The most consequential forecast in B2B for the rest of the decade comes from Gartner, and it has barely registered in the RevOps function it will reshape. By 2028, Gartner expects 90% of B2B buying to be intermediated by AI agents — pushing more than $15 trillion of B2B spend through agent exchanges. By the same horizon, Gartner forecasts that AI agents will outnumber human sellers by ten to one. This is the operating environment your RevOps stack will need to survive. Most stacks will not.
The current RevOps architecture — CRM as the system of record, opportunity stages as the unit of progress, forecast roll-up by territory, attribution stitched together after the fact — was built for a world in which humans took every meaningful action and the system observed them. That world is ending. In the agentic era, agents take meaningful actions: they research accounts, draft messages, qualify replies, schedule meetings, negotiate quotes, and route deals through tooling. The system has to do something different. It has to govern them.
This briefing lays out the four-layer architecture for a RevOps function that survives the transition. It is opinionated. It assumes Gartner's forecast is directionally correct and that operators have somewhere between 12 and 24 months to begin re-architecting before the gap between AI-native operators and everyone else becomes unrecoverable.
The forecast that breaks the architecture
Gartner's top strategic prediction for 2026, delivered at IT Symposium/Xpo in late 2025, is unambiguous: by 2028, 90% of B2B buying will be AI-agent intermediated, pushing over $15 trillion of B2B spend through agent exchanges. That is roughly half of US GDP routed through software acting on behalf of buyers. It is not a software market projection — it is a structural claim about who, or what, will be on the other side of the table from your sales team.
The accompanying predictions sharpen the operational picture. Gartner expects AI agents to outnumber human sellers by ten to one by 2028, while fewer than 40% of sellers will report that AI agents improved their productivity. The implication is not that agents fail to work — it is that the operators who deploy them without an architecture will get less out of them than the operators who deploy them with one. The architecture is the differentiator.
One more figure: by end of 2026, Gartner expects roughly 40% of enterprise applications to have embedded agents, up from less than 5% today. The agents are already arriving inside the tools your team uses. The question is whether your RevOps architecture is governing them or merely hosting them.
Why the current RevOps stack breaks
Three failures explain why the existing architecture cannot govern an agentic GTM.
The data model is opportunity-centric. Salesforce, HubSpot and the wider CRM cohort were built around the opportunity record as the atomic unit. Accounts and contacts are organising metadata; opportunities are where the work happens. In an agentic world, the unit of progress is not the opportunity — it is the signal. Agents act continuously on accounts, surfacing intent, drafting outreach, qualifying responses. The opportunity is a downstream consequence of that activity, not its container. A data model that treats opportunities as primary will continuously miscount what is happening.
The decision layer is implicit. In the human-paced funnel, sales judgement was buried inside the heads of SDRs, AEs and managers — what to research, what to send, who to escalate to, when to discount. That tacit knowledge worked at human speed because there were not many decisions and each one had visible context. Agents cannot inherit tacit knowledge. They need explicit policy. Most RevOps teams have never written down their decision rules; the rules exist as norms, manager coaching, and Slack threads. That is unworkable when an agent is making the call ten thousand times a day.
Observability is retrospective. Attribution platforms stitch the story together after the deal closes. Forecasting accuracy of 75% has become normal because forecasters are reading a partial trace of what actually happened. When agents are taking most of the meaningful actions, retrospective stitching breaks completely — there is no clean record of which agent did what, why, and with what outcome. Observability has to be designed in, not bolted on.
The four-layer agentic RevOps architecture
Four layers, in a clear dependency order. Each is a real architectural commitment, not a tooling category. Get the layers right and the vendors fit around them; get the layers wrong and no vendor stack will rescue the function.
Accounts and signals as primary, not opportunities.
Explicit policy, written down, version-controlled.
Where agents stop and humans take over — defined, not assumed.
Every agent action logged, attributed, forecast-able.
The order matters. Data layer first, because without it the decision layer has nothing to decide on. Decision layer second, because without it the handoff layer has no thresholds to enforce. Handoff layer third, because without it the observability layer is recording chaos. Observability layer last, because it depends on everything below being explicit enough to record.
Vendor landscape — who fits where
The market is fragmenting into clear functional layers. Naming names, because the function deserves a real read on the landscape rather than vendor-agnostic safety.
Data layer. Salesforce and HubSpot still hold the centre of gravity, but neither was architected for signals-as-primary. Salesforce Data Cloud is the company's attempt to retrofit a signal layer onto the opportunity-centric core; HubSpot's Breeze is the parallel move. Clay and Apollo represent the data-layer-native cohort — they treat accounts and contacts as the unit and the opportunity as the consequence. For a greenfield agentic RevOps build in 2026, the choice is no longer "which CRM" but "which signal platform sits in front of the CRM".
