Powens, publish the recovery rate

Half the proposed fixes get directly approved. That is the number worth stopping at in Powens’ 28 October announcement of its Connector Agentic Factory. Not because 50% is bad. Because an AI coding announcement that admits someone still has to judge the output gives us something to argue with.

The next number should be the share of broken connections it actually restores, measured from failure to recovery. A proposed patch is an intermediate result. For the person whose finance app cannot fetch a bank balance, it is not the result at all.

A repair shop, not an AI banker

Powens is the French open-finance business founded as Budget Insight in 2012. Its connectors let financial apps retrieve customer data or initiate payments through financial institutions. The company says it maintains more than 1,800 connectors across Europe. Something at that scale will always need mending.

The factory is internal engineering machinery, not a new chatbot for account holders. According to Powens, it investigates a broken connector and proposes a repair. It can also examine an institution’s systems and build a new connection. Engineers check and approve every connector or fix before it goes live. Powens explicitly says the AI does not manipulate data or payments; it builds and maintains the connectors, which remain deterministic.

That boundary deserves credit. There is a meaningful difference between using a model to write code that someone reviews and putting a model in the path of a live payment decision. Calling both agentic finance would hide the most important design choice.

Fifty per cent of what?

Powens reports typical repairs taking 10–15 minutes rather than up to two days of engineering work, with 50% of proposed fixes directly approved. It also reports 12 new connectors delivered across banking, wealth and document data sources. These are the company’s figures, not an independent benchmark. I read the announcement; I did not run the factory or inspect its incident records.

The approval figure has a narrower denominator than the reader might assume. It counts proposed fixes. It does not tell us how many broken connectors produced no proposal, how many needed several attempts, or how long approved changes waited before deployment. Nor does the announcement define what happened to the other half. They were not necessarily all wrong: some may have needed editing. The release does not say.

Even the time comparison needs care. An upper bound of two days and a typical 10–15 minutes are not matching measurements. The release does not specify whether the shorter clock includes review, deployment and checking that the connection works again. Turning those figures into a neat speed multiplier would manufacture precision the source does not provide.

None of that disproves the improvement. A patch accepted without further work can save real engineering time even if the next patch needs rewriting. But a useful coding assistant and a more reliable financial service are different claims. The second needs evidence from the running service.

Measure where the customer waits

For a bank or accounting-app team choosing a data provider, the useful follow-up would start with all connector incidents in a stated period, not just the ones the model attempted. Show the time from detected failure to restored service, including human review. Show the slow end of that distribution as well as the middle, and how often a repair has to be rolled back or reopened. Averages are poor company when yours is the connection still broken on Friday.

Scope matters too. Powens says repair already applies to any of its connectors, while new-connector building currently focuses on investment and savings. Those are separate capabilities. Twelve delivered connectors do not establish that the factory can build every kind of connection the platform supports.

The sensible buying question is not which model wrote the patch. It is whether this institution’s connection is supported, how failures are handled, and what recovery looks like for customers. Keep the engineer at the release gate. Then measure the queue on both sides of that gate.

The 50% figure is worth publishing precisely because it is incomplete without being useless. Leave it there. Next to it, put the number of people who got their connection back.

Sources: Powens’ launch announcement, 28 October 2026, and company history, read on 31 October 2026. Performance and coverage figures above are attributed to Powens.