The work wasn't accounting. It was waiting.

An accounting firm is a system for moving work from intake to delivery, and like any such system, its throughput is set by its slowest queue. We went into our first engagements expecting that queue to be the accounting itself: the preparation, the review, the technical work that piles up in season. It was not. Map a firm honestly and a second queue comes into focus, usually larger than the first and tracked by nothing: the work that is sitting idle, holding for something it needs before it can move at all. The accounting was rarely what held a firm back. The waiting was.

Two queues, and only one of them is instrumented

Every firm we worked in could see its active queue in fine detail. What is in progress, who owns it, what stage it is at, when it is due. That queue is measured, reviewed, and managed like a real thing, because to the firm it is one.

There is a second queue, and it is the one that actually governs the firm. Documents requested but not returned. Engagement letters sent but unsigned. Questions asked but unanswered. Payroll inputs needed but not provided. It is usually larger than the active queue, and no system in the building tracks it, because the moment a request goes out the door it stops being the firm’s work and becomes the absence of the client’s. When we asked a firm to list everything currently sitting in that queue, across every client, almost none could do it without rebuilding the list by hand. The single most consequential queue in the firm was the one nobody could pull up.

Why that queue has no sensors

The reason is mechanical, not careless. Active work is instrumented because it lives inside the firm’s systems and moves through defined states. Waiting work is uninstrumented because the instant a request leaves the firm, the item exits the firm’s view entirely and sits in a client’s inbox, somewhere the firm’s tools cannot see into. Practice software is built to track what the firm is doing, not what it is not yet able to do. So a firm’s largest single source of delay is precisely the part it has no sensor pointed at. The first thing worth building, before any automation, is that missing sensor.

Collection is a process wearing the costume of a field

On an engagement plan, “gather client information” is one line, a box to be checked. Inside the system it behaves like a process with many states: requested, partially received, clarified, re-requested, verified, rejected, requested again. It has rounds. It has exceptions. It has no terminal state until a person declares one. Onboarding, engagement letters, bookkeeping inputs, payroll, the document-gathering phase of a return: each is a multi-state process modeled as a checkbox. And anything modeled as a checkbox gets no owner, no aging, and no escalation, which is the entire reason it stalls. The work was never mis-sized because people were careless. It was mis-sized because the data model said “field” where reality said “workflow.”

The real build was the operating system underneath

The model was never the difficult part. The work was the layer beneath it: an explicit, shared model of how this particular firm operates. Systems that were never designed to exchange data. Procedures that varied not just firm to firm but partner to partner inside the same firm. Exceptions everyone observed and no one had written down. Client data arriving in whatever shape the client felt like sending. Entire procedures that lived only in one long-tenured person’s memory and had never once been expressed as something a machine could follow. Turning that undocumented set of human habits into an operating system the automation could run on top of, that was the engagement. Anyone selling an “AI solution” that skips this layer is shipping a sensor with nothing underneath it to read.

What instrumenting the wait changed

Give the waiting queue the same instrumentation the active queue already has, an owner, an age, a next action, a state anyone can see, and the firm starts behaving differently before a single reminder is automated.

Chasing stops being an act of memory and becomes a property of the system: every outstanding item is visible, aging, and assigned, so follow-up happens whether or not someone remembers it. Status reporting stops being a meeting and becomes a query. And the firm can finally separate two failure modes that had looked identical from the inside: work that is genuinely slow, and work that is merely waiting. Those are different problems with different fixes, and you cannot tell them apart until the wait has a sensor on it. Most of what improved was not the speed of any task. It was the disappearance of dead time between the states.

Where judgment still matters

Knowing what is outstanding is a systems problem, and we treated it as one. Deciding how hard to lean on a specific client, and when a routine reminder should give way to a partner picking up the phone, is a judgment problem, and it stays with people. The system’s job is to guarantee that nothing is lost and to surface the few items that genuinely need a human, with the context already attached. It is not to decide how to handle the client who is plainly going through something. That distinction held in every firm, and the firms valued it.

Not slow. Waiting.

The firms that pulled ahead were not the ones with the strongest accountants. The talent was comparable across all of them. They were the ones that stopped treating waiting as an absence and started treating it as a queue: a real thing with a state, an owner, and a cost. Once it was visible it was manageable, and once it was manageable the firm got faster without anyone working harder. The accounting was never the constraint. The work was not slow. It was waiting.

Related: how we approach AI for accounting and professional-services firms: instrument the waiting, then drain it.