Decision layer. The least mature of the four. Salesforce Agentforce, HubSpot's agent suite, and Microsoft's Copilot Studio are positioning to own this layer, but most production implementations today are still custom — a combination of internal policy documents, scoring models, and orchestration code. Expect this layer to consolidate fast through 2026–2027. The strategic risk is locking in a decision layer that lives inside a vendor's runtime rather than as code your team can audit and version.
Handoff layer. Outreach, Salesloft, Apollo and Clay all have agent routing today, with varying degrees of sophistication. Gong and Clari sit closer to the handoff layer from the post-call side — both are racing to become the system that decides which conversations a human needs to be on. The vendors who win this layer will be the ones whose handoff logic is most explicit and most auditable, not the ones with the most agents.
Observability layer. Clari and Gong dominate today's incumbent stack, but their observability was built for human action. The next-generation observability stack — what is sometimes called "agent ops" — is being built by a cohort of newer vendors and by Anthropic and OpenAI directly through model-side telemetry. Expect this to be the most contested layer in 2027.
The vendor map will change. The four-layer model will not. Architect to the model, not to a vendor's product roadmap.
The forecasting problem, rebuilt
Forecast accuracy below 75% is now the operating norm for B2B RevOps. In the agentic era, that number will collapse further before it recovers — for a simple reason: agents are taking meaningful actions, and forecasters are reading partial traces. The fix is structural, not procedural.
The new forecasting discipline rests on three commitments. First, the forecast reads from the observability layer, not the opportunity stage. Stage progression is a derived signal; agent actions are the primary signal. Second, the forecast distinguishes between agent-led and human-led commit. Different confidence intervals, different roll-up logic, different exception handling. Third, the forecast carries a model identifier. Which agent, which policy version, which decision threshold produced the prediction. Without that, post-mortem learning is impossible.
RevOps leaders who get this right will restore board-level trust in forecasting within two cycles. Those who do not will see forecast accuracy continue to deteriorate as agents take a larger share of pipeline action.
The RevOps team that emerges
The architecture implies an org chart. Four new (or reshaped) roles sit at the centre.
The RevOps Architect. Owns the four-layer model. Reports to the CRO with parity to VP Sales. Not a tooling specialist — an architect. The senior individual who makes the structural decisions that the rest of the function executes against.
The Decision Engineer. Writes, tests and versions the explicit policy that governs agent behaviour. Part product manager, part RevOps analyst, part prompt engineer. The role is too important to leave to vendor defaults.
The Agent Operator. Runs the day-to-day deployment, tuning and observability of production agents. Most companies will need several. The role replaces, in part, the SDR manager role that managed humans doing similar work.
The Forecast Steward. Owns the integrity of the forecast itself, including the model identifiers and the agent-versus-human distinction. The role exists because the CFO needs a single accountable owner when the forecast misses.
Smaller companies will collapse these into fewer people; larger companies will sub-specialise further. The underlying structural shift is the same: RevOps stops being a tooling function and becomes an architecture function.
Five lessons for RevOps leaders
Drawn from the four-layer model and the operators already building against it.
If your team is still manually creating opportunities to track work, you are operating an opportunity-centric architecture in a signal-centric era. Move the primary unit of progress to the account and the signal. Let opportunities be derived, not declared. Every CRM platform now supports this; very few RevOps teams have actually flipped the model.
If the rules that govern how your team qualifies, prioritises, discounts, and escalates live in manager heads and Slack threads, agents cannot enforce them. The first deliverable of any agentic RevOps programme is a written, version-controlled policy document — qualifying criteria, escalation triggers, discount authority, compliance constraints. Without this, no agent deployment will be safe to scale.
Complex negotiation. Brand-defining customer relationships. Strategic-account exception handling. Regulatory edge cases. That is the list. Everything else is agent territory. RevOps leaders who hedge on this will end up with humans doing agent work and agents stuck behind humans waiting for approvals. Pick the line and enforce it.
If you cannot reconstruct what an agent did, on what input, against what policy, you cannot forecast it, cannot improve it, and cannot defend it in front of the board. Agent observability is not a future feature — it is the entry condition. Do not deploy production agents without it.
The most common 2026 mistake is to bolt agents onto the existing RevOps stack and call the job done. Agents on top of an opportunity-centric data model, an implicit decision layer, an undefined handoff, and retrospective observability will produce more noise, not more revenue. The work is structural. The function itself needs to change.
The operator's takeaway
Gartner's forecast is doing the heavy lifting in this briefing, but the lesson would hold even if the headline figures were half what they project. The structural point is that meaningful actions are migrating from humans to agents, and a function built to observe humans cannot govern agents. RevOps either becomes the architecture function that governs the new system, or it becomes a reporting function on a system someone else is governing.
The four-layer architecture — data, decision, handoff, observability — is the most defensible structure we have seen for the function that survives the transition. It is not the only structure. It is the one consistent with how the leading agentic RevOps deployments are actually being built in 2026. Start with the data layer and work up. Get the order right and the vendors fit around the architecture. Get the order wrong and you will be re-architecting in 2028 with less time and less optionality.
The forecast does not give you decades to think about this.