What the company knows

Palantir’s investment in ontology rests on a proposition: make a company’s distinctive knowledge more useful without surrendering its advantage. But a company does not have one agreed understanding of its business. Whose judgement guides the work, who benefits, and can the organisation keep learning when its systems already have an answer?

Palantir’s CEO, Alex Karp, argues that companies risk surrendering their distinctive knowledge to the AI providers they employ. In his August 2026 shareholder letter, he describes the “language labs” behind LLMs as “structurally designed to capture” their customers’ organisational and business intelligence. He calls that intelligence their ‘alpha’. It is a pointed accusation from a competitor with a different proposition to sell.¹

Palantir’s proposition is that customers should retain and develop that advantage. Its investment in ontology predates the current generative-AI boom: its 2020 stock-market filing already described experts defining the things and relationships that mattered to their organisations, then building applications around them. The ambition reaches beyond answering questions about a business. It is to make the business’s particular understanding available for decisions, actions and new services.²

Karp is a polarising advocate, and Palantir’s military and government work is contested. Its senior leaders openly argue for technology’s role in strengthening American and allied power. Karp has declared that the company has “chosen sides”; chief technology officer Shyam Sankar has taken a programme for defence reform to the US Senate.³ Palantir’s public political positions and government work have become part of the debate over its suitability for public infrastructure, including the NHS.⁴ A supplier offering customers greater control has commitments of its own.

Palantir’s proposition about organisational knowledge deserves examination on its own merits. Making expertise usable across an organisation can improve coordination and open possibilities that were previously too difficult to pursue. But an organisation does not have one uncontested understanding of its work. People disagree about what matters, what constitutes success and who should bear the cost of achieving it. When those judgements shape the software, they influence more of the organisation’s decisions. Deciding whether they should remains the organisation’s responsibility.

The business behind the data

A hospital can have several perfectly legitimate accounts of the same patient. One concerns treatment, another the next place of care, another the staff and facilities required. Bringing the records together does not settle what they mean. Being medically ready to leave, for example, is different from having the arrangements in place to do so. Software that collapses those distinctions gives the organisation a simpler picture at the expense of a useful one.

An ontology makes selected concepts and relationships explicit. It establishes what a term means in a particular setting, how it relates to other things and which distinctions must survive as information moves between people and applications. The concept is older and broader than Palantir. Its product, called Ontology, connects that representation to the calculations, permissions and actions through which work gets done.⁵

This is the substance behind the investment Karp has championed. The company contributes its understanding of the business; the software makes that understanding available beyond the people who originally supplied it. The next team can start with more than another collection of documents. It can investigate how a change in demand, a missing capability or a different priority affects the work already under way.

An LLM brings a different strength. It can interpret unfamiliar material, draw connections and produce useful responses even where the organisation has not explicitly defined the relevant terms and relationships. That flexibility is valuable. It also leaves room for the model to infer a meaning where the organisation needs a settled one. When an answer starts to allocate capacity, authorise a payment or change someone’s access to a service, that difference has consequences beyond the conversation.

Organisations can hold definitions and controls outside the model in databases, code and tools; they do not need Palantir to do so. Palantir itself uses LLMs, including models developed by OpenAI and Anthropic. Its proposition is to connect those models to the customer’s meanings, controls and actions, while protecting what the customer contributes. The choice concerns how much of the surrounding capability to assemble, how much to buy and where the organisation’s own understanding will reside.⁶

Making that understanding explicit takes work. People must reconcile records, distinguish a fact from an assessment and decide what to do when the evidence conflicts. These are consequential judgements. A familiar term can conceal a disagreement that only becomes visible when someone has to specify what the software should do with it.

