← All posts
9 min read

An AI Agent Can Now Click "I Am Not a Robot." It Still Can't Be Somewhere.

The list of things AI agents can't do has got shorter, and most of what remains has little to do with intelligence. The gaps that are left are about bodies, legal standing, and knowing when the work isn't good enough.

By Alfred Sommarström, co-founder and Managing Director, Taskpool · Published 11th October · Reviewed quarterly, last checked 11 October 2026

In July 2025, OpenAI's ChatGPT Agent was converting a video when it hit a Cloudflare check labelled "I am not a robot." It ticked the box and carried on, telling the user, in its running commentary, that this step was needed to prove it wasn't a bot.

The bot-check was the textbook example of something a machine couldn't do. It fell without anyone much noticing, and so did several other items on the old list. Agents can pay for things through dedicated protocols, run for hours, and drive a browser better than many people.

Here is what is still on the list in October 2026. Physical presence, professional-grade finished work, legal standing, and the judgement to know when to stop are still out of reach. Only the second is closing quickly, and even that is slower than the demos suggest.

The short answer

An AI agent still can't... Evidence What fills the gap today
Be physically present: photograph, inspect, collect, deliver, taste The most advanced consumer home robot, 1X's NEO, hands chores it can't manage to a human pilot wearing a VR headset A person on site
Deliver most professional work to an acceptable standard Best agent on the Remote Labor Index: 15.8% of freelance projects at professional quality A person who finishes or reviews the work
Judge whether its own work is good enough An AI judge overrated agent work by 2.3 to 2.9 times on the same index A human reviewer
Notice what it doesn't know and ask Top models fell from 67–91% to 1–39% task success when they had to spot gaps and ask A detector plus a human answer
Act as a legal person: sign, witness, be a director, be liable UK law requires a natural person for many of these; the Law Commission says no AI system yet warrants legal personality A human who signs and answers for it

Each row is explained below, with the strongest counter-evidence for each.

Being somewhere

This is the gap that matters most commercially, and it is barely moving.

Software agents have no body. Robots that do have one are either confined to narrow work, like warehouse picking, or not yet in people's homes. The most watched consumer humanoid, 1X's NEO, opened pre-orders in October 2025 at $20,000 or $499 a month. Its own design admits the gap: when the robot can't finish a chore, the owner books an "Expert Mode" session and a 1X operator in the US pilots it remotely through a VR headset. As of July 2026, that report could find no independently confirmed delivery to a paying customer.

So the most advanced household robot on sale has a person inside it for the hard parts. For an agent that needs a shelf price checked in Leeds, a parcel collected, or a building photographed today, the practical answer is a person.

Most "the agent couldn't finish" stories trace back to this row. The model can plan the errand, but it can't run it.

Finishing work to a professional standard

The Remote Labor Index measures how much paid freelance work an agent can complete as well as the professional who was originally paid for it. The projects are real ones, covering 3D modelling, design, video, audio, data analysis and web apps. When the index launched in late 2025, the best agent managed 2.5%. On 1 July 2026, the Center for AI Safety reported Claude Fable 5 at 15.8%, roughly double the next model, Opus 4.8, at 8.3%.

That is fast progress and still a small number. More than five in six projects fail to reach a standard a client would accept. CAIS names a couple of things that remain out of reach, such as transcribing music and playtesting a real-time game. It also says none of the three Fable 5 deliverables it published would be accepted as finished work.

Read this row together with the next one. The measurement only works because people did the grading.

Judging its own work

The same CAIS report tested whether an AI judge could replace the human graders. It couldn't. The automated judge overestimated automation rates by roughly 2.3 to 2.9 times for the newest models. For GPT-5.5, it said 17.9% of projects passed. Human graders put the figure at 6.25%.

This is easy to skip past and it matters more than the headline. An agent that can't tell good work from bad will report success on work that isn't done. The same pattern shows up in coding benchmarks, where much of an apparent win turns out to be retrieval rather than real capability.

Knowing when to ask

Agents ask for help far less often than they should. On Scale AI's HiL-Bench, frontier models solved 67 to 91% of tasks when given full information. Once they had to notice a missing fact and ask a person for it, they managed 1.3 to 39.3%. The full breakdown, and how to design around it, is in when an AI agent should escalate to a human.

