Your people already have AI writing up their conversations.

We built Fieldnoter, and this page is written from that position. It is one iPhone app that records a conversation, tells the speakers apart and drafts a summary, all of it running on the device. There is no account, no dashboard, and no server that could receive the audio.

That is a narrow product, and it answers a wide problem. The conversations that give organizations the most trouble are the ones that are not meetings: a home visit, an intake at the counter, an inspection, a job interview, a client conversation on location. No calendar item, no tenant, no bot to block. The write-up lands on the employee's own plate, and the quickest route from there is a public AI tool.

Below is what the research actually shows, where the honest limits of that research sit, and where we are the wrong tool. If you are the one who has to sign this off, the two sections you want are the one on what we do not do and the one on how your IT team can check it themselves.

Dutch supervision sees the same pattern in what gets reported to it. In December 2025 the Dutch Data Protection Authority told Het Financieele Dagblad it had received dozens of notifications that year of incidents in which sensitive information was shared through public AI tools.

The clearest Dutch case is the municipality of Eindhoven. A sample study covering 23 September to 23 October 2025 found that staff had uploaded thousands of files containing personal data to public AI websites, among them case files under the Youth Act with health information on minors, and CVs of job applicants. The breach was reported to the supervisor on 23 October 2025 and made public that December. The full scope can no longer be established, because the data are kept at the AI service for at most thirty days.

And the specific act we are talking about has a record too. In April 2023 Samsung locked down generative AI after three incidents in roughly twenty days. In one of them an employee fed an internal meeting into ChatGPT to have it write the minutes. That is one case and not a statistic, and we present it as one. It is also exactly the job this app does.

In the same study, contravening behavior is most common among employees who say their organization has banned generative AI (67 percent) or has a guideline for its use (56 percent), compared with organizations that have neither (33 percent) and employees who do not know whether a policy exists (38 percent). The researchers draw the conclusion themselves: outright bans may be ineffective, and simply having a policy does not guarantee compliance.

This is a correlation and we claim no cause. There is a plain measurement artifact inside it: someone working under a ban knows they are breaking a rule and will report it, while someone in an organization without any policy has no rule to break and reports nothing. Part of that gap is reporting behavior rather than actual conduct. The group with a ban is also small, six percent of the sample, so the subgroup is less precise than the headline figure. We say this ourselves because a security officer will find it in a minute, and it is better found here.

The more useful number sits in the same figure. Only six percent of employees see a ban at all. For most organizations, enforcing one is barely the live question. The live question is this: where does an employee go when the write-up has to be finished today, and is that a place you would have picked yourself.

In March 2025 the Association of Netherlands Municipalities reported that Breda had completed a pilot with generative AI and that the approach would be scaled to Losser, Tilburg and Zeist with funding from the ministry of the interior. One of the named applications is having meeting reports written up.

Three weeks earlier the same association reported that Amsterdam had decided not to start a pilot with Microsoft Copilot because of privacy risks, after a privacy assessment concluded the service could not be used compliantly for the time being. The intended uses were named explicitly, and the first of them was taking minutes of meetings.

Two large Dutch municipalities, the same task, opposite outcomes. The difference sits in where the audio goes and who can be compelled to hand it over. The quality of the transcription plays no part in it.

That assessment also moves. In September 2025 SURF put two of the four high risks around Microsoft 365 Copilot back to orange and advised institutions to weigh the risk per type of use. So we do not claim anywhere that a cloud service is forbidden. Our claim is narrower: for this class of conversation an organization tends not to want the residual risk either, and an app with no server takes the weighing off the table entirely.

Two controlled studies compared speaking with typing and found no time difference. In a controlled study of ten physicians, documentation time was comparable between the two methods, while the dictated notes were longer (320.6 against 180.8 words), used more unique words (170.9 against 120.4) and scored higher on quality (7.7 against 6.6). A randomized experiment with sixty participants found no significant difference in task duration and no loss of quality either.

The best measured practice study is much larger. Across five American academic health systems and 8,581 clinicians, using an automatic scribe was associated with 16.0 fewer minutes of documentation time per eight scheduled patient hours (95 percent confidence interval 13.7 to 18.3). Record work outside working hours did not change significantly, so the evening did not disappear.

The only randomized trial of these tools is more sobering still. Of the two systems tested, one cut time in the note by 9.5 percent and the other produced no significant reduction at all (minus 1.7 percent, p = 0.66). What did improve significantly for both was how the work felt: the Mini-Z burnout score rose by 2.76 points (p < 0.001), work exhaustion fell, and perceived task load fell.

So our claim is the modest one. A transcript with speakers and timestamps makes the record more complete than a note reconstructed from memory afterwards, and it takes the typing out of the evening. How many minutes that is for your work, nobody has measured. We are not going to invent it.

What can be compared without any assumption at all is the price.

Client conversations at investment firms

These fall under the recording obligation in MiFID II. The firm has to keep the recording centrally, retrieve it years later and produce it to the supervisor on request. A recording on a device the firm cannot read works against exactly that. We are the wrong tool here and we advise against it.

Records under the Open Government Act and the Archives Act

A transcript that exists only on one iPhone is a problem the moment an information request arrives, and we offer no archiving path. Where the write-up becomes part of the official record, it belongs in the system your organization set up for that.

Fieldnoter is a free download on the App Store and can be deployed through Apple Business Manager and MDM. The full version is a one-time in-app purchase of 14,99 euro, with no subscription. Today that means the user buys it and expenses it. For larger numbers we quote a price on request.

There is no volume purchase, no licence model and no procurement route, because we have not built one. Apple does not allow in-app purchases to be bought in volume, and on a device with a Managed Apple Account an in-app purchase may not be possible at all. That is worth checking on one of your own devices before you send people to the App Store.

A team of ten or more, a question out of a security review, or a requirement we do not meet yet: write to hello@fieldnoter.app. We answer ourselves, and we will tell you when the answer is no.