Capability or Trust? The Choices Facing Enterprise AI

So who is the “them” in all of this?

It is the frontier model vendors. As Anthropic, Google, OpenAI, and others release updates at a relentless pace, organizations are being forced into a choice they have largely tried to avoid. They now have to decide between capability at all costs and trusted operations. Unfortunately, with the latest gameplay, you cannot have both with uncontrolled, untethered frontier models.

That will not surprise anyone who really understands how this works. Nor will it surprise anyone with a few battle scars from watching the tech economy swing wildly from build-it-yourself, to an app for everything, to SaaS for everything, and now to AI for everything … while scrambling to make sure they do not get architected out of their own business.

What we casually continue to call a “model”, that intelligent orb somewhere in the ether, has quietly become insecure and unreliable agenticware. Make no mistake, today’s frontier systems cannot do what they are doing without a growing tangle of harnesses operating behind the scenes. Hidden scaffolding, state machines, reasoning loops, tool chains, and external dependencies all running behind a curtain you cannot see, never mind inspect. That is a world away from earlier LLMs that relied on relatively straightforward probabilistic passes over their weights.

Once you hand company-sensitive data to a frontier system, you are at its mercy over how that information is collected, routed, and retained. Your accounts, contracts, legal opinions, customer lists, product roadmaps, formulas, and blueprints do not simply pass through a neural network and vanish. They end up in an agent, a log, a scratchpad, or a database controlled by someone else’s architecture. Because the interface still looks like a simple text box, organizations have convinced themselves that the trust problem is equally simple.

It is not!

Capability Is Slipping Sideways

We are also not gaining as much ground as the model release pace implies. Nor are current models setting off the “intelligence threshold alarms” that some of the promotional narrative in Anthropic’s S-1 would have us believe. Yet every vendor will humblebrag that its latest model has dramatically moved the bar. In many practical, real-world scenarios, newer models have actually regressed through overfitting, synthetic training data, biased safety weighting, and other increasingly incestuous tendencies in modern model development. You tune a process around a system that works beautifully on Tuesday, only to find yourself arguing with it on Thursday because an underlying weight, routing rule, safety constraint, or inference budget shifted without you ever being notified.

I’m not aware of any reputable enterprise that can tolerate that level of indifference or swing in outcome.

At the same time, too many organizations are confusing output volume with productivity. They rely almost exclusively on AI to churn out marketing copy, slide decks, emails, and reports, only to produce neutered patterns devoid of novelty that miss the mark with customers and steadily erode their credibility. That failure is happening with increasing frequency as enterprises allow the model to become the rapid author rather than the tuning tool.

In that world, the model has taken the reins and we have become far too complacent about trusting where it is taking us. Tail wagging the dog?

The actual productivity numbers are beginning to challenge the hype as well. In its November 2026 Monetary Policy Report, the Bank of England found that firms estimate AI has increased productivity by only around 0.5% over the past three years. For all of the investment, excitement, infrastructure, and organizational disruption surrounding AI, that is remarkably low. By almost any market standard, it is hardly the return you would expect from something supposedly revolutionizing the economy.

Apparently revolutions take a little longer once someone starts measuring the real results.

Meanwhile, in the back office, organizations are quietly creating an entirely new army of humans to check the work produced by AI. Someone has to review the generated code, validate the analysis, clean up the marketing copy, verify the citations, check the numbers, and make sure the confidently produced answer is not complete nonsense. We are automating the assembly line of work at extraordinary speed while simultaneously creating the AI quality-control function.

Don’t get me wrong, there is real productivity to be gained from AI, and I see it every day. But making a skilled person materially better at their job is very different from simply flooding the enterprise with more AI-generated output. One brilliantly augments human capability. The other can just as easily compound the noise, the rework, and the mediocrity.

Mounting Costs and Premature Execution

Then there is the bill. Those hidden agents behind the curtain can chew through tokens faster than a finance team can say “ouch,” triggering multi-step searches, verification loops, retrieval passes, and subagent calls on a single prompt. More reasoning naturally requires more compute, yet the real architectural hazard is that economic trade-offs are now buried inside someone else’s execution logic.

When a vendor’s harness decides how much reasoning is enough, all by itself, or when some internal spend ceiling has been reached, cost optimization starts turning into massive operational risk. That leads directly to unintended consequences like Premature Execution, where an agent truncates its analysis, skips a verification step, or acts on incomplete context simply because the system has decided it has spent enough.

“Good enough” might pass when drafting an article, but inside financial operations, security remediation, or software deployment, a hidden cost and profit calculation has no business deciding when an action is safe to execute.

Choice Over Dependency

Navigating this does not mean rejecting frontier models. On the contrary, they are enormously valuable and we use them extensively ourselves. We’re just very careful about when and how we use them, and what data we feed them. Secure model routing and sensitive data protection are quickly becoming the new network of choice. Not all models are created equal, and those fancy vendor humblebrags are a terrible way to make a real-world decision. There is also a growing range of choices in the open-weight world, and MCP is making some of that interoperability easier. But, and this is a big but … you still need the internal intelligence to decide what to use, when to use it, why you are using it, and how much you are willing to spend.

There is a decisive difference between model choice and model dependency. Choice means you can swap the reasoning engine tomorrow without redesigning your business. Dependency, on the other hand, means your data structures, state management, security assumptions, and execution logic slowly fuse to one vendor’s brain, and its harness, one convenient integration at a time. It all seems fine and dandy when you are dealing with an immediate requirement and just need to get something done fast, or as some would say, “quick and dirty.”

But in the enterprise, capability is only one part of complex decision-making. A frontier model has no inherent understanding of the fiduciary responsibility of the company, the value of its intellectual capital, a customer’s expectation of privacy, or why one piece of sensitive information should never cross out into the wild unprotected. We humans impose those constraints, and we think about them every day.

The models do not have that care or understanding. They don’t get fired when something goes awry. Nor do they care about the fines, reputational damage, or what happens when regulators or the FBI come calling. Humans inside the enterprise carry that accountability. The model does not.

Increasingly, these frontier models also come with agents attached. You might not have realized that, but they are there. Those agents may be handed a task, but that does not mean they have the authority, clearance, or business context to do whatever is technically necessary to complete it. Being capable of taking an action is not the same thing as having the right to take it, or carrying any accountability for what happens next.

That matters a great deal more today than it did even a few months ago, because the models have stopped simply answering questions. They are increasingly acting on their own logic, using tools, making decisions, and executing on our behalf.

Wrangle Back Control

Building a solid enterprise AI architecture requires designing for choice and trust at the same time. Choice protects you from the volatility of the frontier race as models leap forward, regress, and reprice every few months. Trust is what stays constant underneath, ensuring the enterprise independently controls identity, policy enforcement, sensitive data boundaries, execution authority, and audit evidence regardless of which model happens to be doing the reasoning.

AI has a vital place in the modern enterprise, and the frontier vendors will keep building extraordinary tools. Let them compete to build the smartest reasoning engines in the world. Just do not blindly hand them the keys to how your business operates. If trust is going to be built into enterprise AI, organizations need to wrangle back control of the architecture from the start.