Arvid Madland Lyngnes
field note
2026-11-04
related project:Entriva.ai

What AI research learns about a company — and misses about its market

An AI assistant can produce a credible company analysis before it has examined a single source.

That is useful. It is also a reason to be careful.

I wanted to understand a practical question: when I ask AI to research a company, how much of the answer comes from the company's own description, how much comes from existing model knowledge, and how much is supported by independent information about the market?

There is no honest universal percentage. A language model cannot provide a reliable accounting of how much each sentence was influenced by its training data. Even when live web search is available, a list of cited URLs tells us where evidence was retrieved, not exactly how the model's internal knowledge affected the answer.

So we ran a small, deliberately transparent test on Betong Øst, a Norwegian ready-mixed concrete producer. It was independent public-source desk research, not an assignment from the company or an assessment of its internal commercial work.

We made three passes: first without new searches, then with ordinary company research, and finally with a more systematic search for changes in the market that might create future demand.

First pass: an apparently sensible analysis without research

Before opening any pages about Betong Øst, the AI generated seven hypotheses. It expected a regional ready-mixed concrete supplier serving contractors and developers, competing on proximity, price, delivery reliability and quality. It identified infrastructure projects, transport costs and environmental requirements as likely commercial drivers.

Several of those guesses were broadly right. But their apparent accuracy is not proof of deep knowledge about Betong Øst. They are also what a competent general analysis of this industry might say about many suppliers.

The geographic guess was incomplete: Betong Øst's own material describes operations across Eastern Norway and Trøndelag. The company presents 22 local concrete plants and a much wider operating footprint than the name alone might suggest.

The first pass illustrates a problem I keep encountering: a plausible analysis can look company-specific when it is largely industry knowledge applied to a company name.

That does not make it worthless. It makes it a starting hypothesis, not research evidence.

Second pass: a better company portrait, but where did the facts originate?

Ordinary web research established the registered business, location, ownership claims, products, financial figures and recent coverage. The company's website describes its factory network, DØNN products, low-carbon concrete and ability to provide project-specific environmental product declarations. A financial-data provider reported approximately NOK 951 million in 2025 revenue. Industry coverage reported stronger activity related to large infrastructure projects.

For a small, explicitly selected sample of nine company facts, we traced the immediate supporting sources:

These are counts of selected documented statements, not a measured source mix for AI in general. They do not measure the contribution of model training. They also do not establish that every third-party source independently verified the company's claims.

The source lineage matters more than the source count.

Betong Øst's description of its history, geographic reach and factory network appears in closely matching language on its own website, HamKam's partner page and Concrete Innovation Cluster's member page.

An AI assistant may find three websites supporting the same description. That is not the same as finding three independent examinations of the company's market position. It is likely one description travelling through several channels.

There was also a small reminder about time and definitions. The company website described 22 plants, while an AF Gruppen account of a 2025 industry presentation reported a Betong Øst technology executive referring to 24. Both statements can have reasonable explanations. We cannot resolve the difference merely by choosing the more recent-looking web page.

The second pass was materially more useful than the first. But it remained primarily a portrait of what the company is and what others have published about it.

It still did not tell us where its next customer would come from.

Third pass: start with the market, not the company's description

We then changed the question.

Instead of asking what we could find about Betong Øst, we asked what was changing around its operating locations that could create a reason to buy concrete, technical advice or related services in the future.

Two examples emerged from the project owners' own public information.

Hell–Værnes double track. Bane NOR says detailed planning was completed in autumn 2026, with construction procurement planned for autumn 2027 and a possible start in spring 2028. Betong Øst lists a plant at Stjørdal/Hell. This creates a sensible question about future concrete requirements, logistics and specifications before the contractor procurement is complete.

New Rv. 22 crossing at Fetsund. The Norwegian Public Roads Administration describes a planned road and bridge project with an estimated total cost of NOK 5.1 billion in 2026 prices. Its stated construction start is December 2029, subject to funding and priorities. Betong Øst lists a Fetsund plant. That is a long-term project to monitor, not an imminent sales opportunity.

We also found a possible change in how customers will evaluate suppliers. Norwegian public procurement statistics show a growing role for environmental criteria. Oslo's Water and Wastewater Agency's 2026 instructions explicitly refer to continuing climate requirements for concrete and steel and increasing the use of environmental declarations. Betong Øst already describes relevant product and documentation capabilities.

And we found something more specific than a generic market trend. At an AF Gruppen industry event, a Betong Øst technology executive described the administrative burden of many concrete recipes and environmental declarations prepared during tenders that later go unused when project specifications change.

That raises a useful commercial and operational question: could earlier involvement in material specification reduce rework for both the supplier and its customers?

It is a hypothesis grounded in a reported problem. It is not proof of a new service, a customer willing to pay, or a competitive advantage.

Four different things that AI research must not collapse into one

The third pass identified two named future projects, a change in purchasing criteria, and a documented source of operational friction.

It identified zero verified new sales opportunities.

That may sound disappointing. I think it is the most important result of the exercise.

We must distinguish between four things:

1. An observation: a project is being planned, a rule changes, or a problem is described. 2. A need hypothesis: that observation may create demand for a particular product or capability. 3. A buying opportunity: a relevant buyer has an unresolved need, authority, timing and a possible route to purchase. 4. A Right to Win: this supplier has a credible reason to be chosen over alternatives.

None of the public project pages told us whether Betong Øst is already involved, whether concrete procurement has been allocated, what volumes are required, who influences the specification, or why a buyer would select this particular supplier.

Those are not details to fill in with fluent language. They are questions to verify.

What changed

This test changed the way I think about evaluating AI company research. I no longer want to judge a report mainly by its apparent depth, number of links or how convincing its recommendations sound.

I want to know which claims were observed, where their information originated, whether the sources are genuinely independent, how current they are, and which commercial hypotheses survived scrutiny.

The three-pass method also made the role of the research brief more visible. A general-purpose AI assistant can do substantially better when asked to investigate external market developments instead of merely describing a company. The distinction is not simply “AI versus non-AI”. It is plausible synthesis versus evidence-led investigation.

What I think matters

AI can be correct, detailed and relevant without being commercially useful.

It can know a great deal about a supplier while knowing very little about the decisions its future buyers are making.

And the public information environment is uneven. Companies that publish frequently are easier to describe. Quiet buyers can have substantial needs that leave little accessible trace. Repeated corporate descriptions can make one claim look independently confirmed. A project announcement can look like a lead even when the purchasing decision has already been made.

The answer is not to abandon AI research. It is to make the method stricter: trace sources back to their origin, separate observation from inference, look for early changes beyond the company's own channels, and make verification a step before a prospect enters the pipeline.

Knowing a company is not the same as knowing its market. Knowing its market is not the same as knowing who will buy.

What remains uncertain

This was one company and one sequence of research tasks, not a controlled benchmark across models or industries. The selected nine facts cannot support a universal percentage about where AI answers come from. We did not isolate training-data influence or compare multiple independently repeated runs.

The research used publicly accessible materials. Some trade reporting was available only through headlines and previews, and project schedules can change. We did not interview Betong Øst, its customers, project owners or contractors, and we had no access to internal customer, margin, capacity or sales information.

The named projects may already be known to the company. They may never become available business for it. A public absence of evidence about a relationship is not evidence that the relationship does not exist.

What we have demonstrated is narrower and more useful: changing the method changes the questions AI can answer, but even structured external research must stop where the evidence stops.

We found possible doors. We have not established which ones are open.

Sources