Originally published by The Drum.
As AI agents take on more go-to-market decisions, the quality of the signals feeding them becomes critical. Weak assumptions once tempered by human judgement risk being automated at scale, making signal architecture a new source of competitive advantage.
AI agents are moving into go-to-market teams faster than the conversation around them can keep up. Gartner’s own forecast shows the scale of what’s coming: AI agents are projected to intermediate $15tn in B2B purchases by 2028. That figure describes entire procurement cycles, from supplier identification through option evaluation to order execution, completing without a human buyer ever navigating a vendor’s website. The current debate around that shift sits almost entirely on models, automation and orchestration: which agent framework to adopt, which workflows to hand over first, how much autonomy to grant before a human checks the output. That debate matters, but it skips a harder question. What evidence will these agents actually use to decide what a buyer wants next?
B2B marketing has spent a decade treating buyer signals as broadly interchangeable. Website engagement, review site activity, bidstream intent, publisher network data and first-party content consumption all get folded into a single view of demand, usually inside one dashboard, often under one score. These sources represent fundamentally different types of buyer behavior. A prospect reading a comparison guide on a review site is not doing the same thing as a prospect downloading a whitepaper directly from a vendor. Bidstream activity inferred from ad exchanges is not the same evidentiary category as a logged-in visit to a pricing page. Marketers have been compensating for weak signal architecture for years. Experience lets a lead discount a spike in anonymous intent data while trusting a return visit from a known account. Judgment fills the gap that the data architecture never closed.
AI agents don’t have that judgment and they won’t develop it on their own. An agent optimizes against whatever inputs it’s given, with the confidence of a system that has no concept of which of those inputs deserve less trust. Feed it a signal environment in which inferred behavior sits next to observed behavior, with no distinction between the two, and it will treat both as equally actionable. That’s the real shift worth naming. A system that can no longer tell good evidence from bad, executing at full speed and with total confidence. AI won’t expose a data problem. It will expose a signal architecture problem that has existed for years and been quietly managed by people, not systems.
The automation of flawed assumptions
Hand a go-to-market agent a demand signal built from a blend of inferred intent, weak proxy behavior and genuinely observed engagement, weighted as though all three carry equal evidentiary value, and this is what happens next. A human team running that same blended signal today still makes uneven decisions, but a rep or marketer can sense-check an account that looks unusually active and quietly deprioritize it. An agent making outbound sequencing or budget allocation decisions based on the same signal has no equivalent instinct. It acts on the full data set as given, at whatever speed and scale it’s been deployed to operate. The speed and scale change what a bad assumption costs. It’s the same flawed assumption, replicated across every account the agent touches, without the manual checkpoints that once caught the worst of it.
This is the part of the AI conversation that gets skipped because it isn’t about the technology. Every gap in a signal architecture, every conflation of observed and inferred behavior, every reliance on third-party proxy data that no one on the team fully trusts becomes a gap that an agent will now execute against without hesitation. Organizations that haven’t confronted the quality of their signal inputs are about to find out what happens when those inputs get automated rather than interpreted.
Signal quality as a constraint on decision quality
Models are converging fast enough that model choice is becoming a commodity decision rather than a differentiator. The competitive edge is shifting to a more deliberate hierarchy of buyer intelligence, one based on how the behavior was actually observed rather than which vendor happens to own the data.
That hierarchy starts with distinguishing signals by evidentiary strength. A contact-level engagement with owned content, tied to an identifiable account, sits at a different level than an inferred topic surge pulled from bidstream data with no identifiable buyer attached. Buying group visibility, built from patterns across multiple stakeholders at an account rather than a single anonymous session, carries more decision weight than a one-off visit. None of this is a new idea in principle. What’s new is the consequence of getting it wrong. When a marketer misjudges signal quality, the cost is one bad campaign decision. When an agent misjudges signal quality, the cost compounds across every decision it’s authorized to make, at the speed it’s authorized to make them.
As AI takes on more of the commercial decision-making that used to sit with people, signal quality stops being a data hygiene concern and becomes the defining constraint on decision quality itself. Marketers already lived through what happens when the evidentiary basis for a decision sits inside a system they don’t control. AI agents making GTM decisions on unexamined blended signals are set to repeat that experience, this time on the demand side, unless the signal hierarchy behind the agent is something the organization actually built and can account for.
That’s why the next competitive advantage won’t belong to the organizations with the largest datasets. It will belong to those with the most deliberate signal architecture. Every weakness in today’s signal architecture becomes an automated weakness tomorrow.