A growing share of market research is now conducted on people who do not exist. Instead of fielding a survey to humans, you query a model trained to respond the way they would, which is cheaper, faster and available at any hour. A meaningful portion of the enterprise buyer research shaping product and go-to-market decisions this year was never answered by a person at all.

Andrew Au is in an unusual position to assess that, because his company runs it both ways.

Au is co-founder and co-CEO ofIntercept, a B2B marketing firm that works with global technology companies, including Microsoft, where the relationship runs more than a decade, along with HP. His teams sit inside those companies' marketing organizations. Intercept also builds synthetic audiences of its own.

“There's real opportunity in both, and we use synthetic audiences ourselves in two ways,” Au said. “We build synthetic buyers that we test messaging against for specific audience types, and they're partly trained on the real conversations we collect. We also use a creative-testing model trained on eye-tracking and brainwave data from about 300,000 people, which predicts where attention will go with around 95% accuracy against real eye-tracking.”

His objection is narrower than a skeptic's, and harder to dismiss.

“What makes a simulation accurate is how recent its training data is,” Au said. “Buyers change their minds, and a synthetic buyer trained on last year's data answers with last year's opinions. The published validation studies say the same thing. Synthetic respondents match averages well but struggle when a market is shifting.”

Since a market that is not shifting rarely needs researching in the first place, the question becomes where the current signal is supposed to come from.

Verifying people who never agreed to participate

Intercept's answer is a platform called Watchtower, and the engineering problem it solves has nothing to do with collection, since public conversation is abundant. The difficulty is establishing who is talking and ensuring the sample amounts to more than the loudest accounts in the room.

“We look at an account's entire posting history, not just what it claims in the bio line,” Au said. “AI reviews how that person talks and behaves and the topics they discuss, which lets us validate the bio against their behavior. The same analysis screens out bots and vendor accounts.”

The sampling approach is where it departs from traditional social listening. Without a structured sampling method, crawling a social network tends to surface the most prominent or highly active accounts. Watchtower draws a random sample from a pool Intercept puts at roughly 20 million verified professionals, using a method from statistical physics published in peer-reviewed research. Each account is weighted equally regardless of audience size. Intercept reports findings in aggregate at a 95% confidence level with a margin of error up to 3%.

“Loud doesn't make it true, statistically significant or representative of a segment,” Au said. “A handful of people can be very loud on a topic and still be completely misleading. It's the difference between a popularity contest and what a segment really thinks.”

What the platform counts instead is frequency, meaning how many distinct accounts are discussing a topic and what each one's sentiment toward it is. Reading every account's full history is expensive, which sets the speed. A core topic covering ten to fifteen subtopics takes three to five hours, depending on how much conversation there is to process. Compared with a fielded survey, that is the difference between a research cycle and an afternoon.

Au points to a case where that distinction was expensive. A major enterprise software migration had stalled, and a global research firm concluded the blocker was budget and timing.

“Buyers were telling each other the last migration had been very painful, and they expected this one to be the same,” Au said. “It wasn't, because the new version was built on the same foundation as the one they already ran.”

The global narrative was rebuilt around telling buyers this would be the easiest migration they would ever make.

“The reason this didn't show up in the traditional survey is that no one wants to put ‘we messed up last time’ as a response option.”

That limitation is inherent to the approach, rather than merely an incidental effect. A questionnaire hands people a fixed menu and they tend to pick whatever they think the researcher is after, and a synthetic respondent trained on those answers inherits the same distortion one generation removed.

What the live signal currently shows

Intercept says it used Watchtower to analyze 285,988 conversations among U.S. IT and business decision-makers about enterprise AI vendors over a twelve-month period.

According to Intercept's research, 0.9% of what that group said about AI vendors described them as reliable in production. A further 11.3% said vendors go down more often than the business can accept, and the largest share, 13.8%, said vendors are only as trustworthy as the lab behind them.

Au reads the first two figures as one problem and the third as something else entirely.

