Knowledge management in consulting firms, run by AI

Consulting sells accumulated judgment and then organises the work so that judgment evaporates at the end of every engagement. Only half of that loss is fixable by a system, and it is the operating half rather than the intellectual one.

There is an irony in the phrase knowledge management in consulting that most people inside the industry have noticed. The product is knowledge: what we have seen before, how it went, what we would do differently on the next one of these. And the unit of work is built to lose exactly that. An engagement runs for a fixed period, produces something written for the client's situation rather than for the next engagement, and then stops. The team moves onto other accounts and the reasoning leaves with them.

What survives is the material that had to be written for someone else: the final report, the model, the readout deck, filed in a folder named after a client. What does not survive is everything around it. Why that recommendation rather than the other one that nearly won. What the client pushed back on and what happened when they did. Which two workstreams turned out to be the same problem wearing different labels. What got promised in week three that never made it into the scope.

What goes missing between engagements

The first cost lands in the pipeline. A proposal for a new client of a familiar type should be the cheapest document a firm ever produces, because the firm has answered that question before. In practice it gets written close to from scratch, because the previous version is in somebody's sent items, the assumptions behind it were never written down, and the one person who remembers how the delivery went is billed to another client this week. The firm pays twice for the same thinking and calls the second payment business development.

The second cost lands mid-delivery, and it is sharper in a small firm than a large one. In a boutique or a solo practice the person who sold the work is usually the person delivering it, which means a sentence said in a pitch has a habit of becoming scope. Run several relationships at once, each with its own steering call, its own set of half-promises about what will be ready by when, and its own quiet list of things the client owes you before you can move, and the tracking of all of it lives in one head. It holds until two of the relationships get busy in the same week.

How the record keeps itself current

Mail arrives from a client, a subcontractor, a prospect or a partner. It gets read, filed against the engagement or the relationship it belongs to, and the record for that thread is updated: what the message resolves is closed, what it creates is opened, dates that moved are moved. The next step, usually a reply, is drafted and left for you to approve. Nothing goes out on its own. Over a few months you stop having engagement folders and start having engagement threads, which is a different object: it holds the promises and the reasons, not just the deliverables.

Two constraints to know before you imagine more than exists. The sweep that files and reconciles runs on a schedule, hourly at best, and it needs the machine to be awake, so this is not a live feed. And it is a single seat: one install, one owner, one mail identity. It cannot assign a task to a colleague and it never notifies anyone. A commitment somebody else made is held as your own follow up, phrased as something to check rather than something to delegate.

Where the fit is good, and where it is not

The fit is good at solo and boutique scale, where one person carries the relationship history for the whole firm and there is no knowledge function to lose it to. It suits the practitioner who is at once the seller, the deliverer and the person who has to remember what was said on the last call of every account. It fits worst as a firmwide programme, and pretending otherwise would waste your time: there is no shared workspace, no colleague view, no directory of everyone's expertise. A partnership of two hundred people asking about knowledge management is asking a different question, and the answer to it is a proper firmwide system plus the effort to feed it.

It is also not a document repository, a resourcing tool or a time and billing system. Those hold structured records with owners already. The gap this fills is the unowned one, the layer of commitments and reasons that no application in the firm is responsible for, and which no one has ever been paid to maintain.

Which half of the problem you are trying to fix

Ask what you would need if a familiar client type came back tomorrow with a similar brief. If what you need is the deck and the model, that is retrieval, and your folders will do at your size. If what you need is what actually happened, what the last client of this type resisted, what you underestimated, what you promised and then had to renegotiate, that is the live record, and it currently exists nowhere except in the mail itself.

The industry-neutral version of this argument sits at AI knowledge management, with the concept underneath it at what an AI company brain is. Two neighbouring professions have the same split with different stakes: knowledge management for law firms, where confidentiality decides the architecture, and knowledge management in pharma, where one half of the problem is already inside a validated system. The long-run promise, once enough engagements have gone through it, is AI institutional memory, and the survey of what else is on the market is AI knowledge management tools.

Worth saying, since the demo runs on a drug program: this was built by someone running his own company, and what he knows is that operation, not yours. What carries across is the same in every advisory business: mail, promises, dates and files. Domain judgment stays with you.