Treadstone Associates
Article · Capacity & pricing

AI for capacity planning at a firm

You plan capacity by forecasting the work your existing client book will generate, week by week, against the filing deadlines it is anchored to, and setting that against the hours you actually have. AI’s contribution is building and refreshing that forecast out of data you already hold; the staffing decisions stay with a partner.

Treadstone Associates · Updated 2026

Key takeaways

  • • A practice can forecast demand better than most businesses because the deadlines are statutory.
  • • Forecast in weeks. An annual capacity figure hides every crunch that matters.
  • • The inputs are prior-year job records, year-end dates and staff availability — nothing exotic.
  • • Reviewer hours, not total hours, are usually the binding constraint.

Why the demand side is unusually knowable

Most businesses forecast demand from a pipeline. A practice forecasts it from a calendar, because the dates are written into legislation. Under section 150 of the Income Tax Act, a corporation’s return of income is due within six months after the end of the year; an individual’s is due on or before the following 30 April, except that an individual who carried on a business in the year files on or before 15 June.

Sales tax adds a second set of anchors. Under section 238 of the Excise Tax Act, a registrant whose reporting period is the fiscal year generally files within three months after the end of the year; an individual with a calendar fiscal year who carried on a business files by 15 June; and in every other case a return is due within one month after the end of the reporting period.

Every client in your book therefore carries a set of dates you already know a year in advance. That, not a sales forecast, is the spine of a capacity plan.

Build the demand curve before you buy anything

Four columns get you most of the way: client, year-end, service lines, and last year’s actual hours by role. Export them from whatever you already use. The output you want is not a total — it is a curve showing planned hours by week, split by the role that has to do them.

This is where AI is genuinely useful, and the reason is unglamorous: the data is messy. Time entries are inconsistently coded, job names drift, one client appears under three spellings, and a service line that was scoped as a compilation was actually delivered as something larger. A model is good at reconciling that mess into a job-level history, grouping clients by pattern, and drafting the plain-language summary of what changed since the last run.

Be clear about what you are getting. A forecast built from your history inherits your history, including the years you were understaffed and the jobs that ran long because a client sent everything in the last week. It is a starting position for a conversation, not an answer.

The supply side is arithmetic, and firms overstate it

Available hours are contracted hours minus statutory holidays, vacation, professional development, internal meetings and the non-chargeable administration that every practice pretends is smaller than it is. Run that subtraction honestly once and the plan improves more than any software will improve it.

Then add the line most plans omit: review capacity. Preparation can be spread across more people, hired seasonally, or pushed to a longer week. Review usually cannot, because the reviewer must be someone specific. When a practice says it ran out of capacity in March, it almost always means it ran out of reviewer-hours in March.

What AI changes, and what it does not

It changes the refresh rate. A plan that took two days to rebuild got rebuilt quarterly and was wrong by February. A plan that rebuilds in an hour can be re-run every Monday against actuals, which is the only version that ever changes a decision.

It also helps with drift detection — which jobs are consuming hours faster than planned, which clients have not sent anything yet, which service lines are quietly larger than they were scoped to be. Those are pattern questions over your own data, which is what these tools do well.

It does not decide who works on what, and it should not be allowed to. It does not know that a client is about to leave, that a manager is about to go on parental leave, or that a junior is not yet ready to run a file alone. Those inputs come from a partner, and they change the plan more than any model output will.

What you do when the curve shows a wall

There are only four moves, and a forecast is only worth building if you are prepared to make one of them: pull the work earlier by requesting information sooner, add capacity for a defined window, decline or defer work, or reprice. Doing nothing is a fifth option that shows up in March as unbilled overtime and a missed deadline.

A worked example

A six-person practice with roughly two hundred corporate clients and a personal tax book exports three files: a client list with year-ends, a job list for the last two completed years, and time entries against those jobs. A model reconciles the time entries to jobs, produces average hours per job type by role, and lays the result onto next year’s calendar using each client’s year-end and the statutory due dates.

The curve shows two peaks the partners already felt but had never seen: a review peak in the last three weeks before the personal deadline, and a second one in the sixth month after the most common corporate year-end. The preparation peak is manageable. The review peak is not, because it lands on one person.

The decisions that follow are ordinary practice management, not technology: move a defined set of corporate year-ends into an earlier information-request cycle, bring one manager to reviewer standard on a specific job type over the summer, and stop accepting new personal work after a cut-off date. The plan is re-run weekly against actuals, and the only number anyone looks at is reviewer-hours by week.

Questions we get asked

Do we need to buy something first?
No. Build the first version from exports and a spreadsheet. If it does not change a decision, a tool will not change that. Once the plan is worth keeping, the tooling guides in our Business & Professional Services hub cover what to buy.

Can a model predict how long a job will take?
It can extrapolate from your own history, which is different from knowing. Treat the output as a prior that a manager adjusts, and track how often the adjustment was right — that is the number that tells you whether to trust it next year.

What about the firm’s own bookkeeping and reporting?
That is a different job, covered on our accounting automation and bookkeeping automation pages. Capacity planning is about the work you sell; those pages are about the work you do on yourself.

Where does scheduling and the front desk fit?
Appointment scheduling, reminders and reception are a separate problem with a separate answer — see professional practice owners.

See where AI pays off first in your practice.

A 30-minute call is enough to tell you whether AI pays for itself here.