What three months in Lagos taught us about pro-bono software

The clinic’s generator ran out of diesel at 11:40 on our second morning. The lights went, the router went, and the three paralegals at the intake desk kept working, because the queue outside did not care about the router. One of them wrote notes on paper. One used her own phone’s data. The third, who had opened an AI Interview session fifteen minutes earlier, watched it fail to reconnect and said, without much heat, “This is why we don’t rely on it.”

That was the most useful sentence anyone said to us in three months.

From September to late November 2025, two of us from the Open Bar team sat with a legal-aid clinic in Lagos, Nigeria, as part of the residency we announced in the summer. The clinic handles tenancy disputes, police bail and detention matters, unpaid wages, and family matters, in that order of volume, with eleven paralegals, four supervising lawyers and a queue that starts forming before seven. They had been using Open Bar since early 2024, mostly Recall over their own precedent files and, since May, AI Interview for first intake. What follows is what we learned. It is a field note, not a study; the sample is one clinic and our own eyes.

Connectivity is a state, not a fact

Our product assumed a connection. Interrupted sessions in Interview were discarded on the assumption that the client would start again. In Geneva that assumption costs a minute. In Lagos it cost a client who had waited three hours and would not wait again.

We shipped, in week four, a change that holds an Interview session locally on the device for up to 72 hours and resumes it when a connection returns, with the structured intake brief generated from whatever was captured. The paralegals’ judgement on whether a partial brief is usable turned out to be better than any rule we could write, so we stopped trying to write one and showed them the gaps instead.

Recall queries now queue when offline and return when the connection does, with a timestamp. Nobody asked us for this. We watched three people retype the same query four times and did it.

The phone is shared, and so is the screen

Most intake happened on the paralegals’ own phones, not clinic devices. The phone was sometimes passed to the client to speak into. Sometimes two clients from the same household were interviewed together. Our interface assumed one user, one device, one client, one session.

We did not fully solve this, and we should be honest about it. We added a plain “hand the phone over” mode that locks the paralegal’s view and shows the client only the current question in their language. We added a two-party intake path for household matters. We did not resolve the underlying tension between per-client confidentiality and a device that belongs to nobody in particular, and we are testing approaches rather than claiming one.

Language is not a dropdown

Interview supports 38 languages. The clinic’s clients spoke English, Nigerian Pidgin, Yoruba, Igbo and Hausa, often more than one in the same conversation and often switching mid-sentence. Our language selector asked the paralegal to choose one at the start.

The honest finding is that our coverage of the languages spoken in that room was thin, and our handling of switching was worse. We improved mid-session switching for the languages we do support. We did not add languages during the residency; that is a longer piece of work, with native-speaker legal review, and we would rather do it properly than list a flag. The paralegals’ workaround, interviewing in English and Pidgin and translating for the brief themselves, was better than the tool for the time being, and we said so.

Paper is the delivery format

The intake brief, with its statute-of-limitations computation and its conflict check, was designed as a screen. The clinic’s supervising lawyers read on paper, in the evening, by whatever light there was. A brief that ran to four pages with a sidebar was printed as six, and the sidebar was lost.

We rebuilt the brief export as a two-page, single-column, black-and-white document with the limitation date and the next action at the top. It is less elegant than the screen version. Everyone at the clinic preferred it, and so, once we looked, did a public defender’s office in the US that had never mentioned the problem.

What we removed, what we admitted, and what we brought home

Three things came out of the product during the residency:

• A dashboard of weekly intake volumes that nobody opened.
• The confidence indicator on Recall results, which was read as a judgement of the paralegal rather than of the retrieval.
• An onboarding tour. The paralegals learnt the product from each other in an afternoon and found the tour patronising.

One thing we had to say out loud. Nigeria is not one of our four data regions. The clinic’s data is held in the EU region with the clinic’s informed consent, and we said so plainly in the consent language rather than burying it. Several clinic staff asked about it unprompted. We do not have a date for a West African region and did not pretend to.

And what came home with us: Open Bar now ships offline resumption, queued queries, the hand-over mode and the two-page brief as defaults for every clinic and public-defender account, not as Lagos-specific options. The problems were not Lagos problems. Lagos was simply where they were impossible to ignore.

The paralegal who said “this is why we don’t rely on it” used Interview for her last four intakes of our final day. We take that as a conditional pass, which is the right grade.