The Big Four built AI agents for the work you do by hand. Here is how to copy them.

What the Big Four actually deployed

The headline is that the largest accounting firms are no longer piloting AI. They are deploying it as agents across audit and tax, and the scale is worth reading carefully so you copy the right thing.

As reported across Accounting Today, NYSSCPA, and the 2026 industry roundups, the moves break down like this. EY put roughly 150 AI agents in the hands of its tax professionals. KPMG committed about 2 billion dollars over five years and launched a workbench product to put AI in front of its people. Deloitte introduced an offering called Zora, aimed at invoice and financial analysis. PwC scaled agents, described as GL.ai or an agent operating system, for journal-entry and general-ledger review.

Treat every one of those numbers as reported, not absolute. The exact count of agents and the precise spend matter less than the shape of the strategy, which is consistent across all four. They are pointing software at the lowest, most repetitive layer of the work first.

What is striking is how aligned the four firms are. Nobody led with the partner's judgment work or the client relationship. They all started at the bottom of the engagement, the high-volume, rules-bound tasks that a staff accountant used to grind through, and built the agent there. That alignment is a signal. When four competitors independently reach for the same layer, it tells you that is where the labor cost was hiding and where automation pays back fastest. A solo practice has the same cost in the same place, just at smaller scale.

What an audit or tax agent really does

Strip away the product names and an agent is doing something plain. It takes a task that follows rules, journal-entry review, general-ledger tie-outs, invoice analysis, and runs it across the whole population instead of a sample, then surfaces the items that look wrong.

That is the part worth absorbing. The agent is not replacing judgment. It is replacing the hours a junior used to spend touching every line so that a senior could review a fraction of them. It does the first pass at full coverage and hands up the exceptions. The human reviews what the machine flagged, asks why, and makes the call.

This is exactly the structure of the work a solo CPA or fractional CFO does alone, except you are both the junior and the senior. You do the line-by-line pass yourself, then you review your own work. The agent lets you stop being the junior.

There is a quiet benefit hiding in that shift. When you do the full pass yourself, fatigue is the enemy. By line three hundred of a ledger your attention is gone, and the errors you miss are the ones that show up at the bottom of the file. An agent does line three hundred with the same attention as line one. It does not get tired, it does not get bored, and it does not start skimming. What it lacks is the ability to know which exceptions matter, which is precisely the part you keep.

How a solo CPA copies the pattern at small-firm scale

You do not need an agent fleet. You need to apply the same logic to one task at a time. The general firm pillar in how accounting firms run on AI covers the broader operating model. Here the point is narrower: copy the agent pattern on the tasks that eat your week.

• Pick a rules-based task you already do by hand: monthly reconciliations, expense coding, a ledger review, a flux analysis on the close.
• Give the model the full population, not a tidy sample. The whole bank statement, the whole GL export, the whole expense file.
• Ask it to do the first pass and flag exceptions: duplicates, miscodings, balances that do not tie, entries that break a rule you state.
• Review only what it flags, the same way a partner reviews a junior's exceptions, and own the conclusion.

The AI tools accountants actually use briefing covers which tools fit which task. The skill underneath is the same one the Big Four are buying at scale: defining the task tightly enough that a machine can do the first pass and you can trust the exceptions.

Where this saves real time, and where it does not

The honest version. The biggest gains show up on volume work with clear rules. A reconciliation across hundreds of lines, an expense file with obvious coding patterns, a ledger tie-out, these are where an agent earns its keep, because the rules are stable and the human cost was always the boredom of full coverage.

The gains shrink fast on judgment work. Materiality, a going-concern call, how to treat an ambiguous transaction, the tax position on a genuinely gray fact, these do not hand off. A more capable model will lay out the considerations and miss fewer of them, but it does not carry the professional risk. Your license does.

There is also a trap specific to agents. Because the output looks like finished work, an unchecked agent pass is more dangerous than a blank page, not less. The better it looks, the easier it is to sign without reading. The Big Four's structure handles this with a review layer. As a solo, you are the review layer, and skipping it is the one mistake that turns leverage into liability.

The other limit is data. A reconciliation or a ledger export often holds client identifiers, and pasting an entire file into a public tool is a privacy decision, not just a productivity one. Keep client identity out of any tool that is not covered by your own safeguards, work with de-identified exports where you can, and add sensitive detail back inside your own systems. The point is not to avoid the tool. It is to be deliberate about what crosses into it, the same way you already are with email and your portal.

What to do this week

Take the single most tedious volume task on your desk, the reconciliation you dread, the expense file, the ledger review, and run it as an agent pass once. Give the model the full file, ask for a first pass with flagged exceptions, then review only the flags. Time it against doing it by hand.

If it saves you an hour and you still trust the result after reviewing the exceptions, you have just copied a Big Four pattern at your scale. If the close itself is your bottleneck, the CPA close-week AI system briefing lays out the full sequence.

The skill under the agent

Every firm above is spending real money to buy the same thing: a method for handing the repetitive layer to a machine and keeping the judgment with a licensed human. The agents change. The method does not. The solo CPAs who pull ahead are not the ones with the most tools. They are the ones who know exactly which tasks they hand off, what they check, and where their name goes.

That is the part worth building, because it survives every product launch and every new agent. Claude AI for Accountants teaches that method from the first reconciliation to the signed file, and the two minute course quiz will point you to the right program for your practice.