“That's a reliability problem, and it goes beyond uptime,” he said. “Enterprises want consistency, compliance, and quality control, and a frontier model alone won't deliver that. You need a harness to control and govern agents, provenance to understand what happened, deterministic QA layers to monitor performance, and the right data estate underneath. AI vendors bring a capability, but it takes everything around it to make that capability reliable at scale.”

The trust figure points somewhere different.

“The 13.8% who say vendors are only as trustworthy as the lab behind them are acknowledging that many AI offerings are wrappers around a frontier model,” Au said. “That raises the question of protecting your alpha.”

He cites Satya Nadella's July essay, The Reverse Information Paradox, in which the Microsoft chief executive argued that companies pay for intelligence twice, once with money and again with the proprietary knowledge they have to reveal to make that intelligence useful.

“Even when a lab commits not to train on your data, leaders still wonder whether their alpha is leaking out,” Au said.

Why the agent gets the sandbox

A separate part of the study examined what leaders avoid putting into third-party AI tools, across what Intercept says was 152,083 conversations. Internal databases an agent could reach took the largest share at 29.7%, followed by customer personal data at 28.6% and source code at 21.9%.

“Companies are stuck on whether to build or buy,” Au said. “Do they build their own agentic platforms, or lease them from third-party vendors? There's real fear of vendor lock-in, and they want control over their agentic workflows, because that's where a ton of value is going to be created.”

So the pilots tend to run in the corners of the business where the blast radius is small.

“Connecting it to your ERP, your CRM or the core databases that power your business is different. People don't want to get that wrong, and they don't want to get fired, so there's hesitation.”

He is unwilling to let that be read as the deployments being theatre.

“Companies have real pilots and proofs of concept in pockets of the business. Most of those tools just aren't organization-wide yet, even when the announcement makes it sound that way, because companies haven't settled on the right governance, compliance, and security standards.”

Single player and multiplayer

The sharper version of that observation is the one Au uses internally, and it explains more about the state of enterprise AI than most adoption surveys manage.

“Most organizations are still using AI in what we call single-player mode,” he said. “A team spins up a point solution, it works for that team, and it stops there. The shift we're watching for is multiplayer mode, where one team's creation becomes the floor for the whole organization, so the next team starts from there instead of from scratch.”

“That's where most companies are stuck. They can build a pilot, but they struggle to operationalize it, which is why the ROI of AI is being aggressively debated.”

That recasts the ROI argument as an integration failure rather than a capability failure, which is also how Au reads the readiness question the industry has been circling all year.

“We've reached a point where the frontier models, and frankly the open-weight models, are capable enough. If they never advanced another step, finding the right use cases and deploying them strategically would still unlock trillions of dollars of economic value. So the question isn't whether the models can do it. It's whether we're ready.”

Readiness, as he defines it, spans vision, technology, data, process and people, and the long pole sits in the last three of those.

He reaches back to 1990, when Michael Hammer published “Reengineering Work: Don't Automate, Obliterate” in Harvard Business Review, arguing that most companies were using computers to accelerate processes that should have been demolished. The ones that captured real value rebuilt the workflow around what the technology made possible.

“AI asks the same question,” Au said. “Are you ready to rethink the business?”

What vendors should do about it

Au's advice to the companies selling into this market starts from the premise that the distrust is real rather than a messaging failure.

“Start by recognizing that the distrust is real, and it reaches well beyond the enterprise,” he said. Gallup found this year that seven in ten Americans oppose an AI data center being built in their local area, with 48% strongly opposed. “That skepticism can show up in buying committees. It's the same people.”

“So focus less on proving you have the best model and more on proving you're a vendor worth betting the farm on. The decisions vendors are asking customers to make have downstream implications worth tens or hundreds of millions of dollars, sometimes billions. Who do those customers want in their corner?”

"You earn that trust by showing it. How are you supporting change management? How are you providing transparency? How are you helping govern the models and show the provenance, so customers can see what's happening?"

"The conversation is much more human than it is technical."

Which is the part a simulation cannot reach. A synthetic buyer can tell you what a market believed when it was trained. It cannot tell you who that market has decided to trust.

VentureBeat newsroom and editorial staff were not involved in the creation of this content.