Agentic Is Rewiring Advertising But In B2B It's Theatre. Substrate Is the Fix.
Agentic layers are delivering real value already; 87% reductions in campaign setup time, 70% faster troubleshooting, and 40% more impressions at 30% lower effective CPMs against comparable benchmarks
AdExchanger’s inaugural Programmatic AI event opens in Las Vegas today. Three days, headlined “Mastering the Shift to Intelligent Media,” pitched as the industry’s first dedicated forum for the practical realities of agent-led advertising.
Every DSP, SSP, and analytics vendor that’s shipped an “agent” or a “kit” or an “OS” in the last six months will be on stage, and a credible portion of the buy side will be in the room.
For B2B this is theatre. Substrate is the fix, and in this context, substrate means the underlying layer of data and context that an agent reasons over to make decisions. It's the foundation underneath the agent, not the agent itself. Fixing that is the only way to make agentic work for B2B, as we’ll explore below
Above, join the conversation on LinkedIn
I missed the memo in time to get out there. Genuinely sore about it. The agentic shift in programmatic is the most interesting structural change since real-time bidding, and watching it unfold from just outside of London rather than out on the strip, is suboptimal.
But from the remove, one thing is increasingly clear. The agents being demoed in Vegas this week were built for B2C. The B2B translation hasn’t happened yet, and the gap isn’t the kind that closes by waiting another year. It’s a substrate problem. The agents are clever. The foundation they’re meant to reason over doesn’t exist for B2B in any usable form, and the vendors selling B2B “agentic” capabilities right now are mostly optimising the wrong thing more efficiently.
This piece is about why that is, what an honest B2B agentic stack would actually need to model, and where the next twelve months are likely to take us. Take it as the long-form version of an argument I’ve been making in shorter form for a while, framed against this week’s news cycle.
What you’ll learn in this article
Why the agentic narrative coming out of AdExchanger’s Programmatic AI event in Vegas doesn’t translate cleanly to B2B - YET
The structural difference between B2C and B2B feedback loops that breaks the agent’s ability to learn
Why the substrate (the context an agent reasons over) matters more than the agent itself
What an effective B2B agentic system would actually need to model end to end
Where agentic B2B is most likely heading over the next twelve months
What B2B marketers should be doing now to be ready when the substrate catches up
The Vegas moment is real, but it’s a B2C moment
The Programmatic AI event is the most concrete signal yet that agentic adtech has moved from theory to shipping. The session list reads like a roll call of everything that’s launched since CES: PubMatic on AgenticOS and the Agentic AI Acceleration Program, The Trade Desk on Koa Agents and the Open Agentic Kit, IAB Tech Lab on the Agentic Real-Time Framework, plus the Ad Context Protocol that’s emerging as an industry interop layer alongside Criteo’s Model Context Protocol implementation.
The numbers behind the noise are starting to look credible. PubMatic reported on its Q1 earnings call earlier this month that AgenticOS had run more than 1,000 direct deals and more than 30 fully autonomous end-to-end campaigns by the end of Q1, contributing to 80% year-over-year growth in PubMatic’s emerging revenue category.
Early case studies have claimed 87% reductions in campaign setup time, 70% faster troubleshooting, and 40% more impressions at 30% lower effective CPMs against comparable benchmarks - all stats which combine into a very interesting picture.
The Trade Desk’s Koa Agents launched with Stagwell as the named agency partner in late April. WPP Media, MiQ, Butler/Till, Brkthru, and Untapped Growth’s collective of independent agencies are all running live campaigns. Skyler McGill at Wpromote called it “the biggest transformation in programmatic since real-time bidding.” That sounds like vendor hype, except it’s the kind of language you only really hear when the people in the trenches start saying it.
So the substrate is genuinely changing on the supply and execution side. What you’ll see celebrated in Vegas this week is real, and the practitioners are right to take it seriously. Anyone in B2B should be keeping a watching eye, and I’m sure the masses of newly emerging vendor in-house teams are chomping at the bit to get into this
The catch is that almost all of the early validation, and almost all of the case study performance numbers, come from B2C campaigns. CPG, beverages, retail, performance video, mid-market mass-reach buys. The agent’s job is to optimise within a tight feedback loop that closes inside the platform within minutes. That works because the unit of decision and the unit of buy more or less line up.
