The AI Moat Is Moving and Most Enterprises Are Looking in the Wrong Place

Every time one of your best employees corrects an AI answer, a tiny piece of your company’s judgment is created.

Every rejected recommendation reveals a standard.

Every exception exposes hidden expertise.

Every failure teaches the organization something valuable.

And in most enterprises, that learning disappears the moment the session ends.

Now imagine a competitor that captures it all.

It turns expert corrections into evaluations. Exceptions become reusable agent skills. Decisions become training examples. Production outcomes continuously improve its models and workflows.

Both companies may use the same foundation model.

But one is saving time.

The other is compounding intelligence.

That distinction may define the next era of enterprise AI.

Meet Inkling

On July 15, Thinking Machines Lab released Inkling, a new open-weight model built for text, images, audio, coding, reasoning, and agentic applications.

Inkling is massive: 975 billion total parameters, with 41 billion active for each request, and a context window of up to one million tokens. Its weights are downloadable under the Apache 2.0 license, and the model is explicitly designed for fine-tuning and integration into third-party products. 

But Inkling’s size is not what makes it important.

The real message is ownership.

Unlike closed models that enterprises primarily access through an API, Inkling provides a foundation organizations can modify, customize, and deploy through infrastructure they choose.

Thinking Machines did not design Inkling simply to dominate every benchmark. It describes the model as a broad, adaptable foundation that can be customized across different domains, products, and workflows. 

In other words, Inkling is not merely another model to rent.

It is a model companies can begin to make their own.

That changes the strategic question.

For the past three years, enterprise leaders have asked:

Which AI model should we use?

The more important question now is:

What intelligence should our enterprise own?

Access to AI Is Not a Moat

Most enterprise AI products still use some variation of the same formula:

Foundation model + internal data + prompt + workflow + interface

This architecture has created real productivity.
But it rarely creates durable differentiation.

Your competitors can access the same models. They can adopt the same agent frameworks. They can build similar retrieval systems. They can reproduce many of your prompts and workflows. And with every new model release, capabilities that once looked unique become standard features.

An enterprise can deploy thousands of copilots and still create almost no proprietary intelligence. Employees may work faster. AI adoption may rise. Token consumption may soar. But the organization itself may not become meaningfully smarter.

It is consuming intelligence without compounding it.

That is not a moat.

Bridgewater Showed What the New Moat Looks Like

The strongest evidence comes from a separate collaboration between Thinking Machines and Bridgewater Associates.

Bridgewater wanted AI to filter financial information using the judgment of experienced investors.

The questions sounded simple:

Is this article relevant?
Does this central-bank document signal an important policy change?
Does this research help answer the investment question?

But these were not generic classification tasks.

They required Bridgewater’s definition of relevance.

With better prompting, frontier models improved from roughly coin-flip performance to the mid-70% range. But the models tested still failed to reach Bridgewater’s 80% threshold for trusted daily use. 

Bridgewater then used Thinking Machines’ Tinker platform to fine-tune Qwen3–235B, an open-weight model, using examples shaped and verified by its own investment experts.

This was not Inkling. But it demonstrated the same strategic idea Inkling brings to the market: an organization can teach an adaptable model what only that organization knows.

The customized model reportedly achieved:

• 84.7% average accuracy

• 29.8% fewer mistakes than the best frontier model tested

• 13.8 times lower inference cost per task

These were results reported by Bridgewater and Thinking Machines, not an independent benchmark, so they should not be generalized to every workload. But the direction is unmistakable.

The specialized model did not win because it had more general intelligence.

It won because it learned something the frontier models did not know: How Bridgewater thinks.

That is where the AI moat is moving.

Your Most Valuable Intelligence May Be Walking Out the Door

Executives have heard for years that data is their most valuable AI asset.

That is only partly true. Data is often raw material. The greater advantage is the judgment applied to it.

What does your best underwriter notice that others miss?
How does your senior engineer diagnose a failure that has never appeared in a manual?
When does your medical director recognize that a routine case is becoming dangerous?
Why does an experienced compliance officer question a transaction that technically satisfies every rule?
How does a seasoned executive know that a transformation is failing before the dashboard turns red?

This intelligence rarely lives neatly inside a database.
It lives inside people.

It was built through years of decisions, mistakes, exceptions, relationships, and pattern recognition. When those people retire, resign, or are laid off, much of it disappears. That may be the most expensive form of data loss in the enterprise.

There is a painful contradiction unfolding across industries:

Companies are removing experienced people while investing billions in AI systems intended to reproduce expert work.

They may be eliminating the very knowledge their future AI needs to learn.

That is not merely workforce reduction.
It is organizational amnesia.

The Model Is Not the Asset. The Learning Loop Is.

Satya Nadella recently framed the future of the firm around two forms of capital.

Human capital consists of people’s knowledge, judgment, relationships, ingenuity, and pattern recognition.

Token capital is the AI capability a company builds and owns from that human expertise.

His deeper argument is that companies should not search endlessly for one permanent “best model”. They should build a learning system in which workflows, corrections, decisions, and outcomes continuously improve the enterprise’s intelligence.

The underlying general-purpose model should be replaceable.
The organization’s accumulated expertise should not be.

That leads to a simple sovereignty test:

Could your company replace its foundation model tomorrow without losing what it has learned?

If the answer is no, you do not truly own your intelligence.
You rent access to it.

