ZoomInfo and Gong both shipped agents this week that run a sequence of steps on their own, and Bain has put a dollar figure on the labor that kind of agent replaces. Taken together, the releases say the unit of sales AI is changing from a single task to a complete play, and that changes what a sales leader should measure, buy and control.
A play, in ZoomInfo’s own example
ZoomInfo’s launch post for Agent Teams, written by CEO and founder Henry Schuck and updated October 1, starts from a description of the current state. Most revenue teams, it says, have an assistant in the CRM, enrichment in one tool and sequencing in another. Each handles a task inside its own application, and none coordinates with the others or takes responsibility for the outcome.
The post then walks through one play. A person who helped buy the product at one company changes jobs and becomes a director of RevOps at a Tier 1 account. A traditional workflow, the post says, matches the title, creates a CRM task and hands off at the point where judgment begins. An Agent Team runs on a recurring schedule against an audience of past champions, spots the move, reconstructs the person’s role in the original deal, identifies the current account owner and checks for an open opportunity, because an open opportunity or a different owner can change the route. It drafts outreach grounded in the earlier relationship and sends the account owner a brief in Slack. Two weeks later a pricing page visit triggers the same team to continue, and every step is written back to Salesforce or HubSpot.
Advertisement
ZoomInfo lists five more plays: a renewal at risk, a closed-won handoff, a closed-lost win-back, an inbound buyer and a closed-won deal that creates new pipeline. It ships six templates, including Inbound Lead Activation, Outbound Prospecting and Weekly Sales Briefing. Agent Teams sits inside GTM Studio and, according to the post, needs no new SKU or contract. The post says the product is built on technology from DoubleO.ai, founded by Derek Osgood and Karthik Suresh and now part of ZoomInfo.
We covered the data side of this company’s strategy last week in Sales Intelligence Becomes a Utility for AI Agents. Agent Teams is the orchestration layer that sits on top of that data.
The same shape in Gong’s release
Gong announced Mission Callisto on September 30, and one piece of it follows the same logic. Agent Builder lets teams create Custom Agents that run automatically in response to events and conditions they define. Gong says this turns work that once needed someone to remember and act into an automated workflow. Teams set the trigger, conditions and actions in a visual editor or describe the agent in their own words. Gong adds that automatically triggered Custom Agents are rolling out throughout this year, while manual triggering already exists.
Two vendors with different product histories, one in contact data and one in conversation analysis, ended up in the same place within days of each other. A trigger, a set of conditions and a chain of actions, assembled by an operator and run without a person at each step.
Bain sizes the labor behind it
Bain & Company’s Technology Report 2026 gives the commercial argument. In a section on cross-system labor, authors David Crawford, Chris McLaughlin and Greg Fiore say the large opportunity for agentic AI is converting labor costs into software spending by automating coordination work between systems: employees pulling data from one tool, checking another, interpreting free text and deciding when to escalate. Rules-based software and robotic process automation cannot do that work, Bain says, because it breaks on ambiguity and context.
Bain estimates the potential US market at $100 billion, of which vendors capture about $4 billion to $6 billion today, leaving more than 90% untapped. Sales is the largest single function in the estimate at roughly $20 billion. Bain says that is due more to the number of sales employees than to an exceptionally high potential for automation. It puts sales and IT at 30% to 40% of workflow tasks automatable, below the 40% to 60% it gives customer support and R&D, and names relationship nuance and deal-by-deal variation as the constraints.
In our read, that 30% to 40% range matters for the plays above. A champion-tracking play has a verifiable output, a brief that lands with the right owner, and a low cost of failure. A forecast call or a negotiation has neither. Bain’s six factors for judging a workflow are output verifiability, consequence of failure, digitized knowledge availability, integration and orchestration complexity, process variability and physical-world dependency. Applied to sales, and this is our judgment, the plays vendors lead with, such as routing, enrichment, handoffs and win-back outreach, score well on most of them. The plays they mention less often score worse.
Pricing leaves the seat
Bain also draws a pricing conclusion. When agents deliver end-to-end outcomes, it writes, “the natural pricing unit shifts from seats and logins to outcomes and usage.” It cites a qualified lead as one example of such an outcome.
ZoomInfo’s post is an early case of the usage side. Agent Teams uses consumption-based AI Credits billed against the customer’s existing ZoomInfo contract. An estimate of the credits a play will use appears before the play is activated, and usage is tracked for each Agent Team. Bain’s report frames the macro version of the same shift. David Crawford, Chairman of Bain’s Global Technology, Media, and Telecommunications practice, writes in the report’s introduction that “Opex is migrating from headcount to token spend.”