The return can extend well beyond the first application. A relationship clarified for one purpose may help solve another problem; an exception previously known to one person can become available to a whole team. The same reach can carry an obsolete assumption farther than it travelled before. An ontology gives an organisation something it can examine and improve. It does not make that examination unnecessary.⁷

Alpha has a commercial life

Airbus’s Skywise partnership with Palantir brings the opportunity into view. The manufacturer knew the aircraft; airlines held knowledge of them in daily use. Combining that expertise and operating data with software supported predictive-maintenance services. In 2018, airlines could receive core platform access in exchange for sharing data, while premium subscriptions offered further capability.⁸

The airlines gained a way to anticipate maintenance needs. Airbus gained another service to offer around the aircraft. The participants were making knowledge more useful by bringing it together, and creating a commercial relationship through which that usefulness could be sold.

They also negotiated who could pursue the resulting market. Palantir’s 2021 annual report disclosed restrictions on independently supplying platforms to certain airlines and Airbus competitors.⁹ The software made a capability possible. The agreement helped determine whose business it could become.

Karp’s shareholder letter contrasts value creation with selling “clicks or tokens or chats”.¹ The distinction is straightforward: consuming more AI does not establish that the business is getting a better result. Airbus was developing another use for its expertise, alongside a new reason for airlines to contribute theirs. The value depended on what they could combine and sell. Keeping knowledge confidential would not, on its own, have created that business.

The supplier is developing too. Palantir’s account of its forward-deployed engineers describes people working on customer problems and contributing improvements to shared products.¹⁰ Customers buy access to expertise accumulated elsewhere; their own work can help the supplier become more capable. Both sides have reasons to participate. Their interests need not remain aligned as the next opportunity emerges.

A supplier does not need the company’s secrets to develop a competing service.

The history of platforms makes the distinction familiar. Microsoft used its control over Windows to favour Internet Explorer; the court found that its conduct made Netscape’s alternative routes to customers more expensive and less effective. The European Commission found that Google favoured its own comparison-shopping service and diverted traffic from rivals.¹¹ The products still worked. The platforms changed how customers found them, what they needed to buy separately and who got paid.

AI opens further possibilities for combining tasks that previously supported separate products or professions. A business can improve its current offering while the platform around it changes how much of that offering customers need. Confidentiality terms address one exposure. They cannot reserve a market.

Karp’s company makes a forceful case for protecting customer intelligence, but hosted-model use does not automatically mean that intelligence trains the provider’s models. The service, configuration and terms matter.¹ The wider commercial question remains even where those protections are strong: what new value is the company creating, and which parts of its position does the relationship strengthen?

The expertise has to keep being produced

An organisation’s knowledge is continually renewed through the people doing its work. Experienced staff recognise patterns, newcomers encounter unfamiliar problems, and both discover when the established answer has stopped being useful. AI changes how that experience is acquired as well as how it is applied.

There is evidence of a substantial opportunity. Research involving 5,172 customer-support workers found that AI assistance increased issues resolved per hour by 15% on average, with larger gains among less experienced workers. The assistant brought suggestions into live conversations, making successful practices available at the moment they were needed.¹²

The researchers also examined outages. Workers previously exposed to the assistant continued to handle conversations faster when its recommendations were unavailable. They interpreted the pattern as evidence consistent with learning. The outage analysis measured speed, however, and could not establish whether resolution quality had also improved.¹²

A separate randomised study asked 52 developers to learn an unfamiliar programming library. Participants given AI assistance performed worse on the subsequent comprehension and debugging assessment. The study found no statistically significant improvement in average completion time. Some assisted participants remained actively engaged with explanations; others delegated much of the work.¹³

These are different settings, not competing verdicts on whether AI helps people learn. Together they make an assumption unsafe: successful use does not tell the employer how much capability is accumulating in the person. Neither study tests ontology. They do establish why people’s understanding deserves attention alongside the performance of the tool.

For Karp’s proposition, this is more than a training issue. The alpha he wants companies to retain has to keep being produced. A platform can preserve a rule or a past decision. Someone still has to recognise when the situation has changed, challenge the rule and develop a better response.

A company can use AI to spread expertise and bring more people into demanding work. It can also organise the work so that fewer people encounter the reasoning behind the answer. The immediate output may look satisfactory in either case. The difference becomes visible when an unfamiliar problem arrives, an experienced colleague leaves or the organisation wants to do something it has not done before.

