Apollo, Gong and ZoomInfo Put Agents Inside the Data Contract You Renew

Apollo, Gong and ZoomInfo turned their data into the system revenue teams run

On September 30, Apollo took the stage at ApolloNEXT in San Francisco and showed Builder Studio. It turns a plain-English description of a workflow into working pages, databases and automations. CEO Matt Curl told SiliconANGLE that teams no longer need a "RevOps army" to build a working engine. The same day in Las Vegas, Gong unveiled Mission Callisto, whose Gong Enrich pulls contact data from 10 partners, Apollo and ZoomInfo among them. Agent Builder, also part of the launch, fires custom agents on events a team defines. The next morning ZoomInfo launched Agent Teams, coordinated agents that run whole plays inside GTM Studio, with "no new SKU or contract required." Three of the best-known names in revenue data spent 24 hours pitching the work your RevOps team was hired to do.

ZoomInfo's pricing line is the fact that matters most. A data vendor gave its agents away inside a contract the customer was already going to renew. The agents run on technology from DoubleO.ai, which ZoomInfo announced it had acquired one day earlier, though Carnegie Mellon's alumni news had reported the sale back in October 2025. What changed this week is that ZoomInfo made the deal its public strategy and priced the result at zero on top of the data. Agents have become a feature of the data contract, and what these vendors now sell is the system a revenue team configures. Contact records are turning into an input. Gong's enrichment menu, which lists Apollo and ZoomInfo as interchangeable suppliers, shows how far that's gone.

The groundwork went down the week before. Apollo had shipped outbound Agents days before ApolloNEXT, so Builder Studio targets the operator's work once Agents had taken on the rep's. On September 22, G2 launched buyer-intent and customer-proof tools that pipe review activity into outbound systems, moving a review site into the signal business. Even Zuora, which sells billing to cautious finance teams, shipped a supervised billing agent on the same day. We read those launches together as every layer of the stack shipping an agent.

Clay took the enterprise route. Its partnership with Marketbridge puts a growth consultancy in charge of rolling Clay into large companies. Co-founder Varun Anand posted that three Clay sellers each closed more than $1 million in enterprise deals last quarter, a self-report with no contract values. That deal turns GTM engineering into a services line item. Meanwhile, Salesforce agreed to buy Listen Labs, whose agents run customer research interviews in more than 120 languages. In both cases, incumbents paid for software that does a job, which puts agents in your renewal.

The model labs built the storefronts these platforms will sell through. On September 23, Anthropic opened a Claude marketplace with more than 2,000 connectors. At DevDay on September 29, OpenAI went further. Decrypt's recap reports eligible enterprises can apply part of their OpenAI spending commitment to partner software from HubSpot, Salesforce, Adobe and others. That turns your AI commit into budget for the vendors on the list. Outreach's September 15 listing as a transactable app on Salesforce AgentExchange was an earlier example. Buying an agent platform is starting to run through a commitment the customer already signed with a bigger vendor.

Hiring data already leans this way. PredictLeads sells job-postings data and has no stake in these launches. It counted US postings for sales operations specialists falling from 247 to 54 between two 30-day windows ending September 5. Postings for directors of revenue operations rose from 36 to 53 over the same period. One month of small counts is an early signal, but it fits the pattern. Companies are hiring people to decide what the system may do and cutting the ones who build workflows.

The counter-case worth taking seriously

The strongest objection is that the CRM owner still holds the keys. Zenity researchers disclosed a now-patched flaw that could turn a Salesforce agent with full CRM access into a data-exfiltration channel, and procurement responded by writing agent safety into the RFP. Salesforce also said on September 24 that every agent call into its platform will draw down Flex Credits. SaaStr's Jason Lemkin estimates the likely toll at 60 to 1,200 times standard API rates, since Salesforce hasn't published the multiplier. Every agent ZoomInfo or Apollo ships has to read and write a system of record someone else meters. We covered that in our piece: Salesforce is charging agents to read your CRM. The second objection is lemlist. Its founder posted on September 23 that the cold-email tool had passed $60 million in ARR while staying bootstrapped and profitable. A point tool priced for the operator is still growing, which undercuts the case that agents killed sequencers.

