The Knowledge We Buried Is Moving Again

Salesforce has crossed an interesting architectural boundary.

In September 2026, Salesforce and NVIDIA announced Koa, Salesforce's first CRM-specific reasoning model. Salesforce says the model was post-trained on a proprietary synthetic dataset representing nearly three decades of CRM processes, workflows and operational policies.[1]

Marc Benioff described Salesforce's accumulated knowledge of how enterprise business operates as one of the company's most valuable assets. Koa, he said, puts that knowledge into the model.

One qualification is essential:

Salesforce says no customer data was used to train Koa.

The training scenarios were synthetic. They were designed to reflect the reasoning, tool use and decision-making involved in CRM work across industries. The argument is therefore not that Salesforce has taken its customers' proprietary knowledge or intellectual property.

The more defensible question is also more consequential:

How does decades of exposure to and participation in enterprise operating practices become proprietary model intelligence?

This is not only a question about Salesforce. It is a question about the direction of enterprise AI.

From operating patterns to model capability

Koa is significant because Salesforce describes a clear progression.

Enterprise activity produces processes, workflows, operational policies and recurring patterns of work. A vendor participates in that activity over decades by building applications, encoding common practices and learning what enterprise users need. Synthetic scenarios can then represent those practices without reproducing customer data. Those representations become training material. Training produces specialised model capability.

The progression looks something like this:

Institutional work → enterprise applications → accumulated operating patterns → synthetic representations → proprietary model capability

Every step may be legitimate. Every step may create value.

Salesforce has built substantial expertise in customer relationship management. It can reasonably convert its own expertise into intellectual property. A specialised model may perform CRM work more reliably, use tools more effectively and reduce the burden on people navigating complex systems.

The architectural question emerges from the aggregate direction. As accumulated operating knowledge becomes model capability, the model may become more than a mechanism working with the institution. It may become an increasingly important location of the intelligence through which the institution understands and conducts its work.

The knowledge we buried can move another layer deeper

In The Knowledge We Buried, DataMPowered argued that decades of enterprise technology created distance between people and institutional logic.

Rules became configuration.

Processes became workflows.

Authority became permissions and approval structures.

Organisational context became tickets, records and dashboards.

Institutional memory became distributed across applications, integrations, documents, vendors and people.

The knowledge did not disappear. It became difficult to reach, understand and challenge at the moment work occurred.

AI appeared to offer a way back. Natural language could help people begin with their situation rather than the name of the application, field, form or process they were expected to navigate. It could reconnect people with capability their institutions already possessed.

But another possibility now deserves attention:

The buried knowledge can move one layer deeper, into the model.

That movement may initially feel like recovery. The model can interpret the records, traverse the workflows and explain the process. It can make the systems beneath it easier to use.

Yet accessibility and authority are different properties.

If the model becomes the principal place where rules are interpreted, operating patterns are understood and work is coordinated, the institution may gain a better interface while becoming more dependent on an intelligence layer it does not independently possess.

Do not infer what can be established

The earlier perspective Are We Building Institutional Capability Inside AI When We Don't Need To? proposed a simple principle:

If something can be authoritatively established, establish it. Use inference for what genuinely needs to be inferred.

If the institution can establish a record, retrieve the record.

If it can execute a rule deterministically, execute the rule.

If it maintains a delegation, verify the delegation.

If an API provides a capability, invoke the capability.

If a workflow has a known state, inspect the state.

If a policy is authoritative, preserve the policy as an independently governed institutional source.

AI should reason where reasoning adds value: interpreting language, resolving ambiguity, synthesising evidence, recognising patterns and working through genuine uncertainty.

This does not mean models should contain no knowledge. Specialised knowledge is part of what makes models useful. The distinction is between model capability and independently authoritative institutional capability.

Capability should not migrate into the model merely because the model has become capable of reproducing it.

More capable intelligence makes the boundary more important

The wider frontier-AI trajectory makes this architectural choice more consequential, not less.

In The Gentle Singularity, Sam Altman describes a trajectory towards digital superintelligence, rapidly improving systems and intelligence becoming dramatically more abundant.[3] These are forecasts and ambitions, not evidence that artificial general intelligence or superintelligence has already been achieved.

Even so, the direction matters.

As models become more capable, the argument for allowing them to interpret more context, coordinate more work and make stronger recommendations becomes increasingly persuasive. Each additional delegation may be rational because the model performs the task well.

The question is not merely:

How capable will AI become?

It is also:

What should remain outside the model even when the model becomes capable of doing it?

That question cannot be answered by model performance alone. It concerns institutional authority, resilience, contestability and the ability of an organisation to understand how it works.

Authority can be surrendered without being seized

Altman described another path during a July 2025 fireside discussion hosted by the US Federal Reserve.[4]

He asked what might happen if an AI recommendation became so consistently strong that a human decision-maker could not improve on it and could no longer fully understand it. In any individual case, following the recommendation might be reasonable. Across society, however, repeated deference could transfer a significant share of decision-making to systems evolving in ways people do not completely understand.

This is sometimes usefully described as silent surrender. It is not the title of an Altman essay. It is a description of the phenomenon he outlined: consequential authority can move gradually because deferral appears rational, not because a machine dramatically seizes control.

The institutional version can unfold just as quietly:

• enterprise AI begins with useful assistance;
• specialised models understand domains more deeply;
• agents coordinate a larger share of work;
• recommendations become more reliable;
• human intervention adds less apparent value;
• organisations defer more frequently;
• independent institutional capability begins to atrophy.