Represented does not mean influential

Understanding the work is one condition. Having influence over it is another.

HCA Healthcare developed Timpani with Palantir to tackle a difficult scheduling problem across a large hospital group. HCA’s account describes staff preferences, competencies and scheduling rules, with nurses involved in development. Its product owner acknowledged the compromise directly: “Creating a schedule that is perfect for everyone is impossible.”¹⁴¹⁵

A system able to consider more combinations can help. Less time preparing schedules can leave nurse leaders more time with staff and patients. HCA reports a 6% reduction in turnover in departments using Timpani and reduced scheduling work across a deployment spanning more than 130 hospitals.¹⁶ These are provider-reported results; they do not establish how much of the change the software caused.

Nurses and their union contest aspects of the arrangement. National Nurses United alleges that decisions have moved away from local managers towards a central office in Nashville. Nurses interviewed by WIRED describe difficulty changing schedules. HCA maintains that nursing leaders review, adjust and make the final decisions.¹⁷

HCA says those preferences are represented. The dispute concerns what happens afterwards: how much that knowledge influences the schedule, and how readily a local leader can change the result. Recording expertise and giving it authority are different decisions.

Central coordination has a reason to exist: a hospital group has to reconcile demand, scarce staff and resources across many units. A local leader encounters the particular combination of patients, colleagues and experience on a shift. Standardisation can reduce arbitrary decisions and expose avoidable imbalances. It can also make an exception harder to secure when the person requesting it understands something the central view does not.

Sometimes the relevant knowledge is missing. Sometimes it is represented inadequately. Sometimes it is understood perfectly well but loses when competing objectives are resolved. Those are different problems with different remedies. Better data may help the first two; it does not settle the third.

The numbers can make the distinction harder to see. HCA reports honouring 98% of requests not to work on particular days, and separately reports that more than 98% of shifts have a mix of skills and experience.¹⁴¹⁶ Those measures can be useful while leaving important questions unanswered. A mixed shift might still depend heavily on one experienced person. The remaining inconvenience might repeatedly fall on the same staff. Neither possibility is established by the aggregate result.

As a hypothetical illustration, consider the experienced nurse who quietly makes a difficult arrangement work. The final record may show successful delivery without the extra effort, negotiation or improvisation that produced it. If the organisation learns only from that result, it risks treating the rescue as evidence that the original arrangement was sound.

Buying outcomes still leaves a harder question: which outcomes count?¹⁸ Faster scheduling and lower turnover are valuable. So are continuity of care, the development of experienced teams and confidence that professional judgement matters. A monthly performance report may capture a saving well before the organisation can tell whether it is strengthening the people and relationships on which future care depends.

Stakeholders may disagree about what constitutes a good result, not merely about how to achieve an agreed one. If staff or patients lose confidence in the way care is organised, improved headline measures may coexist with a loss of social licence, the trust and acceptance that allow the organisation to operate. That possibility belongs in the assessment of results, alongside the gains.

The organisation can retain the knowledge and still make it difficult for that knowledge to change a decision.

The terms of participation are already being negotiated. In a separate health setting, the University of California’s agreement with the California Nurses Association gives professional practice representatives a role in selecting, designing, building and validating technology affecting nursing care.¹⁹ Their participation begins while the technology is taking shape, when professional judgement can influence the design. The agreement gives them a role in that work, without granting a general veto.

The incentives can change as the capability develops. Expertise contributed to help colleagues might later support an expanded service, fewer staff or a product sold elsewhere. Retaining the company’s alpha does not settle how the people who created it will benefit. The organisation needs to hear about the difficult shift, the workaround and the apparently minor exception. People’s willingness to contribute those observations depends partly on what happened to the last one. A nominal right to challenge means little if the cost of using it falls on the person already carrying the problem.

Shareholders, patients, employees, professional bodies and regulators do not have identical interests in the service. Some disagreements remain choices about whose costs it can reasonably impose. Others may expose an opportunity: different training, earlier coordination or a change in how care is organised could reduce a shortage that scheduling alone can only distribute. The knowledge becomes more valuable when it changes the arrangement that created the difficulty. Boards have to revisit how success is defined, whose experience tests that definition and how affected groups can influence it as the technology and the operating model change.