For a sales leader the practical effect is a new line in the budget that behaves like variable cost. A seat license costs the same whether the rep logs in or not. A play costs credits every time it runs, so the number of triggers, the breadth of an audience and the length of the chain of agents all move the bill. Our October 1 opinion piece on carrying the 2026 AI stack into FY27 is a companion read. In our view, consumption pricing turns a line-by-line audit of this spend into a monthly task.
Control moves inside the play
If a play runs without a person at each step, the guardrails have to live inside it. ZoomInfo’s post says as much: a single misfire in a complex GTM workflow can set off a chain of incorrect actions and bad data. Its answer is play-level monitoring, admin dashboards, role-based access, credit estimates before activation and organization-level knowledge bases. Its FAQ also says a Pause for Review step for human-in-the-loop approval is on the roadmap for a future release.
Salesforce’s account of customer rollouts, drawn from Dreamforce 2026 speakers, shows what the control layer looks like in production. Most examples there come from service and retail, but the pattern carries over. Sammons Financial Group ran more than 200 guardrails and tests before going live and chose to stop the model learning from live conversations. Andrew Walling, AVP of Capability Planning and Delivery at Sammons, described the fail-safe: “We have a kill switch, a supervisor agent that’s constantly listening to all of those phone calls, checking the values in the CRM against what was said on the line.” AT&T, in the same account, kept the ability to turn agents on or off at any time and runs its own LLM for anything touching proprietary IP.
Two controls stand out for a sales motion. One is a supervising check that compares what an agent wrote to the CRM with what actually happened. The other is a switch that one named person can use without a ticket. Neither appears in most sales AI demos, and we raised the missing-numbers problem in Dreamforce Sold Trust. The Numbers Are Still Missing.
Data is the binding constraint
Bain’s list of six factors includes one that it singles out. Digitized knowledge availability, meaning whether an agent has well-labeled, machine-readable inputs at runtime, including tacit knowledge that lives in people’s heads, is in Bain’s words “almost always the binding constraint on automation.” Even in data-rich settings, Bain says, automation stalls when the context for a decision was never documented.
Gong’s release reads as a response to that constraint. Gong Enrich extends the Gong Revenue Graph with third-party account and contact data. Customers connect supported providers, set the order in which Gong searches them and fill remaining gaps with Gong Credits and data from Gong Enrich Data Partners. Gong lists Apollo, Findymail, Firmable, Kernel, LeadIQ, Lusha, RocketReach, Wiza, ZeroBounce and ZoomInfo. Gong’s Chief Product Officer and Co-Founder, Eilon Reshef, put the rationale this way: “Every revenue decision is only as good as the context behind it.”
The detail worth noting is that ZoomInfo appears on Gong’s partner list in the same week ZoomInfo launched its own agent product. In our read, data suppliers, orchestration layers and systems of record are now overlapping, and a buyer could end up paying for the same contact lookup through two contracts. Bain makes the competitive version of the point. The advantage in agentic software, it argues, is cross-workflow decision context, the ability to see and act across workflows that span several systems, and it expects that to matter more than owning a single system of record. It names Salesforce’s Agentforce among the incumbents moving quickly. For the vendor-side view, see our piece on revenue platforms that all claim unification.
What it means for the sales leader
Everything in this section is our read, not the vendors’ or Bain’s.
First, pick the unit you will measure. A task-level agent is judged by time saved. A play-level agent should be judged by the outcome the play exists for: a champion meeting booked, a renewal saved, a handoff completed with no missing commitments. Divide the credits or fees a play consumed by the outcomes it produced and you have a cost per outcome that you can compare with a rep’s cost for the same work. Bain’s pricing point says vendors are heading toward charging this way, so buyers need the baseline before the invoice arrives.
Second, start with plays Bain’s factors favor. Output that can be checked, a mistake that can be undone and data that is already in the CRM describe inbound routing, closed-won handoffs and meeting follow-ups. Renewal saves and win-backs touch revenue directly and deserve a shadow period in which the agent drafts and a person sends.
Third, write down who holds the off switch and what the supervising check compares. If a vendor’s control list is monitoring and role-based access, ask when human approval steps ship and what runs in the meantime. Ask which data partners and which credit pools a play draws on, because the same contact lookup could be billed twice across two tools.
Fourth, give RevOps ownership of the library of plays, not each rep. A play that works becomes a shared asset, and one that misfires needs one place where it is paused and fixed. Start with one play, a baseline from the last two quarters of manual work and a review date.
Source: ZoomInfo
