AI Has Intelligence to Spare, but Not Enough Trust: Two Quiet OpenAI Moves Address the Bottlenecks in Enterprise Adoption

In July, OpenAI quietly launched an enterprise product called Presence. Go back a little further, and in February this year it released another product, Frontier.

Neither product received much fanfare. Neither has reached general availability or a self-service stage, and both still require FDEs and systems integrators to help deploy them for enterprises. So many people may not have noticed them.

But having taken a look, I think they offer a glimpse of how a leading large-model company is placing its ‘bets’ on enterprise AI adoption.

This company, which has long made its business out of ‘the smartest large language models,’ is now most concerned with finding ways to constrain them, control them, and make them auditable.

This may also represent a future direction for the AI industry. Today, let's talk about this interesting change.

The smarter AI gets, the less enterprises dare to use it

For the past two or three years, the AI industry has been doing almost one thing: making models smarter.

Early on, everyone talked about ‘model as a service’ (MaaS) and ‘the model as the product or capability.’ The model itself could make money. Then came an upgrade every few months: stronger reasoning, longer context, better coding, higher benchmark scores.

But have you noticed? By now, everyone is numb to ‘the model got smarter again.’ It's like... ‘Oh, okay. And then what?’

What does GPT becoming another 20% smarter actually mean for a company? Often, people feel nothing. There is no direct significance.

Because what really holds enterprises back is no longer ‘Can AI do it?’ but ‘Do I really dare hand this over to it?’

There is a central contradiction here:

AI intelligence is starting to become abundant, but AI trust remains scarce.

Take the simplest example.

If AI does 90% of my work, but I still have to recheck every single thing it did before I can feel reassured, then it hasn't really reduced my workload. It has only changed my work from ‘doing’ to ‘checking.’ And checking is sometimes no less burdensome than doing it myself.

Enterprises want more than intelligence: they want a closed loop of responsibility

Why is this shift happening?

We don't need a long list of reasons. I think two are enough.

Reason one: AI has moved from ‘giving answers’ to ‘taking action,’ and the stakes have changed.

In the chatbot era, AI giving a wrong answer had a limited impact.

In the agent era, AI sends emails, changes data, issues refunds, operates systems, contacts customers, executes approvals, and manages accounts.

Once AI has the authority to act, the nature of errors changes completely. A wrong answer is embarrassing; an email sent to the wrong person or an incorrect refund is an incident.

So the more capable AI becomes, the more enterprises need to know how to constrain it.

Reason two: enterprises aren't buying tokens. They are buying work outcomes.

No enterprise will pay for ‘how many billions of parameters this model has.’

What they will actually pay for is: how much are customer-service costs reduced? How many support tickets are resolved automatically? How much shorter is the claims cycle? How many sales leads are processed? Does an employee who previously handled twenty things a day now only need to handle five?

So the unit of value changes: from tokens (compute), to tasks, to workflows, to outcomes, and finally to responsibility.

That brings us to a central judgment:

The commercial value of enterprise AI is moving from ‘providing capability’ toward ‘taking responsibility.’

Whether AI can truly take a job depends on containing the 10% of exceptions

Here is something I've felt very deeply over the years.

What really determines whether AI can take a job isn't the 90% of normal situations. It's the remaining 10% of exceptions.

The truly valuable capability is knowing what to do when that other 10% occurs. It's what we often call fault tolerance, or a fallback mechanism.

The same applies in real life. Here's a simple example.

When hiring an employee, you naturally want someone competent. Your default expectation is that they can do more than 90% of the job correctly. If they can't do, or get wrong, fifty out of a hundred routine tasks every day, you certainly won't hire them. So that 90% of competence is just a standard, or a baseline.

But what really separates an ordinary employee from a highly valuable one has never been that 90%. It's what happens in the remaining 10%, where the SOP doesn't cover the situation and an error is possible. Do they know that they don't know? Can they identify an exception, assess risk, stop, find the right person, explain the problem clearly, preserve evidence, propose a solution, and turn the experience into accumulated knowledge afterward?

A mature person doesn't never make mistakes. They know how to stop a mistake from becoming an incident, and can build a complete fallback mechanism so that it doesn't affect the operation of the entire company or business process. That is the core of truly valuable talent.

The same logic applies when we write code. With vibe coding, the code may run normally 99% of the time. But there is still a 1% possibility that falls outside everything you could anticipate while designing it: the network goes down, the database times out, a third-party API returns data you've never seen, or a user enters something you never imagined.

What demonstrates engineering maturity isn't whether the happy path works. It's whether, when that 1% failure occurs, you can catch the error, keep AI from hiding it, make it report the failure quickly, and handle it effectively so it doesn't keep affecting users. An engineer with that mindset and the ability to design fallbacks is truly valuable.

The happy path demonstrates capability. The exception path determines trust.

Product design used to focus heavily on the happy path. In the agent era, we need to devote substantial effort to designing the exception path instead.

This is what a ‘closed loop of responsibility’ looks like in practice. Fundamentally: I don't require you never to make a mistake, but I want every kind of error to have somewhere to go, as far as possible.

