Your company has an
AI-readable shadow strategy

We were reading those signals long before AI

Decades ago, when I worked at the Evening Post, we looked at advertiser activity to build a view of businesses: how well they appeared to be doing, where they might be under pressure and what they might need. Job advertisements provided another set of clues.

That was commercial intelligence, even when the work was selective and fairly informal. An advertisement was both a message to its intended readers and evidence about the business that placed it. A plausible profile was not the same as knowing the company’s circumstances, but it could inform an approach.

AI did not invent that discipline. The consequential question is what happens when that kind of attention becomes economical across far more companies and commercial conversations.

Scale is the game changer.

That changes the management problem. Traditional disclosure controls tend to ask whether a particular announcement, filing, presentation or web page is accurate and appropriate. AI makes it cheaper to ask a different question: what becomes inferable when hundreds of individually acceptable fragments are analysed together?

The HBR signal is useful, but the underlying change is larger

A 15 December 2025 Harvard Business Review article by Yi Cao and Long Chen argues that AI is changing corporate disclosure because machines can extract strategic insights from information that managers may regard as innocuous. The article points to sources such as job postings and earnings calls. ¹

That is a useful signal, but the Andrew’s Intel angle is not “companies should disclose less”. The deeper issue is what I would call the aggregation gap: the difference between what each function believes it has disclosed and what an outsider can infer after combining the organisation’s public traces.

Legal, regulatory and investor disclosures still have to be made. Recruitment still requires useful job descriptions. Customers still need product information. Marketing still needs to explain the offer. Trying to make the company informationally invisible would be impractical and, in many cases, undesirable.

The problem is that the cost of synthesis is falling. Work that once required substantial specialist time and tedious cross-referencing can increasingly be assisted by general-purpose AI. The commercial implication is not just faster preparation for an important account. It is the possibility that accounts previously too small to justify that preparation become worth researching properly.

Machines were already an audience before generative AI

The idea that corporate information has machine readers is not new. A 2023 Review of Financial Studies paper by Sean Cao, Wei Jiang, Baozhong Yang and Alan Zhang examined the rise of machine readership of corporate filings. Machine readership was proxied through machine downloads and ownership by AI-equipped investors. The researchers found evidence that firms adjusted disclosures as algorithmic readers became more important, including changes in language associated with machine interpretation. ²

That evidence predates the current generation of reasoning models and agents. It does not prove that today’s systems can reconstruct a company’s strategy reliably. It does establish something more basic: managers have been publishing into a mixed human-and-machine audience for years, and firms already respond to that reality.

Generative AI changes the economics again. It is not only a parser. It can summarise, compare, classify, hypothesise and explain. It can move across source types and turn a collection of clues into a narrative an ordinary manager can use.

The analytical skill barrier is falling

A 2025 Review of Accounting Studies paper by Yi Cao, Long Chen, Jennifer Wu Tucker and Chi Wan tested whether generative AI could identify product-market competitors. Machine-generated peer sets overlapped substantially with those selected by human experts and established identification systems; the authors concluded that the method worked reasonably well, especially for large firms. ³

That is a bounded peer-identification task, not evidence that a general-purpose model can reconstruct a company’s strategy. Its relevance here is narrower: some analysis that benefits from specialist judgement can be assisted by a general model and public data.

The management implication is that some forms of analysis may become cheaper and more accessible. The output remains a hypothesis to test, not a fact to act on without challenge.

The same pattern is visible in job-market research. Acemoglu, Autor, Hazell and Restrepo used the near-universe of United States online vacancies to construct an establishment-level measure of AI activity and adoption footprints from AI-related vacancies and changes in skill requirements. ⁴ In August 2026, researchers at the Federal Reserve Bank of Dallas reported using an LLM to classify whether nearly 600 million vacancy postings across five English-speaking countries explicitly offered at least one day a week of remote work, assessing the model against 30,000 human classifications. The working paper reports 98 per cent classification accuracy and remains preliminary rather than peer-reviewed. ⁵

Again, those studies are not about stealing strategy. They show that public operational traces can be used to construct measures of behaviour beyond their immediate purpose—and that AI can make extracting those measures much cheaper.

The disclosure unit has changed from the document to the surface

A communications team can approve an earnings statement. HR can approve a vacancy. Product can approve a web page. Engineering can publish documentation. A senior manager can give an interview. Procurement can issue a public tender. Each item may be legitimate and unremarkable in isolation.

The outside observer sees one company.

That is why the relevant unit is no longer only the document. It is the public information surface: the collection of facts, signals, changes and absences that the organisation exposes over time.

Three properties make that surface especially useful to machines.

First, repetition creates confidence. One job ad can be noise; 14 related vacancies across two locations may look like a programme.

Second, sequence creates meaning. A new executive hire followed by supplier recruitment, altered product language and a new market-facing role can suggest direction even when no single item announces it.

