Why are accounting firms under pressure to use AI now?
Private equity is consolidating the profession fast. Fewer than 200 PE investments in accounting firms led to some 900 follow-on deals in 2025, most of them roll-ups, according to IFAC figures reported by CFO Brew (March 2026). Deals per PE investment have risen fourfold since 2021.
Thrive Holdings shows where this goes. It raised $2B in August 2026 and runs an accounting platform called Current, with more than 50 firms and 2,000 professionals. According to OpenAI (May 2026), OpenAI's forward deployed engineers worked with Thrive's engineers on tax-preparation agents for those firms.
An independent firm competes with platforms like that for clients and for staff. Keeping up without selling takes one workflow done inside the systems you already run, built by someone who sits with your staff long enough to see how the work gets done.
Which workflows are worth automating at an accounting or CPA firm?
Start with work that arrives in volume and already gets checked by someone senior. That review step is where the human stays, so you aren't inventing a new control.
Client document intake and chasing
Busy season starts with chasing. Clients send documents late and in pieces: phone photos, forwarded emails, portal uploads, the occasional shoebox. An agent can read what arrived, check it against that client's list (built from last year's return and the engagement letter), file it in the right folder and draft the reminder for what's still missing. That list is the hard part, so build it from data the firm already has.
Staff approve the reminders at first. When the drafts are reliable, routine ones can go out on their own, inside rules the firm sets. Anything sensitive stays with a person: if a client's note mentions a death or a divorce, someone who knows them replies.
Bookkeeping categorization and reconciliation
For client books in QuickBooks Online or Xero, an agent proposes categories for uncategorized transactions from the client's history and the bank memo, and flags the ones it isn't sure about. Recurring transactions get a plain rule instead. A rule is cheaper than a model and never changes its mind.
On reconciliation, the agent lists the unmatched items with a likely explanation for each, and the bookkeeper clears them. Each suggestion shows its reason, so a correction is quick and can become a rule for that client. A reviewer approves anything below the confidence bar the firm sets, plus anything that changes the tax treatment.
Tax prep data extraction and workpaper prep
An agent pulls figures from source documents, such as brokerage statements and partnership schedules, into the workpaper or the input sheet your preparers use. Plain code ties out the totals and flags any figure that moved more than the firm's threshold against last year.
Keep a link from every extracted figure back to the page it came from, so the preparer can check it against the source. The preparer reviews whatever the tie-out can't confirm, and the reviewer signs, same as today.
Month-end close support
For client accounting work, an agent can pull the transactions behind each variance against last month and draft the commentary. It can also track the close checklist and chase the client for missing items, in whatever tool holds the checklist today. A person posts every journal entry, and the commentary gets a review before the client sees it.
AP for clients
Firms that run bill pay for clients have the same problem as any AP team, repeated for every client. Each client has its own vendors and approval rules, so the agent works client by client and never mixes them. It reads the invoices and proposes the coding. A person approves every payment, and changes to vendor bank details always go to a person. We sorted each AP step, from intake to payment run, in our guide to AI accounts payable automation.
Client email drafts
Clients ask the same questions all season: where's my return, and what do you still need from me? An agent can draft the answer from the client file, and staff edit and send it in the firm's voice. The draft shows which part of the file it relied on, so checking it is fast.
Advice is different. Anything that tells a client what to do gets written by a professional, or reviewed line by line by one.
What stays human, and why standards come first
AI doesn't sign anything. The professional who signs keeps the judgment that goes with the signature, and the firm's standards and data rules decide what the AI may touch.
Settle those rules before anyone picks a model, and write down the answers:
- Which client data each workflow needs. Send the minimum.
- The model provider that processes it, under terms that rule out training on your clients' data.
- Where the logs live and who can read them. Retention follows your existing policy.
- Who at the firm approves a new use of client data. One name, not a committee.
- What your engagement letters say about technology and third parties, and whether they need updating.
- If you prepare tax returns, how the rules on tax return information apply. Ask whoever owns compliance at the firm before client returns go near a model.
Reviewers also need to see what the AI saw. Every draft should point back to its source documents, so review stays fast and the audit trail holds up when someone asks how a number got there.
We run this review in the Diagnostic, before the build. A security review left to the end is a classic way for an AI pilot to stall before production.
Build on the stack the firm already runs
You don't need a new platform to use AI. The agent should work inside your practice management system, tax software, ledger and client portal, where staff already spend the day.
Client books stay in QuickBooks Online or Xero. Intake runs through the portal clients already know, and checklists and deadlines stay in practice management. Clients already have a login to your portal, so don't give them another one for the AI.
Where a system has an API, the agent reads and writes through it, with an account that has only the rights the workflow needs. If there's no API, we look for an export or an email route before anyone suggests screen scraping.
The same goes for AI features your vendors already offer. If your tax or practice management software added something recently that does the job, use it. Building a copy of a feature you already pay for is waste, and we'll say so in the Diagnostic.
Mistakes accounting firms make with AI
The expensive ones come from starting in the wrong place.
- Starting with advice. Client-facing advice is the riskiest work in the firm. Automate the prep around it first.
- Leaving staff to find their own tools. If the firm hasn't given people an approved AI tool, assume some are pasting client data into an unapproved one. Give them a sanctioned option that follows the data rules above, and say plainly what's off limits.
- Buying a platform before picking a workflow. A license with no workflow behind it is shelfware by the next filing deadline.
- Skipping the baseline. Without the hours and turnaround from before, the partners can't tell whether it worked, and the next budget conversation runs on opinions.
- Mixing client data. An agent that works across clients has to check access client by client. A client's data should never show up in another client's draft.
Start with one high-volume workflow
Pick the workflow where staff spend the most hours on work a reviewer already checks. Measure it before you change anything.
In a tax practice, look at document intake and chasing first. A client accounting practice should look at categorization and reconciliation. Either way, the baseline is simple: hours per client spent on the workflow, days from engagement letter to a complete file, reminder emails per client, and review notes per return. Choosing that workflow is the first step of our AI implementation work.
Timing matters in a firm. Don't put a new workflow live in the middle of busy season. Run shadow mode in the quieter months, so the AI has seen real work before the volume arrives and your staff already trust it when the deadlines bite.
AI implementation for accounting firms: timeline and price
One workflow takes a 2-3 week Deployment Diagnostic, then a 6-8 week Production Sprint, at fixed prices.
The Deployment Diagnostic ($7,500 fixed) puts a senior engineer with the staff who do the work. They compare the documented process with the real one, get access to the systems, set the baseline numbers, run the data review above, and write an eval plan and a fixed quote. If you go ahead with the Sprint, that $7,500 is credited against it.
Then the Production Sprint ($25,000-$60,000 fixed) puts that one workflow live inside your existing tools. It comes with evals and an audit trail, human approval wherever money or risk is involved, a shadow-mode period before go-live and a written handover.
After launch, our AI Team retainer keeps it running from $3,000 per month. If you'd rather have a senior engineer inside the firm for longer, the Embedded Engineer option is 2-3 days a week for $16,000-$20,000 per month, with the studio behind them. That's our forward-deployed engineering model. Our CTO, Christian Vismara, is based in New York and can work on site with your team. Everyone else at the studio works remotely.
If the partners haven't agreed which workflow comes first, start with our five-day AI Readiness Audit (from $800) instead. It's built for owners deciding what to build first. The Diagnostic assumes you've picked the workflow and can give us access to the systems behind it.