The moment you swap in a B2B campaign, that alignment falls apart.
The B2C feedback loop is why agents work
Here’s the loop the Vegas vendors are extending into the open programmatic ecosystem this week.
A consumer triggers an intent signal. The signal might be a search term, a product page visit, a cart abandonment, or a CRM trigger from a logged-in session. The signal is read by the bid stack in real time. The agent reasons over the available inventory, the audience cohort, the creative variants, and the predicted return. It bids. The impression serves. The same consumer, often in the same session, clicks. They convert at the checkout. The conversion pixel fires back to the platform. The agent reads the outcome. It updates its model. It tweaks.
That entire loop, from signal to closed conversion, often runs in single-digit minutes for the high-velocity verticals where agentic optimisation is being trialled hardest. Even for slower considered purchases like a holiday booking or a sofa, the loop closes within days, not months. The unit of decision (a single person, often a single session, ending in an online conversion) and the unit of buy (the impression) sit inside the same identity, on the same domain, in the same data environment. The agent learns from the outcome because the outcome is observable, attributable, and quick.
Performance Max has been running on something close to this logic since 2021. Meta Advantage+ has been doing the same since 2022. Both work, within their constraints, because the loop closes fast and the optimisation signal is honest. The agents being shipped on AgenticOS, Koa, and the rest are extending the same logic into the open web and across SSPs. That’s a meaningful unlock for B2C, because it removes the walled-garden tax on agent-led optimisation. But it doesn’t change the underlying assumption that the feedback loop is tight, observable, and closes inside the platform.
This is the critical assumption to flag, because B2B violates almost all of it.
The B2B loop doesn’t close, and the agent has nothing to learn from
In B2B, the unit of decision is not a person. It’s a buying committee, procuring on behalf of a business, and the decisions they’re making are expensive and slow. Gartner’s research on technology procurement has consistently put the average B2B buying committee at six to ten stakeholders, and the buying cycle for mid-market and enterprise software at six to eighteen months. The 6sense Buyer Experience Reports have shown that more than 90% of buying committees have shortlisted vendors before any direct sales engagement. The buyer journey, in other words, runs in the dark for most of its life and converges into a formal procurement process when it’s nearly over.
The unit of buy is still the impression. The agent can still see who it bid for, what it served, what creative it ran, what site it ran on. That hasn’t changed. What has changed is the relationship between the impression and the outcome that matters. The outcome that matters is the deal, and the deal closes via a DocuSign in the CRM, attached to an account record, weeks or months after the impression that may have started it.
The agent doesn’t see that DocuSign event. Even if it did, the time delay between the impression and the close means the agent can’t reasonably attribute marginal lift to any single bid. The signal layer (clicks, page views, content downloads, video completions) has a thin and noisy correlation with whether a deal eventually closes. The agent ends up optimising on what it can measure, which is the surface engagement, while the actual outcome lives in a system it isn’t connected to.
So when an agent “optimises” a B2B campaign today, it is optimising the wrong thing more efficiently. Cheaper clicks. Better engagement rates. Media performance metrics that have weak predictive power for pipeline. The agent runs faster, the dashboards look better, and the pipeline doesn’t materially move.
This is the heart of the substrate problem. The agentic infrastructure is good. The agents themselves are clever. They were just trained against a feedback loop that closes inside the platform in minutes, and the B2B world they’re being asked to operate in has a feedback loop that closes outside the platform, in months, in a system the agent has no read access to.
The substrate matters more than the agent
The point worth labouring, because it’s the conclusion that gets things in the right order, is that the substrate matters more than the agent.
By substrate, I mean the data context the agent reasons over. The signals it can read. The state it can update. The outcome it can observe. The unit of decision it’s optimising toward. For a B2C agent, the substrate is impressions, clicks, cookies, on-domain sessions, and conversion pixels. That substrate exists in clean form across the open programmatic ecosystem. Standards are emerging fast (the Ad Context Protocol, the IAB Tech Lab Agentic Real-Time Framework, the Open Agentic Kit) to make it interoperable across vendors.