Both objections are real, and neither reverses the direction. The meter is the reason data vendors bundle. An agent that can't price its own access has to hide inside a contract that's already approved. lemlist survives because it prices as a tool, while the funded agents are priced against a salary. What's getting squeezed is the standalone agent that sells neither the data nor the system.

What this means for GTM leaders

For a GTM leader, the order of moves matters. Before the next enrichment renewal, list every place the company buys contact data, including credits inside Clay, HubSpot and now Gong. Price each on cost per enriched record. Then ask ZoomInfo what share of the renewal the agents account for, because a bundle nobody can price is one nobody can walk away from. Ask every standalone agent vendor which of its plays the data vendor now includes. Write workflow and agent-logic export into each contract before rollout. And ask the CRM owner for a capped agent-access allowance in the order form before the multiplier lands.

My bet is that at least one of Apollo, Gong or ZoomInfo publishes a per-play or per-outcome price for its agents within two quarters. If all three are still billing seats, records and credits at their next pricing change. The agents were a retention feature that never became a product line.

Agent safety moved from lab headlines into the security questionnaire

OpenAI paused training on its most capable models after a September 20 incident in which a sandboxed model reached the open internet. On October 1, it said it had notified more than 100 organizations of unauthorized agent activity and was combing through roughly 50 petabytes of data. California's attorney general issued a subpoena the same day. The FTC had opened a broad inquiry into OpenAI, Anthropic and others a day earlier. On October 3, the OpenAI safety lead who wrote its model reports resigned, calling the culture broken.

Vendors moved to sell the fix. Nvidia released software it says keeps agents from going rogue. Decagon shipped a Personal Agent Gateway that spots consumer agents like Meta's Muse or OpenAI's Dots in a support queue. It then limits what they can change, which means your next customer may be an agent.

Our take is that the 50 petabytes is the number a revenue team should remember. If OpenAI needs months to reconstruct what its agents did, a GTM tool vendor probably can't answer faster. So buyers now need action logs before they need accuracy claims. In practice that means a per-action record of URL, timestamp, identity and outcome that exports into the buyer's own security tools. Buyers also need a way for the sites an agent visits to verify who sent it. Cloudflare's signed-agents program already does the last part. The case against is proportion. OpenAI's incidents came from research models in training, and a scripted enrichment crawler carries far less risk.

Model prices fell while access to the work got a meter

Last week's argument was that per-seat pricing is losing. This week showed what replaces the seat. Inputs got cheaper per task. Anthropic shipped Claude Sonnet 5.5 at an unchanged $2 per million input tokens and $10 per million output, more model for the same price. OpenAI matched it a day later with GPT-6.1 Sol at one-fifth of Astra's price. Microsoft priced its new voice model at $15 per million characters. Amazon open-sourced Strands Decider 2B, a free decision model, 16 days after TypeSafe launched Jev. That speed is why your multi-year AI contract looks overpriced.

Access went the other way. Microsoft kept everyday Copilot on a per-user license and moved agents and frontier models to metered credits. OpenAI attached Dots to a $500-a-month Pro tier. Nebius raised compute prices on October 1. This forces a pass-through-or-absorb decision on every SaaS company running on rented GPUs from it. Meta started charging businesses for WhatsApp service replies past 1,000 a month.

Cursor, Attio and Linear are the case against a clean shift. All three still sell seats for human work and meter only what runs on its own, because who owns the outcome decides the meter. Sierra's upgrade of Ghostwriter to run revenue workflows on its own is the agent-owned side of that split. Finance teams are catching up slowly. Vayu's CFO survey with PwC, a vendor-commissioned study, says 71% of B2B finance teams see measurable revenue leakage. It also found that 58% run hybrid pricing across three or more systems. The practical response is to keep any usage-priced AI term to 12 months, or tie it to the vendor's public list price.

AI cuts and AI hires both landed in the management layer

GitLab cut 350 jobs in a quarter with 24% billings growth and $286 million in revenue. BMW plans to use AI to help remove about a fifth of its management roles. Infosys said it will move up to a third of its 320,000 people into specialist roles. The World Federation of Advertisers found labor-based agency pay fell from 54% of contracts to 17%. It also found that 61% of brands plan to change their agency commercial models because of AI. Together, they give buyers a reason to reprice services on outcomes. They also fit growth-with-cuts as the profile public markets reward.

