Deploying intelligence is different from deploying software

The firm of the future is being built right now. Not in strategy decks — in month-end closes, at firms that will tell you themselves they didn't feel ready when they started.

A year ago, we argued that intelligence becoming abundant would not make deployment easy. Now, after working inside firms already making the shift, we're seeing what successful deployment actually asks of a firm.

For most leaders we work with, the strategic question is settled: AI will be part of how the firm reaches its goals. The consequential question is the one behind it: is the firm ready to make the operating-model changes deployment requires?

That question deserves careful consideration, because it comes from experience. Every practice runs on more than its documentation. SOPs rarely capture 100% of reality; processes inevitably vary a little by service line, by office, by manager; and a great deal of what makes client work excellent lives as tribal knowledge, carried by your best people, without necessarily being written down. In the software era, all of that stood between a firm and a new system — software needed everything defined before it could be put to use.

The reason lies in the nature of accounting work itself, which breaks down into three parts:

1. Context — the client data, documents, and history the work depends on
2. Procedures — how each part of the work should be done for this client, in this period
3. Judgment — the professional call that makes the output right and client-ready

Software demanded the first two in finished form: context captured in predetermined, structured systems, procedures written down tightly and uniformly. If your practice serves every client a little differently — and every good practice does — that was the friction, and it often led to software breaking when it encountered edge cases.

Intelligence is different. Agents take context from wherever it lives — the shared folder, the email, the report in whatever format it arrives — and structure it themselves. Procedures no longer have to be perfectly documented before work begins. Agents take a first pass using the context available; as reviewers correct the work and teach client-specific rules, that knowledge becomes organized into repeatable procedures. Judgment is never handed over — your people still decide what posts to the ledger and what makes the work client-ready.

Whereas software deployments would have punished firms that started with less defined operating models, intelligence deployments actually help shape those operating models.

Feeling unready is not a reason to wait. It's the reason to start.

Inside a deployment, the change shows up in roles more than in systems. Preparers have agents take the first pass and apply judgment to finalize the output, in effect shifting to become reviewers. Managers become reviewers twice over — of the work people and agents produce together, and of the deeper client service their freed-up time makes possible. Whether a client's data sits in three disconnected systems or arrives by email doesn't decide readiness; agents work across whatever exists. And the deployment tells you it's working when the questions change: in week one, a first-time user asks whether the agents can handle a task. Within a few weeks, it's what else can they take on? Soon after, the conversation is with managers and partners about what it means for the whole practice.

The firms already building the future practice all start from a similar position, and it isn't perfect processes. They have a clear vision of the practice they want, the conviction to make bets on it, and a partner in redesigning what it means to prepare and to review. The best practitioners we know are making those bets now — on their clients, their people, and their own future.

So what does readiness actually require? Context — which doesn't need to be structured, because agents take what exists today and structure it for you. Procedures — which don't need to be written down first, because reviewer corrections and client-specific rules become organized and repeatable as the agents run. And judgment — which stays where it has always belonged: with your people. Work this way for a few months and you'll hold the thing you were waiting for: an AI-native operating model, built on live work.

You're more ready than you think.

Next in this series: what it actually feels like to have agents orchestrating work inside a practice.