For a B2B agent, the substrate would need to be account-level intent signals, engagement velocity across the buying committee, CRM stage progression, opportunity creation events, deal stage transitions, and ultimately the closed-won outcome. That substrate exists, but it lives in fragments. Intent data lives at Bombora, G2, and a handful of co-ops. Engagement data lives in marketing automation platforms. Account graph and committee mapping live in 6sense, Demandbase, ZoomInfo, and similar. CRM stage data lives in Salesforce and HubSpot. The closed-won signal lives in CRM and is gated behind RevOps reporting that often runs on a quarterly cadence.
None of that substrate is currently exposed to the bid stack in a form the agent can reason over in real time. The vendors selling B2B agentic capabilities today are almost without exception layering an agent on top of B2C-shaped substrate (clicks, impressions, engagement scores) and calling it B2B. It produces the same kind of optimisation it would produce for a B2C campaign, with the same shortcomings, dressed up in B2B language.
This isn’t a criticism of agents. It’s a criticism of the order things are being built in. Building the agent first and assuming the substrate will catch up is the same mistake the industry made when it built ABM platforms before account graphs were stable, or when it built intent-data activation before intent-data quality was good enough to act on without a human in the loop. The platforms shipped. The outcomes lagged. The customers wrote the disappointment up as the technology being immature, when in fact the substrate underneath was the immature thing.
Agents are only as useful as the context they reason over, and the context for B2B isn’t users. It’s accounts, buying committees, deal stages, and the velocity at which they move through the funnel.
If you fix the substrate, the agent gets useful very quickly. If you don’t, you can put the smartest agent in the world on top of the wrong context and the outcome will not move.
What a B2B agentic system would actually need to reason over
So what would the right substrate actually look like? This is where the conversation gets interesting, because it stops being a critique and starts being a roadmap.
A real B2B agentic system would need to reason over at least five things, in real time, with read and write access to each.
First, account-level intent. Not user-level. The agent needs to know which accounts are showing surge signals across third-party content topics, first-party engagement, and review-site traffic. Raw signal access matters here, because aggregated and pre-scored intent (”MEDIUM intent on Cybersecurity”) strips most of the information the agent could use. The agent should be reading the underlying signals and forming its own judgments about salience, not consuming someone else’s threshold model.
Second, engagement velocity across the buying committee. Not just whether the account is engaged, but which roles inside the account are engaging, in what sequence, with what content. A CTO reading a security whitepaper followed two weeks later by a procurement lead requesting a vendor questionnaire is a different signal than three random employees clicking a retargeting ad. The agent should know the difference, and it should have a model of what a real buying committee looks like by ICP segment.
Third, CRM stage progression. The agent needs read access to the deal pipeline. It needs to know when an account moves from MQL to SQL, from SQL to opportunity, from opportunity to closed-won or closed-lost. Without that, the agent has no truth signal to learn from. With it, the agent can start asking the right question: did this impression contribute to this account moving forward in the funnel?
Fourth, the high-value action layer. The B2B equivalent of a conversion isn’t a checkout. It’s an HVA: a demo request, a pricing-page visit, a sales engagement, a contract review. These need to be tracked, scored, and exposed to the agent as the proximate outcome it can optimise against in the absence of a closed-won signal.
Fifth, win probability by segment. The agent needs a model of which kinds of accounts, at which stages, with which engagement patterns, are most likely to close. That model has to be built off historic CRM data, not borrowed from someone else’s benchmarks. It’s the model that lets the agent answer the bid-time question that actually matters: what’s the marginal lift to account progression from this impression?
Different question. Different math. Different infrastructure. Probably different vendors winning at the end of it.
The next twelve months: from optimising campaigns to orchestrating accounts
Where this gets genuinely interesting is the move from agents that optimise inside a campaign to agents that orchestrate across accounts. I think we’ll see that shift gain traction over the next twelve months, and the vendors who get there first will have an outsized claim on the next cycle of B2B media buying.
The first move is budget reallocation agents. Instead of optimising a fixed campaign budget against engagement signals, the agent reallocates spend across accounts based on progression probability. Account A moves from cold to warm based on a surge signal: the agent shifts budget toward it. Account B sits at the bottom of the funnel and goes quiet for three weeks: the agent pulls spend back. The optimisation target stops being CPM or CTR and starts being “marginal change in pipeline probability per pound of spend.” That’s a fundamentally different objective function, and a fundamentally different agent.
