What AI still can't do: a working list from people who ship it

Public conversation about AI swings between two moods. On one side: it can do everything, and soon. On the other: it's all hype and nothing works.

The people who build AI products for a living tend to sit somewhere much more boring and much more useful. They know exactly where the tools shine, and exactly where they fall over — because that's where the support tickets come from.

We asked engineers, founders and operators in our community to keep a running list. This is it, as of mid-2026. It will change. That's the point of keeping it.

1. Know when it doesn't know

Models are far better than they were at saying "I'm not sure." But in production, they still regularly give confident answers to questions they shouldn't answer at all — especially at the edges of their knowledge, or when the question is slightly malformed.

The practical consequence: any system where a wrong answer is expensive still needs a human, or at least a hard check, in the loop. One engineer summed it up: "We don't trust the model's confidence. We trust our evals."

2. Own an outcome

AI can complete tasks. It can even chain tasks together into impressive workflows. What it can't do is be accountable. When a deal falls through, a patient is harmed or a launch goes badly, someone has to own it, explain it and fix the relationship. That's still a human job, and organisations are discovering that the accountability layer is where much of the real value always was.

3. Read the politics of a room

An agent can summarise a meeting perfectly and completely miss what happened in it. That the VP was quietly unhappy. That two teams are protecting territory. That "let's revisit next quarter" meant "no." Operators in our community say this is the single biggest reason fully automated workflows fail inside large organisations: the work was never only about the information.

4. Decide what matters

Give an AI system a clear goal and it's remarkably capable. Ask it to decide which goal is worth pursuing — which customer to prioritise, which feature to cut, which risk to accept — and the answers become generic. Prioritisation depends on context, values and trade-offs that aren't written down anywhere. Taste and judgement are still distinctly human advantages.

5. Stay consistent over long, messy work

Short tasks are largely solved. Long-running work — a multi-week project with changing requirements, partial information and shifting stakeholders — is where agents still drift. They lose the thread, forget earlier decisions or quietly optimise for the wrong thing. This is improving quickly, but anyone promising fully autonomous long-horizon work today is overselling.

6. Earn trust

People don't trust outputs. They trust people and institutions with a track record. When the stakes are high — medical, legal, financial, emotional — buyers want to know who stands behind a decision. Several founders told us their biggest sales unlock wasn't a better model, it was putting a named, qualified human in the loop.

7. Handle genuinely new situations

AI is very good at situations that resemble what it's seen before. It's much weaker when something is truly new: a novel crisis, an unprecedented market shift, a first-of-its-kind product. Humans aren't perfect here either, but we're better at reasoning from first principles when the patterns run out.

8. Care

This one sounds sentimental, but it came up again and again. A support agent can be polite. It can't care whether your day got better. A manager who notices someone is struggling, a founder who calls a churned customer to genuinely understand what went wrong, a mentor who stays invested for years — none of that is automatable, and people can tell the difference.

What this list is for

This isn't a comfort blanket. Several items on it will shrink over the next few years, and we'll update it when they do. It's a map — a way to decide where to invest your own skills, and where to design humans into the systems you build.

If you notice, the list has a pattern. What AI still can't do is mostly about judgement, accountability, relationships and meaning. Those were always the hardest and most valuable parts of work. AI didn't create that truth. It just made it obvious.

We can't out-compute the machines. We don't need to. We need to get very good at the parts that were always ours.

---

This list is maintained by members of TopAImate, an invite-only community of people building with AI. If you have something to add from your own work, request an invitation and bring it to the room.