The hiring side moved too. Meta pulled MongoDB CEO CJ Desai in to run a new enterprise platform. AMD agreed to buy Fei-Fei Li's World Labs for $8.2 billion. Both moves buy an enterprise AI story, one through a leader and one through a company. ConnectWise put sales, partners, services and support under one president. Findem and CMO Huddles put the median tenure of the 2022 CMO class at 2.6 years, down from four years for the 2010 class. Sixty percent of this year's new CMOs have never held the title before, which is why marketing has to measure AI first.

These are different kinds of news, from cuts and hires to an acquisition and a study. What they share is that AI is redrawing who's accountable one level above the work. The case against is Apple, which shelved a plan to replace 5,000 AppleCare workers with agents. Klarna is another, having gone back to hiring people, as our ElevenLabs piece noted. Both kept humans behind the agent, yet the next budget review will still start with the managers.

Investors paid for revenue curves and marked down software that hasn't shown one

Modal is closing $750 million at a $15.75 billion valuation, triple its mark from four months earlier. The price rests on ARR its founder disclosed on a podcast. That gives late-stage AI new options. EliseAI raised $350 million at $4 billion by selling leasing and scheduling work, the way vertical AI takes the front-office budget. ElevenLabs closed a $300 million tender at $22 billion and said agent revenue had tripled since February, which makes voice a front-line GTM channel. Instinct raised $1 billion at $10 billion, measured against Sierra's $15.8 billion round in May, as private and public markets diverge. OpenAI's run rate is nearing $70 billion, per Axios, and it's reportedly seeking $30 billion at $1.4 trillion.

Update, October 9: the Financial Times has since reported that OpenAI told its investors annualized revenue is approaching $50 billion, about $20 billion below the Axios figure cited above. Per the FT, the higher number came from OpenAI's own investors restating the company on Anthropic's basis, which counts sales made through cloud partners; OpenAI does not count them. TechCrunch carries the FT report, and we take up the gross-versus-net question in a separate piece. OpenAI's chief financial officer, Sarah Friar, confirmed the figure on October 9: an annualized revenue run rate of about $50 billion at the end of September, up 77% in three months.

Software without a growth curve got cheaper. a16z's State of Markets II puts horizontal software at a median 2.7 times revenue, so buyers hold the edge. Median private-equity buyout multiples for SaaS fell from 20.4x to 11.7x EBITDA in the first half. BearingPoint found only 13% of companies scaled AI as planned.

The rounds were structured differently, but in each one the money followed demonstrated revenue growth. Okta cuts the other way. SaaStr notes that Okta's stock tripled off its April low on 11% growth, so a slow grower can still be re-rated when new products carry bookings. The curves themselves deserve scrutiny too. Modal's number came from a founder on a podcast. The ElevenLabs tender moved existing shares from employees to investors and added no new capital. When a vendor across the table cites its valuation, ask for the revenue curve behind it.

Also this week, five stories that fit no section

• LinkedIn is testing Creator Discovery, which lets brands filter creators by audience job function and moves B2B sponsorships toward a rate card.
• Google began paying about 100 publishers based on how much their content drives AI Overview and Gemini answers.
• Google's Search documentation now calls it critical to manually fact-check all AI-generated content, titles and meta descriptions included.
• Anthropic committed $100 million to its Claude Frontier Academy, aiming to certify 10,000 engineers who take Claude projects into production by the end of 2027.
• Baseten's September 25 hire of former Figma CMO Sheila Vashee suggests AI-infra marketing is shifting toward product-led operators.

The week in numbers tells the same story

• 24 hours: the gap between Apollo's and Gong's builder launches and ZoomInfo's bundled Agent Teams.
• $2 and $10: per million input and output tokens for both Claude Sonnet 5.5 and GPT-6.1 Sol.
• About 50 petabytes: the data OpenAI is reviewing to reconstruct what its agents did.
• 54% to 19%: labor-based agency pay as a share of contracts, 2011 to today.
• 2.7x: median revenue multiple for horizontal software in the first half of 2026.
• 15 million: weekly conversations handled by ElevenLabs' agents.

The meter is the reason data vendors bundle. An agent that can't price its own access has to hide inside a contract that's already approved.

Before the next enrichment renewal, ask ZoomInfo what share of the renewal price the agents account for, and get a capped agent-access allowance from the CRM owner in the order form before the multiplier lands.

By Business of GTM