The second move is buying committee orchestration agents. Once the agent has a model of which roles inside an account need to be engaged in what sequence, it can start making channel and message decisions per role. A senior leadership target gets a CTV reach play with branded content. A practitioner-level target gets a contextual display campaign on G2 or a topic-relevant content site. A procurement target gets a competitor comparison or a TCO calculator. The orchestration becomes a decision the agent makes in real time, based on account state, rather than a campaign architecture a human builds in advance.
The third move is closing-the-loop agents. Once CRM stage changes are exposed to the bid stack as event triggers, the agent can adjust media activation in real time without a human in the middle. An account moves to opportunity stage: media spend ramps. An opportunity slips: a recapture sequence fires. A competitor wins: the account moves into a different long-cycle nurture flow. The orchestration layer collapses into a single decision loop, run on account context rather than user clicks.
None of this is exotic. The protocols emerging this week in Vegas (the Open Agentic Kit, the Ad Context Protocol, the IAB Tech Lab framework) could plausibly carry this kind of agentic logic if the substrate underneath them gets built out properly. The technology is there. The question is whether the vendors solving for B2B will build the substrate first or layer the agent on top of insufficient context and call it done. That choice will decide whether the next cycle of B2B media buying actually changes, or whether we get a generation of B2B agents that produce the same dashboards faster.
What to do about it now
If you’re running B2B marketing or media right now, you don’t need to wait for the substrate to mature before you start positioning for it. There are three things worth doing.
Audit your account-level data plumbing. If you’re not currently feeding CRM stage changes back into your media activation stack, that’s the gap. The closed-loop reporting is the first piece of substrate any agentic system will need to operate on, and most B2B marketing teams still treat it as a quarterly reporting exercise rather than a real-time signal.
Get serious about engagement velocity, not engagement volume. The metric that should be on your dashboard is the rate at which an account is escalating through the buying committee, not the total number of impressions or clicks. If your reporting still treats every engagement as equal, you’ll get agentic optimisation against the wrong objective the moment you deploy it.
Stop letting vendors define “intent” as a pre-scored topic surge and start asking for raw signal access. Pre-scored intent is the equivalent of letting someone else’s model decide what your agent gets to reason over. It’s a substrate compromise dressed up as a feature. The competitive advantage in the next cycle will accrue to the buyers who can see the raw signals and build their own judgement layer on top.
The agents are coming. The substrate is the lever you can pull now.
At FunnelFuel, we’ve been building the account graph underneath our activation precisely because we could see this coming. Our position has been that the agent layer is a follower, not a leader: the team that wins the next cycle is the team whose agents are pointed at the right substrate. We’ve spent the last eighteen months on the data infrastructure, the raw signal access, and the closed-loop reporting that makes account-level optimisation possible. The agents are downstream of that. They’ll get clever fast once the substrate is in place.
If you’re building toward this and want to compare notes, hit reply. Always interested in talking to the people working on the substrate problem from the inside.
The B2B Stack is written by Mike Harty, founder of FunnelFuel and a fifteen-year veteran of B2B programmatic advertising. If this resonated, the easiest thing you can do is forward it to one person in your network who’s currently being pitched an agentic adtech product. Want to connect with me? You can email me here or connect on Linkedin here
Sources and further reading
AdExchanger, “AdExchanger Launches Programmatic AI, Expanding Its Leadership In Intelligent Media” (January 2026)
PubMatic, “PubMatic Launches AgenticOS, the Operating System for Agentic Advertising” (January 5, 2026)
AdExchanger, “PubMatic’s Agentic AI Is Going Beyond Direct Deals” (May 8, 2026)
Marketing Dive, “PubMatic offers agentic platform access to indie agencies in new deal” (March 30, 2026)
The Trade Desk and AdAge, Koa Agents and Open Agentic Kit announcement (April 22, 2026)
6sense, “2025 B2B Buyer Experience Report” (94% of buying committees rank vendors before first contact)
IAB Tech Lab, Agentic Real-Time Framework and Programmatic Governance Council formation (April 2026)