Presence: OpenAI starts trying to hand AI a real job

That's why these two recent OpenAI products deserve a closer look.

Let's start with Presence.

Many people's first reaction is: OpenAI has built another agent that handles phone calls, claims, and IT tickets. But honestly, a demo that can handle those tasks isn't especially remarkable anymore.

What is worth looking at in Presence is how much product capability OpenAI has devoted to what happens when an agent ‘can't do it, gets it wrong, is unsure, crosses a boundary, or needs approval.’

Think of hiring an employee. You don't just ask whether they're smart. You go through a whole process: define the job, train them, provide knowledge, create system accounts, assign permissions, establish SOPs, specify what needs a supervisor's approval and who to contact when something goes wrong, review performance, and provide ongoing training. Presence essentially brings that entire process to the agent.

So Presence isn't selling a smarter agent. It's trying to train an agent into an employee who can truly take up a job.

Previously, when building agent products, we wrote a Feature Spec: what it can do.

What Presence tells us is that a real agent product needs a Job Contract: what it is responsible for, what permissions it has, where the boundaries lie, when approval is required, when it must hand over to a human, what counts as doing the job well, and how it keeps improving.

That is the ‘closed loop of responsibility.’

Frontier: with more AI employees, companies must start building an organization

Now let's talk about Frontier.

With one agent, there is no management problem. But when an enterprise has ten, fifty, or hundreds of agents, the problem changes.

It is no longer ‘Can each agent do the work?’ It becomes:

How many agents does our company actually have?

Who is responsible for what?

Who created them?

Who is accountable for them?

What can they access?

Which permissions can be granted?

Which require approval?

What have they done recently?

How much have they spent?

How are they performing?

Which agent should be taken offline?

These are the things Frontier is trying to solve.

Put that way, does Frontier sound like an ‘HR department for AI employees’? I think even that analogy is too narrow.

Running a company in the real world requires organizational structure, a job framework, permissions, IT systems, SOPs, audits, performance management, workflows, and so on. Simply put, Frontier is trying to rebuild that entire organizational operating infrastructure for AI employees. I think that's a more accurate description: an operating foundation for an AI workforce of agents.

At this point, the problems each product solves are clear—

Presence: how to get one AI employee genuinely on the job.
Frontier: once you have a group of AI employees, how to build a company to manage them.

OpenAI is betting on trustworthiness, not greater intelligence

Put Presence and Frontier together, and what OpenAI is really betting on becomes obvious.

It is betting not on ‘smarter AI,’ but on ‘execution that can be trusted.’

OpenAI's route used to be clear: model → interface → application.

Now that chain is getting longer: model → agent → workflow → responsibility → an entire AI workforce.

This also answers the question at the beginning. The smarter AI gets, the less enterprises dare to use it. This isn't because enterprises are conservative. There is an enormous gap between ‘smart’ and ‘able to be safely entrusted with a whole job.’

The unit of competition in AI products is shifting from ‘model capability’ to a ‘closed loop of responsibility.’

This isn't far removed from us

For enterprises, this changes how AI is put into practice.

In the future, enterprises shouldn't start with ‘We need to deploy an agent.’ They should start with ‘Which jobs can we hand to AI?’ Then divide them into three levels by risk:

Level 1: AI does the work; humans check it.

Level 2: AI completes the work; humans approve key steps.

Level 3: AI handles most situations autonomously, involving humans only for exceptions.

Putting AI into practice is no longer fundamentally about deploying a model. It is about redrawing the boundaries of responsibility between humans and AI.

For an AI builder or a one-person company (OPC), this is even closer to home. We should, in fact, already be experiencing it.

When a person has only one ChatGPT, there is no management problem. But once there are several tools—for orchestration, coding, research, and content—the question immediately changes from ‘Which model is best?’ to ‘Who exactly is responsible for what?’

So for builders, what should we really focus on when creating AI products in the future? I think there are three points.

First, don't start with a ‘universal agent.’ Start with ‘a job.’ Instead of asking ‘What can my agent do?’ ask ‘What outcome am I willing to make it responsible for?’ Write it a job contract first. Define the product through the job.

Second, product defensibility may increasingly come from the exception path, rather than the happy path. Demos are easy. What's genuinely hard is: what happens when it gets something wrong, doesn't know, exceeds its permissions, or the system fails? When does a human step in?

Third, new opportunities may emerge in ‘AI workforce infrastructure,’ but that doesn't necessarily mean building a general-purpose platform to compete directly with OpenAI. The bigger opportunity may be going deep into a particular industry on ‘how an AI employee can genuinely take responsibility for a job.’ In other words, industry-specific AI workforces and vertical agent operations.

Finally, let's return to the beginning.

For the past few years, we've kept asking: how much smarter can AI get?

In the years ahead, the more important question will be: how much responsibility do we dare give it?

The sign that AI has truly entered the enterprise won't be another benchmark score, or how many tools it can call. It will be the day a manager can hand over an entire job and go do something else with peace of mind.

On that day, the destination of AI commercialization will no longer be ‘smarter,’ but ‘more worthy of trust.’