Legal AI Should Know What It Does Not Know

On August 26, Flowt closed an undisclosed pre-seed round from Delta40 Fund I, Impacc and Argidius Foundation. The Nairobi-based startup, founded by Elana Laichena, uses AI to turn messy financial records into lender-ready data, and its first facility went to GreenBay, a Kenyan appliance refurbisher whose growth had been capped by inventory it could not afford to hold.

Strip away the funding figure and what is left is a specific kind of product: a system whose entire value lies in converting something illegible, a business's actual, disorderly financial life, into something a lender can act on with confidence. That is a hard problem and a useful one. A lender who cannot read a business's records cannot extend it credit, and a business that cannot borrow cannot hold the inventory it needs to grow. Flowt's bet is that AI can close that gap faster than an accountant can.

The same conversion happens in a law office

Legal work runs on an identical structure. A client's situation, a set of facts, a stack of correspondence, a half-remembered agreement, arrives messy. A lawyer's job is to convert that mess into something a court, a regulator or a counterparty can act on: an opinion, a pleading, a clause that will hold. The lawyer is doing, in a different register, exactly what Flowt does for GreenBay. Both are in the business of making the illegible legible enough for someone else to rely on.

That is precisely where the risk sits, and it is a risk that looks identical to the help. A financial intelligence platform that smooths over a gap in a business's records, rather than flagging it, produces a number that looks lender-ready and is not lender-safe. A legal research tool that resolves an ambiguous line of authority, rather than naming it as unresolved, produces a citation that looks brief-ready and is not brief-safe. In both cases the danger is not that the system gets something wrong. It is that the system presents an unresolved question as a resolved one, and the person downstream, the loan officer, the judge, the client, has no way to tell the difference from the output alone.

The objection worth taking seriously

The obvious objection is that hedging is not a feature, it is a failure to do the job. A lender does not want a memo full of qualifications, it wants a number it can lend against. A partner does not want a research note that says possibly six times, she wants an answer she can put in front of a client. Push a tool toward constant disclosure of uncertainty and you get something too timid to be useful, and the market will simply reward whichever competitor sounds more confident.

That objection holds for Flowt, up to a point, because Flowt is making its own commercial decision. It underwrites the facility, it carries the credit risk, and if its model is wrong the cost lands on Flowt and its investors. A legal AI tool does not carry that risk. The lawyer who signs the pleading carries it, personally, on a practising certificate that does not transfer to a vendor. The profession has already watched what happens when a filing cites a case that reads well and does not exist: the sanction falls on the person who signed, not the tool that suggested it. That asymmetry is the whole argument for restraint. A tool that hides its own uncertainty is not saving the lawyer time, it is quietly moving risk onto her without telling her it has done so.

LEX, HERDelight's legal AI associate for East African practice, is built around that same discipline. Where it can tell you a decision has been cited in later judgments, it says so, and it says plainly when that treatment is unverified, rather than manufacturing the confidence a busy lawyer might be tempted to accept. It does not resolve the ambiguity. It marks it, and leaves the judgment call to the person who holds the practising certificate.

The question worth asking a vendor now is not how accurate their tool claims to be. It is what the tool has been built to refuse to tell you. A system that always sounds certain is not more capable than one that sometimes says it does not know, it has simply moved the risk of that uncertainty onto you, silently, and left you to discover it the way the profession usually does: in front of a judge, after the fact.

Sources

Kenyan AI startup Flowt raises pre-seed funding round (Disrupt Africa, 2026-08-26)

SA AI startup Verascient raises $1.2m pre-seed funding to help companies build better teams (Disrupt Africa, 2026-08-25)

Nigeria’s Intron launches new version to bring natural African language mixing to Voice AI (Disrupt Africa, 2026-08-27)