A model — even a fine-tuned model — is only a snapshot.

The enduring asset is the loop:

1. An expert makes a decision.

2. The system captures the context and outcome.

3. A correction becomes an evaluation.

4. An exception becomes an agent skill.

5. A failure becomes a guardrail.

6. Production evidence improves the next model, workflow, or policy.

7. The organization becomes more capable with every cycle.

The model can change.
The cloud can change.
The agent framework can change.

The learning loop must survive them all.

This Is No Longer Just a CIO Decision

Inkling may look like a technical release. Its implications reach the entire leadership team.

For the CEO, it changes competitive strategy.

For the board, it raises questions about intellectual property, workforce risk, accountability, capital allocation, and enterprise value.

For the CFO, it shifts the conversation from token cost to the economics of proprietary intelligence.

For business executives, it changes how scarce expertise can be preserved and scaled.

For CIOs, CTOs, CDOs, and AI leaders, it changes model architecture, infrastructure, vendor dependency, and operating models.

For the CHRO, it changes succession planning, knowledge retention, and workforce transformation.

For legal, security, risk, and compliance leaders, it moves more accountability for AI behavior inside the enterprise.

Boards should therefore ask more than:

• How many AI pilots have we launched?

• How many employees use copilots?

• Which model provider have we selected?

• How much productivity have we gained?

They should ask:

What proprietary intelligence are we creating — and who owns the learning loop?

That may become one of the most important board questions of the AI era.

Open Weights Do Not Automatically Create Sovereignty

Inkling gives enterprises more freedom. It does not automatically give them control.

A company can own model weights and still be locked into:

• A fine-tuning platform

• A cloud provider

• An inference runtime

• A proprietary agent framework

• A memory architecture

• An evaluation platform

• A vendor-controlled governance layer

Real AI sovereignty means the enterprise can:

• Replace the base model

• Move workloads between providers

• Preserve its evaluations and agent skills

• Reproduce previous versions

• Roll back unsafe changes

• Control sensitive data flows

• Continue operating during vendor disruption

• Audit consequential decisions

Sovereignty is not a license.
It is not self-hosting.
It is not owning a checkpoint.

It is an engineered property of the complete intelligence system.

Freedom Also Transfers Accountability

When an enterprise customizes a model, it changes the model’s behavior.

Accuracy may improve.
Safety may deteriorate.
Bias, hallucinations, refusal behavior, security exposure, and tool use may all change.

Thinking Machines’ own model documentation emphasizes that downstream applications require use-case-specific controls, evaluation, monitoring, and human oversight. 

The enterprise therefore inherits more responsibility for the resulting system.

Governance can no longer remain a policy document, a procurement checklist, or a quarterly committee conversation.

It must operate inside the learning loop:

• Which data may be used?

• Which expert corrections are trusted?

• Which evaluations define acceptable performance?

• What evidence is required before release?

• When must a human intervene?

• How are failures detected and contained?

• How can the system be rolled back?

• Who is accountable for the outcome?

Open models create freedom. They also transfer accountability.

Why Agentic Engineering Becomes the Differentiator

Most enterprises still build AI through disconnected activities.

A model team selects the model.
A data team prepares the information.
A product team builds the interface.
A cloud team deploys it.
Security performs a review.
Legal writes a policy.
Business leaders sponsor a pilot.

No one owns the behavior — or the learning — of the complete system.

That fragmentation was already limiting for copilots.

It becomes dangerous when customized models and autonomous agents begin using tools, retaining memory, making decisions, and acting across enterprise workflows.

The enterprise needs an end-to-end discipline connecting:

• Models

• Data and context

• Memory

• Tools and workflows

• Agent skills

• Evaluation

• Human oversight

• Runtime controls

• Security

• Governance

• Continuous learning

That discipline is Agentic Engineering.

Agentic Engineering is not another name for prompt engineering, fine-tuning, or building agents.

It is the discipline for designing, deploying, governing, and continuously improving AI systems that pursue goals inside real organizations while remaining reliable, secure, observable, and economically accountable.

This is why the Agentic Engineering Institute (AEI) is codifying the discipline through the Agentic Engineering Body of Practices (AEBOP™), the Agentic Enterprise Governance Framework (AEGF™), the Agentic Operating Model Framework (AOMF™), professional education, and its global expert community.

Because the enterprise bottleneck has moved.

Yesterday, it was access to powerful models.

Today, it is turning models into trustworthy production systems.

Tomorrow, it will be converting organizational experience into continuously improving intelligence — without surrendering control.

The AI Moat Is Moving

Imagine two competitors using Inkling — or any other capable foundation model.

The first connects it to internal documents and launches another assistant.

The second captures how its best people think. It builds proprietary evaluations, reusable agent skills, workflow memory, governance controls, and a learning loop that improves with every real-world outcome.

The first gains productivity.
The second builds an appreciating asset.

The model may be identical. The enterprise capability is not.

The first era of enterprise AI was about accessing intelligence.
The next will be about owning differentiated intelligence.

OpenAI, Anthropic, Google, Inkling, and the models that follow will remain important.

But none of them will be the ultimate enterprise moat.

The moat will be the organization’s ability to turn human judgment, proprietary data, workflow experience, and production outcomes into intelligence that competitors cannot easily reproduce.

You can outsource a task. You can replace a model.

But you cannot outsource your organization’s ability to learn.