The agent works. Almost nobody uses it.

Uber built an agent that books rides. Then it told reporters that almost no one uses it. The feature works. The demand is not there.

In the same week, Airbnb said its AI travel agent arrives in the first half of 2027, and Meta opened video chat with an avatar to US users. Three companies, three announcements, one shared assumption: that people want software to finish the job rather than start it.

We have spent the last year watching demos of "autonomous procurement" and "self-running sourcing desks". In every single one, a human still pressed send. The vendor called it a guardrail. We think it is the actual product.

The task is not the bottleneck

Ask an agent to draft a reply to an RFQ and it does a decent job. Ask it to send that reply and something in the room tightens. The draft is a draft. The sent version carries a price, a lead time, a commitment to a date, and a liability if the container does not sail.

That is not a limitation of the model. It is a property of the decision. Some actions can be taken back and some cannot, and the difference is what separates a useful agent from an expensive one.

Sort your own workflow this way and the picture gets clearer than any capability benchmark:

• Reversible, frequent — rewriting a listing title, tagging a supplier email, extracting line items from a PDF, translating a spec sheet for internal review. Agents are already good here and the cost of being wrong is a few minutes.
• Reversible, rare — drafting a quote, summarising a factory audit, assembling a compliance folder. Worth doing, worth reviewing, low blast radius.
• Irreversible, frequent — sending price confirmations, updating stock, filing customs paperwork. This is where most "agentic" pitches point, and where almost nobody deploys.
• Irreversible, rare — signing a contract, releasing payment, choosing a freight forwarder. Humans will keep these for a while, and should.

The pattern is not that agents are too dumb for the bottom two rows. It is that the bottom two rows have no undo button, and an undo button is cheaper to build than a guarantee.

What Uber's number actually says

The tempting read is that people do not trust AI yet, and trust will arrive with better models. We do not buy that. Trust in a booking agent is not a feeling about intelligence. It is a calculation about what happens when the driver does not show up and the fare was charged to a card the agent picked.

Airbnb's 2027 travel agent will face the same wall, and so will every sourcing agent sold to an export team. The feature that decides adoption is not the model's reasoning. It is the receipt, the confirmation, the "this action can be reversed for 30 minutes" line, and a phone number to call when it cannot.

One vendor we tested last quarter had built a genuinely good agent for chasing supplier confirmations. It drafted the follow-up, picked the right thread, and stopped one click short of sending. When we asked why, the product lead said: "Because the first time it emails the wrong buyer about the wrong price, we lose the account." That is not a missing feature. That is a company that has thought about the last mile.

Three questions we now ask before buying anything described as an agent

We have started putting these in the first call rather than the fifth, because they sort vendors faster than any demo:

1. What is the undo? Not "can we review it" — what happens after the action has been taken. If the answer is a support ticket, the action is not reversible, whatever the marketing says.
2. Who is liable when it is wrong? If the vendor's contract says the customer is responsible for all output, then the customer is the agent's real quality-control department, and should price the tool accordingly.
3. What is the audit trail? A log with a timestamp and the exact prompt is table stakes. A log that shows which supplier record, which price, and which previous message it read is the thing that lets a human sign off in ten seconds instead of ten minutes.

None of these are AI questions. All of them decide whether the AI gets used.

Where this leaves the export desk

The realistic 2027 version of an agentic sourcing desk is not a system that runs procurement. It is a system that removes the boring 60% — the extraction, the translation, the chasing, the comparison — and hands a human a decision with the evidence already assembled.

That is less exciting than the demos and considerably more useful. It also changes what you should measure. Stop asking whether the agent completed the task. Ask how many seconds it took a person to approve the result, and how often that approval was wrong.

The number that matters is not task completion. It is the cost of a signature.

---

We test agents on the same four tasks every quarter: a 40-SKU listing in three languages, a supplier negotiation email, an RFQ comparison sheet, and a freight quote reconciliation. If a tool cannot survive that brief with a human in the loop, it does not go in the library.