Third, contradiction creates its own signal. If a company says a market is peripheral while hiring aggressively into it, an analyst has a reason to investigate further.

The risk is not limited to a competitor discovering a true secret. A plausible but wrong inference can also spread. Once an AI-generated account of the company is repeated by analysts, candidates, customers or journalists, the organisation may find itself correcting a strategic story it never told.

Scale changes which accounts are worth researching

My newspaper experience was one version of a familiar commercial practice: assemble fragments, form a view and use it to prepare. The constraint was how much attention a particular opportunity could justify.

Lower-cost AI-assisted research could change that calculation. A major prospect may already receive careful preparation; the more consequential gain may be extending useful preparation to ordinary accounts that currently receive little. That is a change in coverage, not merely a time saving on work already being done.

There is also evidence that public disclosure can carry competitive cost. A 2021 Journal of Accounting Research study of investment companies, 13F and EDGAR disclosures and trading strategies found that sophisticated ‘copycat’ firms could exploit disclosed strategies and that this could harm the firms being copied. ⁶ That is a narrow investment-company setting, not direct evidence about operating-company strategy. The management lesson here is an analogy: information can be individually public, lawfully disclosed and still become more strategically valuable when someone has better tools for combining it.

The important distinction is therefore not public versus secret. It is visible versus inferable — and whether useful inference is cheap enough to become routine.

Making the process more explicit, consistent and testable offers a route towards greater objectivity. Producing a confident profile automatically does not. Observed facts, provisional interpretations and conclusions tested against further evidence must remain distinguishable.

A counterparty need not discover a secret

Consider a hypothetical manufacturer. It announces a new facility, advertises for particular production skills and appears in a supplier’s case study about an installation. An outsider might infer an expansion timetable and possible implementation pressure. None of those disclosures says what terms the manufacturer would value, and the inference could be wrong.

A prospective seller could use that hypothesis to ask about delivery certainty or phased implementation rather than assume that price is the only issue. A harder negotiator might instead infer urgency and adjust its demands. Better understanding can reveal mutually preferable terms as well as leverage.

The disclosure question is therefore wider than whether a rival can reconstruct the company’s strategy. It also concerns what an ordinary customer or supplier might conclude about timing, capacity or limited alternatives — and how that changes the conversation.

Do not solve an aggregation problem with blanket secrecy

The obvious bad response is to tell every department to say less.

That would create its own commercial damage. Thin job descriptions reduce recruitment quality. Vague product pages make buying harder. Investor communication cannot be managed as though transparency were optional. Employees and partners need clear information.

A better response is selective discipline.

Some information is required and must be disclosed. Some is commercially useful and earns its place in public. Some is low-value housekeeping that creates little benefit but adds disproportionately to an outside inference. Some is simply stale, contradictory or uncontrolled.

The job is to distinguish them, not to suppress everything.

This is also why the problem belongs across functions. Investor relations cannot see every recruitment signal. HR does not own product documentation. Marketing may not know that the same capability appears in a patent filing. Security teams may focus correctly on confidential data while ignoring what public data becomes when aggregated.

Someone needs responsibility for the combined surface.

The Public Information Aggregation Audit

Before a sensitive launch, market entry, acquisition programme or major organisational shift, run a public-only test against your own company.

On a narrow screen, each question is shown as a separate card.

This is a management screen, not a secrecy score. The output should be decisions: leave required disclosures alone; improve factual consistency; remove genuinely unnecessary exposure where appropriate; and prepare for inferences that cannot or should not be prevented.

Competitive intelligence is becoming easier to democratise

The most important commercial implication may sit on the other side of the same argument.

If your public surface is more machine-readable to rivals, their public surface is more machine-readable to you.

That does not mean every company should build an automated surveillance operation. It does mean competitive intelligence is becoming less dependent on teams of specialists manually reading scattered material.

A product manager can ask for a monitored map of competitor hiring. A procurement team can track supplier capacity signals. A strategy group can compare public investment patterns across peers. A salesperson can prepare an account hypothesis from public changes before a meeting.

The capability creates both exposure and opportunity. Treating it only as a disclosure risk would miss half the management decision.

The strategy you did not publish may still be the strategy the machine sees

Leaders are used to reviewing what the company says.

They now need to review what the company adds up to.

AI does not abolish uncertainty. It can connect the wrong dots, and a polished synthesis can hide weak evidence. The reason to pay attention is not that the system must outperform your most experienced commercial researcher. It is that more counterparties may be able to justify doing the research at all.

The next disclosure question is therefore not merely: is this individual item safe, accurate and necessary?

It is: when this joins everything else we have already made public, what becomes easier to infer about our strategy, operating pressures or negotiating position?

A counterparty need not discover a secret to change how it approaches your business. That is the practical consequence of the aggregation gap — and why scale matters.