These interests also shape systems built around LLMs: their instructions, evaluation targets and permitted actions. Making selected judgements explicit in an ontology gives people something to examine. The organisation still decides whose evidence counts and what can change as a result.

The supplier’s politics enter the customer’s choice

Karp’s proposition asks customers to consider who controls the intelligence on which their businesses depend. Palantir itself participates in decisions about technological power. Its submission to the US AI Action Plan advocated stronger restrictions on advanced-chip exports to China.²⁰ Its commitments help explain why some customers choose it and why others question the relationship.

In June 2026, the UK Parliament’s Science, Innovation and Technology Committee recommended moving away from the Palantir-led NHS Federated Data Platform. The committee raised concerns about the company’s political positioning, US military and immigration work, and vendor dependence, while explicitly distinguishing those concerns from product quality.⁴

A service can be useful and its supplier’s wider role still become unacceptable to those accountable for it. In this case, a company’s participation in one state’s exercise of power became part of the debate about another state’s public infrastructure. The objection did not establish misuse of NHS data. It concerned the relationship through which public capability was being built.

By then, the NHS had invested in its data model, operational applications and the people who understood the work. The government has stated that the NHS retains rights in its data model and operational products, with provisions for replatforming. In evidence to Parliament, NHS leaders nevertheless warned that transition could mean pausing development while a replacement caught up.²¹

That puts more than a switching bill into the decision. Patients and staff have an interest in confidence in the supplier, continuity of service and improvements that might be delayed. Rights to retain the accumulated work are valuable. Exercising those rights still requires the means to make that work useful elsewhere.

Nor can the incumbent’s claimed benefits settle the matter without examination. The Office for Statistics Regulation found that the NHS’s before-and-after comparisons could not establish causation and secured clearer reporting commitments.²² Political acceptability, demonstrated performance and the cost of transition each deserve their own evidence. A favourable answer to one does not answer the others.

This is the sovereignty question at the level of organisational knowledge. What a company has learned may be its own, yet its ability to use that learning can depend on software, expertise and a political relationship it does not control. The companion Resource examines what can actually be carried, substituted or rebuilt when that relationship changes.

What the investment leaves behind

Karp is right to direct attention to the organisation’s distinctive understanding. Making more of it usable can expand what the business offers, help people become more capable and expose constraints that were previously worked around one decision at a time. Those are substantial reasons to invest.

That advantage is produced through relationships. It draws on what customers reveal, what employees discover and what suppliers make possible. Keeping the resulting records is only part of keeping those contributions productive. Rights, incentives and the means to question an established answer matter too.

An investment can protect the record of those contributions while weakening the conditions that make the next contribution possible. It can also improve those conditions: giving staff access to expertise, making disagreements examinable and shortening the route from a useful observation to a different decision.

Palantir’s own guidance recommends testing an ontology with unfamiliar business questions and people who did not build it.²³ The original team may know how to work around the gaps; a new team has to rely on what the investment has made available. Difficulty developing a new service can reveal a missing relationship, an unexplained decision or a commercial right nobody secured. A team that understands the problem but cannot change the rule exposes a different limit again.

The company’s advantage has to survive a changing business. Customers need different things, suppliers acquire new capabilities and people discover better ways to work. Ontology offers a way to make accumulated understanding available for those changes. The return depends on the organisation remaining able to develop that understanding and to act on what it learns.

As AI gets better at the company’s business, what is the company getting better at?

