We’ve all done it.
We’ve all seen those flashes of brilliance, the moment you sit back and a bit of a cold sweat washes over you, because the AI has just produced something that is genuinely world-class in your industry. An advice letter, a set of board-pack commentary, a proposal… something you would have struggled to produce with ten times the effort it took the machine.
And for a moment there’s dread. Maybe, just maybe, this thing is going to take your job and destroy your industry.
You’re quickly disabused of that notion.
Because you run it a second time, and a third, and a fourth, and a fifth, thinking you’ll never have to do this work again, and hopefully it’ll take a little while for your clients to notice so they still keep paying you. That’s when the mirage comes crashing to the ground. Every output is different. Some are okay. Some are absolute disasters. Mistakes start creeping in, and you can never seem to get it to produce, consistently, the way you want.
And you start to realise you’re going to have to put a lot more effort into shepherding it. Being the conductor, as we call it, rather than the audience.
Every output is a dart
Unconstrained, AI throws all over the place. A bullseye one minute, the wall of the pub the next.
Think of every output as a dart.
Say you’ve asked AI for a client advice letter. The bullseye is the world-class version, ten out of ten, couldn’t do better yourself. It’s spotted the point you’d have missed, it’s picked up the risk the client hasn’t mentioned, and it’s written so the client actually understands it. Around the bullseye you’ve got the trebles, the doubles and the normal scores. Then the bit just off the board. And then the whole back wall of the pub.
Now picture who’s throwing, because this is the bit that matters… the most talented player you’ve ever seen, who turns up drunk half the time. On a good night it’s bullseye after bullseye, and on a bad night there are darts in the wall, the bar and somebody’s pint.
You wouldn’t hire that person. Not because they lack talent, but because you can’t rely on them, and sooner or later they’d do something in front of a client that you can’t take back.
That’s AI with nothing around it.
The ripples you don’t see
One small mistake doesn’t stay small. It reaches the client, then your team, then your plans for AI.
Here’s what I see in firm after firm. Somebody sets up a prompt or an assistant, the first week is magic, everyone’s impressed, and then nobody watches it because as far as anyone’s concerned it’s “done”, so nobody checks the hundredth output the way they checked the first.
So the small misses start, a client’s name spelt wrong here, last year’s tax threshold there, a confident sentence about the law that isn’t quite true, a tone you’d never use with someone who pays you for your judgement.
Each one is a pebble in a pond. A client notices, and trusts you a little less. Your team notices, and quietly goes back to doing it by hand “because the AI gets it wrong”. The tool is still on the invoice every month, and the transformation you bought it for never happens.
You haven’t just had a bad output. You’ve killed your own AI revolution before it’s even begun.
And in a firm where your reputation is the product, one mistake in front of a client costs more than every hour the AI ever saved you.
It’s not a capability problem
Every time the output disappoints, the instinct is to reach for a better thrower, a cleverer prompt, a newer model, whatever came out on Tuesday that everyone says changes everything.
But the thrower can already hit the bullseye. You’ve seen it do it.
When you run a firm, it’s not a capability problem. It’s a reliability problem. You aren’t doing PhD-level mathematics. You aren’t discovering unknown physics. You don’t need all of that unconstrained intelligence pointed at your engagement letters, your month-end commentary and your client advice. What you need is a process that runs 99.99% of the time to the satisfaction of you, your clients, and your regulator, whether that’s the Solicitors Regulation Authority (SRA), the Institute of Chartered Accountants in England and Wales (ICAEW) or whoever sits over your profession.
That’s what a harness is for.
What a harness does to the darts
Each part of the harness takes away a slice of the bad outcomes. First the wall, then the edge of the board, until every throw scores.
The harness is everything you put around the AI to control what comes out. It doesn’t make the AI any cleverer. It does something more useful. It shrinks the range of possible outcomes.
Take away the wall of the pub and every dart at least lands on the board, and it might not always be a bullseye, but it’s on the board. Control it a bit more and every throw scores. A bit more again and it’s a double, a treble or a bullseye.
That’s something you can put your firm’s name to.
There are five parts. None of them needs a developer.
1. The template
Most of a client document is the same every time. Lock that part down and only let the AI write the rest.
Look at an engagement letter. The scope wording, the fees section, the terms, the complaints procedure, the sign-off… it’s the same every single time unless the client asks for something different. So why are you letting the AI vibe it?
Hard-code it. Then the instruction to the AI is simple. Use all of your creative power, but only on these parts.
Take the last three versions of whatever your firm produces most, an advice letter, a month-end pack, a coaching session summary, and highlight everything that was identical, or should have been. That’s your template. What’s left is where the AI earns its keep.
2. The rules
Your best people carry hundreds of unwritten rules in their heads. Figures only ever come from the client’s books, never from memory. Never state a filing deadline without checking it against this year’s date. Never give a view on tax we haven’t been engaged to give. The AI knows none of them, so it guesses, and every guess is a dart that might hit the wall.
Write them down in a plain document your AI reads every time it does the job. You write it once, and from then on it works to your standard instead of the internet’s.
3. The reviewer
A second AI checks the draft against your standards. If it fails, it goes back. Only a clean draft reaches a person to sign off.
This is the one almost everyone misses, and it’s the one that changes the most.
Before anything goes to a client, a second AI reviews the draft, and its only job is to break it. Not improve it. Not rewrite it. Fail it. The technical name is an adversarial review, which is a grand name for something very simple… a sceptic with a checklist.
You give it one checklist for each standard you care about:
- Brand voice: does this sound like us, and would the client understand every word?
- Firm values: does anything here go against how we treat clients?
- Facts: can every name, figure and date be traced back to the client file?
- Regulatory compliance: could anything read as advice we haven’t signed off, a guarantee, or a claim we can’t evidence?
If it fails, the draft goes back with the notes and gets fixed. A person still signs it off, but now they’re reading something that’s already been checked four ways, not catching basic errors at eleven o’clock at night.
That’s how you keep the drunk half of your darts player away from the client.
4. The workflow
A template, rules and a reviewer only work if they happen in the same order every time. Gather the inputs, draft from the template, review it, fix what was flagged, and then a person signs it off. Write those steps down, and the AI does the steps. It doesn’t decide them. You do.
That’s what being the conductor actually means.
5. Watching the darts
Left alone, AI quietly gets worse. Watched, every mistake you catch makes the harness tighter.
The first output isn’t the test. The hundredth is.
Put a regular slot in the diary where somebody looks at a sample of what the AI has produced. When you find a mistake, don’t just fix that letter. Ask which part of the harness let it through, and fix that. Wrong standard wording, update the template. A guess it shouldn’t have made, add a rule. Slipped past review, add a line to the checklist.
Every mistake you catch tightens the harness for good. That’s the difference between AI that quietly gets worse and AI that gets more reliable the longer you use it.
Don’t wait for the next model
There will always be a next model, twice as good and twice as disruptive as the last one, and when it comes you’ll plug it in like everybody else.
But a firm with a harness plugs it in and every dart still lands on the board, thrown by a better player. A firm without one gets a stronger arm and the same wall.
Your harness is not making your AI more capable. Your harness is controlling your AI so that it can be more reliable.
The five parts of a harness. Build them in this order.
So pick the one document your firm hands to AI more than any other, and this week:
- Template it. Hard-code everything that’s the same every time.
- Write down the five rules your best person would correct in a junior’s draft.
- Give a reviewer a checklist for voice, values, facts and compliance.
- Write the workflow as numbered steps.
- Put the first review of the darts in the diary.
Stop polishing the prompt. Start building the world your AI works in.
This is the game. Build your harness.
Written by