What a Court Cannot Automate

A court has neither an army nor a treasury. It cannot enforce its own orders or fund its own existence. Its authority rests on one thing alone: that people believe in it.

That sentence — a near-quotation of Hamilton, writing in 1788 that the judiciary holds "neither the purse nor the sword" — was the premise beneath five briefings I was asked to prepare for the senior judiciary of the Federal Republic of Nigeria: for the Supreme Court, the Court of Appeal and the Federal High Court. Two of the five were about artificial intelligence. The other three were about procedural fairness, the culture of court administration, and ethics and public trust. On the face of it, only two were about technology.

But preparing them together, one thread pulled through all five — and it turned out to be the same question that now sits in front of every board, every executive team and every leader I advise. Not can we automate this? The honest answer to that is almost always yes. The real question is narrower and harder: how far may the human step back before the thing we were trying to protect quietly disappears?

Start with what a court actually is. It commands no soldiers and controls no budget of its own. When it rules, it relies on everyone else — the executive, the police, the parties, the public — choosing to comply. That compliance is not automatic; it is granted, and it is granted because the institution is trusted. Strip the trust away and a judgment becomes a piece of paper.

This is not sentiment. In the ethics session, evidence from across the Nigerian judiciary itself made the point sharply: asked how transparent the system is, fewer than one in five respondents said "completely" — and among those with more than thirty years at the Bar, not a single one did. Confidence, once spent, is spent at the counter, one interaction at a time. Everything that follows is about protecting the source of that confidence.

Here is the finding that should unsettle anyone who thinks justice is only about getting the answer right: people judge a court less on whether they won, and more on whether the process felt fair. Perceived fairness of process is the single strongest predictor of whether people accept a decision and comply with it — stronger, in the research, than the perceived fairness of the outcome itself. Litigants who felt heard and respected were more than fifty per cent more likely to comply with an order, regardless of how the case went.

Fairness, in other words, is not a feeling the court can afford to leave to chance. It is observable behaviour, and it rests on four things a participant can actually see: the chance to voice their side and have it recorded; a neutral decision-maker who explains their reasoning; respectful, dignified treatment; and an authority that visibly acts in good faith.

A litigant who loses but felt heard trusts the court more than one who wins but felt ignored.

That is why the administration around a case matters as much as the ruling inside it. A citizen's reading of a court is shaped long before a word is spoken in the hearing — by the form they could not decode, the summons that never came, the corridor they waited in. In the service-culture session I drew on England and Wales's own reform programme, which is instructive precisely because it is honest about its mistakes.

Over nine years, His Majesty's Courts & Tribunals Service spent £1.23 billion modernising the system: fourteen digital services, millions of cases filed online, probate processed four times faster. Real gains. But the programme's hardest-won lesson was a warning, not a boast — that efficiency is not justice. Online-only services quietly excluded citizens without connectivity or confidence. Systems designed without the profession introduced friction that slowed the very courts they were meant to speed up. The remedy each time was the same: keep a human route, and design with the people who use it.

Hold that thought, because it is the hinge of the whole essay. Every one of these first three briefings — fairness, service, ethics — was really about the same thing: the parts of justice that are irreducibly human. Being heard. Being treated with dignity. Being able to trust. None of it can be delegated to a process, let alone to a machine. Which is exactly why the arrival of AI in the courtroom is not a technology question. It is a question about what we are willing to protect.

The two AI briefings opened on a single, deliberately unambiguous line, and I want to give it to you exactly as their Lordships heard it: artificial intelligence may assist the judge; it must never become the judge.

That is not my opinion; it is the settled position of the UK judiciary's own AI guidance, and it rests on three reasons any leader will recognise in their own domain. Independence: no system and no vendor may stand between the decision-maker and the decision. Impartiality: a model trained on another country's history carries that history into your courtroom — ask an ungrounded model a Nigerian question and it will often answer in American doctrine. Integrity: the judge remains personally accountable for every word issued in their name, including any a machine drafted.