The institution may continue to hold legal responsibility. People may still approve the result. Policies may still state who has authority.

But formal authority does not by itself preserve practical capability.

An organisation may still formally possess authority while becoming increasingly dependent on intelligence it does not independently possess.

A person asked to approve a recommendation they cannot understand is not necessarily exercising meaningful judgement. An institution unable to operate when a model or provider changes does not fully control its own operating capability.

Whose knowledge becomes valuable?

The economic question requires care because not all organisational knowledge is the same.

Institution-specific knowledge includes local delegations, internal interpretations of policy, institution-owned definitions, organisational constraints, historical decisions and context particular to one institution.

Generalised operating knowledge includes common CRM patterns, sales processes, service-management practices, recurring organisational workflows and reusable reasoning about enterprise work.

A vendor can legitimately accumulate expertise in the second category. Building products for many organisations creates insight, intellectual property and reusable capability. Synthetic training can transform that expertise into a model without training on customer data.

The question is not whether vendors are allowed to learn.

The question is what happens economically when decades of enterprise activity make possible increasingly valuable proprietary operating intelligence.

When accumulated organisational practice becomes proprietary model capability, who owns the resulting intelligence and who ultimately captures the productivity dividend?

Customers may receive better software, lower error rates and more productive work. Vendors may receive a defensible model asset whose value compounds across customers. Both can benefit.

But institutions should still ask whether the productivity improvement leaves them with more durable capability of their own, or with a deeper dependency on the provider through which the work is interpreted.

Infrastructure control is not operational sovereignty

Salesforce's announcement also highlights control of model weights, inference inside its trust boundary, and support for private-cloud or air-gapped environments for government and regulated organisations.[1]

Those capabilities matter. Infrastructure location, data protection, security controls and deployment boundaries are legitimate dimensions of sovereignty.

Salesforce's AIforce announcement makes a related proposition. It brings data, workflows, business logic, semantics, permissions, security and governance already held in Salesforce to AI interfaces, with requests governed through existing permissions and business rules.[2]

That can preserve important controls. It also reveals why infrastructure control alone is incomplete.

Infrastructure sovereignty does not automatically establish semantic or operational sovereignty.

A model can operate inside an organisation's trust boundary while still becoming the place where institutional meaning is interpreted and operating logic is assembled. A system can respect access controls while concentrating practical understanding in a proprietary intelligence layer.

The Sovereign Operational Intelligence question is therefore broader:

• Where does institutional meaning live?
• Where are rules authoritative?
• Where does organisational memory reside?
• Can the institution explain why work occurred?
• Can it change models without reconstructing how the organisation works?
• Can people still challenge the model?
• Can authorised people still exercise meaningful judgement independently of it?

Operational sovereignty does not require institutions to build every model or reject global technology providers. It requires the institution to remain the authoritative source of its own meaning and action.

An alternative architecture

The alternative is not to isolate models from institutional work.

It is to place them within an architecture where the institution remains authoritative.

A practical direction is:

Human intent → relevant context → institution-owned capability, knowledge, rules and authority → appropriate mechanism → governed work → human judgement where required → accountable outcome

The appropriate mechanism may be a model. It may also be an API, application, deterministic rule, workflow, database, document, calculation, person or specialised tool.

Models can interpret what someone means. They can synthesise material, identify uncertainty, propose pathways and reason about situations that do not have predetermined answers.

Institutional capabilities can establish records, apply authorised rules, execute transactions, preserve evidence and maintain the identities of responsibilities and delegations.

The model does not need to become the authoritative home of the institution in order to make institutional capability accessible.

This architecture also allows models to improve without requiring the institution to migrate its identity into each new generation of intelligence. Models can remain powerful and replaceable. Institutional capability can remain durable and governable.

The boundary institutions must choose

Salesforce Koa does not prove that enterprise knowledge should never become model capability.

It demonstrates that the movement is already technically and economically plausible.

Accumulated operating patterns can be represented synthetically. Synthetic representations can become specialised reasoning capability. That capability can make enterprise work easier, faster and more reliable.

It can also create a new centre of dependency if institutions do not decide what must remain independently authoritative.

We spent decades burying institutional knowledge inside systems.

AI gives us an opportunity to recover it.

We should be careful that we do not instead bury it one layer deeper, inside models that eventually become easier to trust than ourselves.

The primary question is:

What knowledge should become model capability, and what knowledge must remain an independently authoritative capability of the institution?

And behind it sits another:

Are we using AI to make institutions more capable of understanding and governing themselves, or are we gradually creating intelligence they cannot operate without?

References

[1] Salesforce, "Announcing Koa: Salesforce's First CRM Reasoning Model, Built on NVIDIA Nemotron", 15 September 2026. https://www.salesforce.com/au/news/stories/koa-reasoning-model/

[2] Salesforce, "Salesforce Unveils AIforce, Bringing the Full Power of Its Platform to Any Interface", 15 September 2026. https://www.salesforce.com/news/stories/aiforce-announcement/

[3] Sam Altman, "The Gentle Singularity", 10 June 2025. https://blog.samaltman.com/the-gentle-singularity

[4] Board of Governors of the Federal Reserve System, "Fireside Chat: Vice Chair for Supervision Michelle W. Bowman and Sam Altman, OpenAI CEO", 22 July 2025. https://www.federalreserve.gov/mediacenter/files/capital-framework-conference-fireside-chat-transcript.pdf