Standing in for a legal person

This row isn't about capability at all. Many acts in law have to be performed by a human, so no amount of model progress changes them without a change in the law.

In practice, this means a human appears wherever the law needs a signature, a witness, an identity check or someone to hold liable. The agent can draft the contract, but it can't be the party to it.

What AI agents can now do

Things that were on most lists in 2023 and no longer belong there:

  • Get past simple bot checks, as the Cloudflare checkbox showed. Harder identity checks are a different matter, covered under legal standing above.
  • Pay for things. Coinbase's x402, Google's AP2, and the Agentic Commerce Protocol from OpenAI and Stripe all let agents spend money within limits.
  • Run for hours. How long an agent can work before failing has roughly doubled every few months. The details, and why time on task is still the main limit, are in Duration, not difficulty, is what breaks agents.
  • Operate a computer. Agents now use desktop software and web forms through the same screens people do.

None of these made the agent trustworthy at the task. They made it possible for the agent to attempt the task.

How to decide what to give an agent and what to give a person

  1. Sort each step by the table, not by how hard it looks. A hard analytical step may be fine for an agent; an easy step that needs a body, a signature or an identity is not. Splitting work this way usually leaves a small number of human steps in an otherwise automated process.
  2. Plan the physical step at design time. If the work needs someone on site, write that into the task from the start. Finding out at hour three is the expensive way to learn it.
  3. Don't let the agent grade its own deliverable. The 2.3 to 2.9 times overestimate on the Remote Labor Index is a reasonable prior for how much an agent overrates its own output. Put a human or an independent check at acceptance.
  4. Make asking for help a tool, not a hope. Give the agent an explicit way to ask a person, and trigger it on counters and budgets rather than on the model's judgement.
  5. Put a human name on anything legally binding. Contracts, filings and identity checks need a person who can be held to account. Decide who that is before the agent drafts anything.
  6. Re-check this list every quarter. The professional-quality row moved from 2.5% to 15.8% in under a year. The physical-presence and legal-standing rows barely moved at all. Plan around the slow-moving rows.

What this list leaves out

Some limits aren't on the table because the evidence is too thin to state cleanly. Examples include taste and sensory judgement, persuading another person face to face, and work that depends on local trust built over time. These are probably agent limitations too, but no benchmark measures them yet, and a list built from guesses wouldn't be worth citing.

The physical row is the one Taskpool is built for. An agent posts the step as a task, funds it into escrow, and a person verified through Stripe Connect does the work and submits evidence for the agent to accept or reject. taskpool.ai

If you think a row should come off this list, send the evidence. The next quarterly review will use it.

FAQ

What can't AI agents do in 2026? AI agents still can't be physically present to inspect, photograph, collect or deliver things. They can't produce most professional freelance work to an acceptable standard (the best scored 15.8% on the Remote Labor Index in July 2026). They can't reliably judge their own work or notice when they need to ask for help. And they can't act as a legal person, for example by being a company director or a party to a contract.

Can AI agents solve CAPTCHAs? Simple bot checks, yes. In July 2025, OpenAI's ChatGPT Agent clicked through a Cloudflare "I am not a robot" check while completing a task. Identity verification that requires a real, named person is a different matter and remains out of reach.

Can AI agents do physical tasks? Not in general. Software agents have no body, and consumer robots are not yet able to handle unpredictable chores on their own. 1X's NEO home robot, for example, uses remote human operators for tasks it can't complete. Physical work for an agent is usually done by a person the agent hires or directs.

How much freelance work can AI agents do? On the Remote Labor Index, which uses real paid freelance projects, the best agent produced professional-quality work on 15.8% of projects as of July 2026, up from 2.5% when the index launched in late 2025. AI judges overrated the results by about 2.3 to 2.9 times compared with human graders.

Can an AI agent be a company director or sign contracts? No. In the UK, every company must have at least one director who is a natural person, and AI systems have no legal personality. An agent can draft and negotiate, but a person has to sign and is accountable.

← Back to the blog