When knowing things is free

I hold a belief about where professional services are heading, and this piece sets out what I think follows from it. The belief is directional and its timing is uncertain, but I would put it this way: the value of knowing something is falling towards zero, and it is falling faster than most of us will be comfortable with.

For a growing share of questions, the best available answer is now “ask the AI”, and the share grows with every model release. Best-practice advice on a technical question in tax, law, engineering or design used to require finding someone who knew and paying them. Increasingly it requires neither. Whatever a competent professional could tell a client from general expertise, the client can get from a machine, cheaply, and the gap between the two is closing. Accountability and licensing survive this, but in a stripped-down form: the professional’s role migrates from checking the AI’s work to vouching for the AI’s ability to do the work. None of this happens overnight, but the direction is not in doubt.

Two strategies that will not work

If that belief is right, two of the strategies that small and mid-tier firms reach for most often ultimately will not work.

The first is to compete on the quality of the service. This is the traditional position, and it was sound while expertise was scarce. It stops being sound when best practice is a commodity. Being better than your competitors at something the client can get for nothing is not an advantage worth much.

The second is to compete on being the best at implementing AI. This is the one I see firms grab for most often, and it is understandable: if the technology is the threat, master the technology first. The problem is that implementation skill is knowledge like any other, and it goes the same way. If knowing how to do trust accounting is going to be free, then knowing how to configure an AI to do trust accounting is going to be free too, and not much later. Being six months ahead of a competitor on tooling is an advantage that expires with the next model release. You will still need to implement AI well, but it will not distinguish you.

What survives is the knowledge the AI does not have

What survives is the knowledge the AI does not have. By definition, that is knowledge that exists only inside your firm. There is more of it than most firms realise, and it comes in three kinds.

The first is how you do what you do. I mean the method as it is actually practised rather than as it appears in a proposal: the judgement calls, the sequence, the questions you ask a new client in the first hour, the work you decline, how conservative you are on a particular class of advice, what you tell a client when the honest answer is “don’t”. Much of this is implicit. It lives in the heads of your senior people and has never been written down because nobody needed it written down.

The second is what you know about your clients. Their history, their structure, their appetite for risk, what went wrong last time, what they said they wanted and what they actually meant. A generic model knows none of this, and no competitor can buy it.

The third is who you are: your purpose, your values, what you stand for. This functions as branding, and branding matters more, not less, when the services being compared are otherwise interchangeable. But it does something more useful than branding. Given a firm’s actual judgement rules, an AI stops producing generic output and starts producing output that is recognisably yours: answers given the way your firm would give them.

Why this is a knowledge management problem

Capturing these three kinds of knowledge, keeping them current and making them available to whatever AI you use is the whole task. It is a knowledge management task, and I use that term deliberately, knowing that knowledge management has a poor record. Firms tried it in the 2000s and mostly failed. They failed because the administrative overhead of capturing and maintaining knowledge was unsustainable; the knowledge went stale because keeping it fresh was too onerous; and nobody used it because it was stale. All three are labour problems, and AI is good at that kind of labour. It can draft the capture from the material a firm already generates, flag what has gone stale, and use what has been captured without anyone needing to go looking for it. The overhead that sank knowledge management the first time is now the part that can be delegated.

Why timing matters

This is also why timing matters. Most advantages a firm can build with AI decay as the models improve, because the improvement is available to everyone. This one compounds. A better model extracts more value from the same body of captured knowledge, so a firm that has done the work is better placed with every release, and a firm that has not falls further behind with every release. At some point, probably sooner than we would like, firms will be competing on entirely new terms, and those still playing by the old rules will lose relevance quickly. I know of no firm that has yet done this well.

How it should be built

How it is built matters as much as whether it is built, and two principles apply. The first is separation: knowledge about how the firm works must be kept apart from knowledge about individual clients, so that each can be brought to the AI only when needed, the resulting artefacts can be managed, and conversational traces can be deleted. Done this way, the approach sits comfortably with confidentiality obligations. The second is portability: the knowledge must live in a form that any model can consume and no vendor owns. The models will keep changing. A firm’s knowledge base should not have to change with them, and it should never depend on whichever platform happened to be current when the work was done.

What this amounts to

If I had one sentence for a partner group, it would be this: stop asking how to use AI better than your competitors, and start asking what you know that the AI does not. Then write it down, in a form you own.