The sentence at the bottom of the page
On 22 July, OpenAI announced a product that does exactly what a lot of mid-sized companies have been circling for two years. AI assistants that answer customer questions, resolve issues, reach into your own systems, take approved actions, and hand over to a person when they should. On the phone and in chat.
Which is to say: the thing we build. From the largest possible supplier, as a finished product.
Then you read to the bottom of the page, where there is a sentence that is easy to miss. It is not available self-serve. Every rollout is led by the supplier’s own engineers, who sit down with the customer and pick the workflow. They connect the systems and set the rules for what the assistant may do and when a person takes over.
The company with the best AI models in the world is not selling software. It is selling software with consultants around it.
Quick disclosure before I go on: I make my living from exactly the work that is left over here. Apply the appropriate discount to everything that follows.
If the supplier with the best models on earth cannot ship this as a self-serve product, a mid-sized company is unlikely to click one together on its own.
The bottleneck moved
The first line of their own announcement says it: the challenge is no longer proving that AI assistants can work. It is making them reliable enough to do real work.
That matches what Forrester wrote in June. A research firm that sells studies rather than AI assistants, so it has nothing to promote here. The title of their stocktake says it all: companies are chasing, few are catching. Most leaders report they are adopting AI assistants. Only a small minority have anything genuinely running day to day.
Two reasons, and neither is technical. First, the step from a working trial to a running system does not add up on paper, so it stays a trial. Second, proving afterwards what it did costs money: every action the assistant takes on its own has to be logged and defensible to an auditor.
The question almost nobody asks
While everyone argues about what AI can do, a team at Stanford did something obvious that almost nobody does. They asked the people.
1,500 workers across 104 occupations. 844 individual tasks. For each one: would you hand this to a machine? Not an opinion poll, these are people who do the work themselves, every day. And they were explicitly asked to think about job loss and about the work becoming less enjoyable before answering.
Then they compared those answers against what experts consider technically feasible. That produces a simple grid: tasks people want to hand over and machines can do; tasks machines can do that nobody wants to give up; tasks people would happily hand over that machines cannot manage yet; and the rest.
Then they checked what young AI companies are actually building. The result: 41 per cent of those efforts aim at tasks that either nobody wants to hand over or nobody particularly cares about.
Two in five funded attempts are building past a question nobody asked.
The surprise is in the answers
You expect the workers to be the obstacle here. They are not.
For 46 per cent of tasks, the people doing them say: yes, please, take it. And the most common reason is not laziness. 69 per cent say: freeing up time for work that is worth more.
So why does it fail when it fails? By some distance the most common reason for resistance is a lack of trust, 45 per cent. Fear for the job comes second, at 23 per cent.
That is good news, because it changes what you have to fix. You cannot talk missing trust away. You build it away. Visible limits. A person at the right point. And the ability to look afterwards at what happened and why. That is the same construction I wrote about last week. The separation that makes an assistant safe is also what makes it trusted enough that people actually use it.
What you can buy, and what you cannot
The engine has become a catalogue item. It gets better and cheaper every quarter, and that is a good thing. That decision will soon stop being a big one, whichever way you make it.
What is not in the catalogue is the answer to five questions. Which task? Whose exceptions? Which access? Whose approval? And what happens when it gets it wrong?
A comparison from the shop floor: you can order a CNC machine. Whether it pays for itself depends on which part you make on it and where it sits in your process. That is in no catalogue, because it is in your building, and mostly in your people’s heads.
Which is exactly why it is not written down anywhere. If it were written down, your existing software would already handle it.
This holds for everything we build. An assistant on the phone, a chat that answers from your own documents, a helper in the inbox: the engine has become interchangeable in all three. What makes the difference is the question before it. Which call should reach the assistant at all? Which documents are the ones that matter? Which email may it answer itself, and which never?
The test for Monday
Before you buy anything, this costs you an afternoon. Find three people who actually do the work and ask them two questions:
- Which two tasks would you hand over tomorrow?
- Which one would you never hand over, and why?
The second answer is the more valuable one. It shows you where trust breaks, and a lack of trust is the most common reason people do not come along.
No supplier will do that for you. Not even the biggest one. That is what it sends its engineers for.