Demystified, the technology makes its own limits obvious. A large language model does not know the law; it predicts the next most probable word. That has three consequences a court — or a compliance function — has to respect. It is strong at compression and weak at discovery: excellent at summarising what is already in front of it, unreliable at finding anything it cannot be made to verify. Hallucination is not a bug but the method — a system that completes patterns will invent a plausible citation as readily as recall a real one. And so trust can never live in the machine's fluency; it lives in the citation, the named human who checks it, and the contract that binds the vendor.

Which is why there is proper, proven work for it to do — all of it on the administration of a case, never its merits:

Notice the safeguard hiding in the Tanzania figure: the machine drafts, and a human verifies before it counts. That is human-in-the-loop made concrete — and it is the difference between a tool and a risk.

The seductive case is prediction — feed a model years of records and let it forecast the outcome. South Africa's eLAA study is the most-cited African evidence on exactly this, and it deserves to be read carefully, because its headline is a trap. The best model forecast a binary outcome — prison or acquittal — about seventy-five per cent of the time. Say that plainly: a system right three times in four is wrong in one matter in four. That might help a lawyer weigh whether to run a case. It comes nowhere near a standard of proof, and the researchers say so themselves.

The deeper problem is what the number hides. That model was trained on a single firm's caseload — it had learned one chambers' files, not a nation's law. Add the risks that come with any imported system — a default towards foreign precedent that falls hardest on the marginalised; instructions hidden in a filing as invisible white-on-white text that the machine obeys and the judge never sees — and you arrive at the accountability gap that should worry every leader: a human answering personally for the output of a system they did not build and cannot inspect.

The reassuring news, which formed the second AI briefing, is that this is not uncharted territory. When you place the serious jurisdictions side by side — Estonia, Singapore, China, Germany, India, Brazil, France, the European Union, the United States — they reach the same boundary by entirely different routes. Twenty-one jurisdictions on the map; one line.

Everything reduces to one organising idea, and it is the sentence I would lift out of the entire engagement for any leader deploying AI in any high-stakes setting. Every framework is really a decision about where on a single line a given task may sit:

Human in the loop — the human reviews and signs off every output before it has any effect (Tanzania's transcripts). Human on the loop — the system acts, but a named human supervises and holds an absolute right to intervene or halt it (Estonia's non-binding rulings, which route to a judge the moment either party objects). Human out of the loop — the system decides, with no routine human check between it and the person affected. That last position, for a verdict, is ruled out everywhere.

Think of a cockpit. Hand-flying is human-in-the-loop. Autopilot engaged with the pilot watching, ready to seize control, is human-on-the-loop. An aircraft with no pilot and passengers aboard is human-out-of-the-loop — unthinkable, for precisely the reason a court will not automate a verdict. Singapore's Chief Justice put the same idea in five words a board could adopt tomorrow: technology augments judgement; it does not replace it. Different systems, different tools — but not one of these jurisdictions lets the machine give the verdict.

None of this argues for keeping AI out. It argues for a gate — a repeatable test applied before any tool is adopted, and long before an incident. The whole engagement collapses into three questions asked in order; fail any one and the tool stops at the door.

And whatever passes the gate should enter on four conditions — the same four I would write into any procurement of a system that acts on your behalf. They are far cheaper to insist on in a contract than to retrofit after an incident.

I closed both AI briefings on the same line, and it is the line I will leave you with, because it reaches well beyond any courtroom: automate the administration; preserve the judgement. The authority of a court is not a workflow — and neither, in the end, is the authority of a leader.

Strip the robes and the ritual away and the judiciary is simply the sharpest version of a test every organisation now faces. There is enormous, legitimate value in letting AI carry the administration — the listing, the drafting, the transcription, the triage. There is real peril in letting it drift, one convenient step at a time, into the judgement itself: the weighing of facts, the reading of a person, the decision someone must answer for. The four "human" briefings built the case for what must be protected. The two AI briefings showed how to protect it. The join between them is the whole point.

So the question to hold — in a court, in a boardroom, in your own week — is not whether the machine can do it. It is whether, after it has, a named human still has the last word, and can still stand behind the result. Keep that person in the loop, and AI becomes the most powerful assistant judgement has ever had. Let them step out of it, and you have automated away the one thing that made the institution worth trusting.

Automate the administration; preserve the judgement. The authority of a court — like the authority of a leader — is not a workflow.