To Autonomous
Creator programme execution including its commercial terms, campaign construction and launch across paid channels, and the reporting and forecasting layer all crossed from assisted to autonomous inside roughly four weeks, and each of those has historically been answered with headcount rather than software. The question arriving in planning conversations over the next quarter is therefore not whether any of this works, because enough customers are now on record to suggest it does something. It is which of these categories a given organisation is prepared to let software handle without a person in the middle, and what the approval architecture around that decision looks like in practice. An agent that can negotiate a fee and release a payment belongs to a different category of thing than an agent that drafts an email, and September was the month several vendors decided their customers were ready for the first kind.