Notes

Palantir, AI sovereignty is your alpha, 27 July 2026, and Alex Karp’s Q2 2026 shareholder letter. This is the supplier’s commercial and strategic proposition, not independent proof that adopting its platform protects or increases a customer’s advantage. “Alpha” here concerns distinctive organisational knowledge and tradecraft, rather than a measured investment return. The capture allegation and short quotations are Karp’s claims in the 3 August letter, not findings endorsed by this Signal. His value-creation argument does not establish that every Palantir contract is priced by measured outcomes. The letter names “language labs” rather than individual firms. The July article extends the argument to labs and hyperscalers hosting models; OpenAI and Anthropic are examples of model developers, Amazon Web Services and Microsoft Azure of cloud delivery routes. That article does not name these four companies individually. It concerns permitted use, retention, contractual changes and continued access as well as training. Hosted-model use does not in itself establish transfer of that advantage or training on customer inputs; the service, implementation and terms matter. ↩ ↩² ↩³
Palantir, registration statement, amendment of 3 September 2020, “Making Data Intelligible”, printed pp. 131–132. The claim concerns the documented architecture before the current generative-AI boom, not its invention by Karp or a causal explanation based on his doctorate. ↩
Alex Karp, founder’s letter in Palantir’s 2020 registration statement, printed pp. i–ii, sets out the company’s support for US and allied institutions and says it has “chosen sides”. Shyam Sankar, testimony to the US Senate Armed Services Committee, 28 January 2025, official hearing record, printed pp. 4–7, with his programme, The Defense Reformation, appended at pp. 60–78. The hearing record was printed in 2026; the testimony was delivered in 2025. These are the leaders’ stated positions, not an assessment that every allied government shares them. Sankar’s 18 theses are distinct from the 22-point manifesto discussed in the NHS report below. ↩
House of Commons Science, Innovation and Technology Committee, Rewiring the state: Delivering digital government, 3 June 2026, paragraphs 84–98, especially 84–85 and 93–95. The committee discusses Palantir’s military and immigration work, Peter Thiel’s comments about the NHS and the company’s 22-point manifesto. This supports the connection between public political positioning and the NHS debate; it does not establish that Karp’s 2020 letter or Sankar’s 2025 testimony caused those objections. A parliamentary recommendation, not a government termination or proof of misuse of NHS data. The NHS FDP concerns England; it should not be described as one uniform UK-wide deployment. ↩ ↩²
Natalya F. Noy and Deborah L. McGuinness, Ontology Development 101, Stanford, 2001, pp. 1–3, on the generic concept; Palantir, Ontology overview. The generic concept describes explicit domain concepts and relationships; Palantir’s product includes a broader operational implementation. No ontology contains the whole of an organisation’s knowledge or culture. The generic concept and Palantir’s operational implementation should not be treated as synonyms. ↩
Anthropic, Effective context engineering for AI agents, 29 September 2025; OpenAI, Retrieval; Palantir, AIP architecture and Model Catalog. The catalog explicitly distinguishes model creator from hosting provider, including OpenAI through Microsoft Azure and Anthropic through Amazon Web Services. Availability depends on the deployment. These document context, tools and control around models. They support an architectural comparison, not a performance ranking or a claim that explicit meaning requires one vendor’s ontology product. ↩
Palantir, Ontology system and Ontology permissions. Representation, decision logic, enforcement and operating authority are related but different. Permissions require implementation; ontology does not confer automatic compliance, perfect source data or freedom from model error. ↩
Airbus, Skywise market development, February 2018. Historical service model, not a generative-AI outcome or a demonstration of ontology’s isolated contribution. ↩
Palantir, 2021 Form 10-K, printed pp. 30–31. Disclosed restrictions on independent platform supply to specified categories; current terms and the complete private agreement are not established. ↩
Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 8 April 2019. Describes engineers working with customers and contributing improvements to shared products; not evidence of reusing confidential information. The discussion of changing scarcity and competing offerings is QURKI analysis, not an allegation about a named contract. ↩
US District Court, Microsoft findings of fact, 5 November 1999; European Commission, Google Shopping decision announcement, 27 June 2017. Different documented mechanisms; no inferred copying motive or assertion of equivalent present-day AI conduct. ↩
Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work, Quarterly Journal of Economics 140(2), 2025, especially section VI.B, printed pp. 921–923. The published full sample is 5,172 workers; resolutions-per-hour data are available for a smaller subset. The overall estimate concerns successful resolutions per hour. The outage analysis concerns chat duration, with no chat-level resolution-quality measure. Rare outages, noisy estimates and possible changes in query mix limit the learning inference. ↩ ↩²
Judy Hanwen Shen and Alex Tamkin, How AI Impacts Skill Formation, version 2, 1 February 2026, preprint, main study and qualitative analysis. Randomised study of 52 developers learning Python Trio; immediate assessment, not longitudinal evidence of deskilling. Differences between interaction styles within the assisted group were observational, not independently randomised effects. Authors are affiliated with Anthropic. ↩
HCA Healthcare Magazine, Innovative nurse scheduling solutions transforming patient care, 25 April 2025. The 98% statistic concerns requests not to work on particular days. Its sample, period and distribution are not specified sufficiently for stronger inferences. HCA describes clinical input and human review; these are its accounts of the implementation. ↩ ↩²
HCA’s collaboration with Palantir is documented in the reporting and statements linked under hca-dispute. The public sources do not expose Timpani’s complete ontology, optimisation weights or enforcement rules. The case illustrates the importance of representation, objectives and authority; it does not isolate ontology’s causal contribution to benefits or problems. ↩
Michael Schlosser, HCA Healthcare, Strategic approach to scaling artificial intelligence, 21 July 2026, “Timpani”. Provider-reported outcomes. Earlier monthly hours and later per-cycle hours are not directly comparable. HCA does not specify the calculation basis or reporting period for its 6% turnover figure; it should not be converted into percentage points or treated as a causal estimate. ↩ ↩²
National Nurses United, Nationwide protests against Palantir, 24 August 2026; Paresh Dave, AI is making a mess of nurses’ schedules. They say it’s a safety issue, 2 October 2026. The former is the union’s statement; the latter reports nurses’ accounts and HCA’s response. HCA reiterated its position in Becker’s reporting of 6 October. The April statistic concerns requested days off; the October claim covers scheduling preferences more broadly. The accounts do not establish an undisclosed objective function, a deliberate intent to understaff or a quantified effect on patient outcomes. Requests for particular working shifts, requested days off and mixed-experience shifts must not be treated as the same metric. ↩
Palantir, G-Cloud 15 Service Definition, section 16, printed pp. 27–28, describes capability licences for workflows enabling a specified outcome or set of outcomes. Its accompanying Pricing Document, printed p. 2, lists annual charges for those licences. These support the outcome-focused commercial proposition; they do not establish payment conditional on realised results or disclose HCA’s payment terms. ↩
University of California–California Nurses Association agreement, 22 November 2025–2 January 2030, Article 42, printed p. 144. The employer’s contract page supplies this as the complete agreement despite the PDF’s draft metadata. The article entitles Professional Practice Committee representatives to participate and recommend improvements; it does not establish a general veto, mandatory adoption of their recommendations or improved outcomes. This agreement concerns the University of California, a separate employer from HCA. ↩
Palantir, Submission to the US AI Action Plan, 15 March 2025, printed pp. 11–12. Policy advocacy, not an enacted rule or evidence of a particular customer’s interrupted access. ↩
James Frith, letter to the Health and Social Care Committee, 21 August 2026, pp. 2, 4–5; NHS oral evidence, 16 June 2026, Q100–105. Stated contractual rights do not amount to ownership of Palantir’s underlying platform or an executed migration. The development pause is Rob Thompson’s prospective assessment, not an observed halt or a universal requirement. These documents establish rights and the transition argument, not a completed decision to terminate. ↩
Office for Statistics Regulation, FDP presentation and communication of information and performance metrics, 22 July 2026. Distinguishes action counts from observational before/after comparisons. Its 10 August response to the Health Foundation records the Foundation’s finding of no measurable discharge improvement, but also different patient cohorts in the competing analyses. Neither establishes that all benefits are administrative, that the software is ineffective, or that public claims reflect deliberate manipulation. ↩
Palantir, Ontology design: Validation, accessed 13 October 2026. Supplier guidance, not measured organisational learning. Extending the test to changing an assumption is QURKI’s proposal; the companion develops practical examples. ↩