The difference between using AI and being AI-native
Almost every firm now uses AI somewhere. A partner drafts a client email with it, a senior asks it to explain a code section, someone summarizes a long document. That is real value and it is worth having. It is also not what makes a firm AI-native, because nothing about how the firm works actually changed. The same steps happen in the same order, a few of them slightly faster.
A firm becomes AI-native when the process is rebuilt on the assumption that the mechanical parts are handled. That is a different question than which tool to buy. It means looking at an engagement and asking which steps exist only because a person had to move information from one place to another, and then removing those steps rather than accelerating them.
The test I use: if you turned the AI off tomorrow, would your firm be slower, or would it be broken? A firm that would merely be slower is using AI. A firm that would have to rebuild its process is AI-native. Neither answer is wrong, but they are not the same thing, and only one of them changes your capacity.
What it looks like in a real engagement
Here is the shape of a return in my own practice, which is the honest version rather than the demo version:
- Intake runs itself. The engagement folder is scaffolded, the client record is drafted, and the document requests go out with completeness checks, so the chasing is mostly automated rather than mostly mine.
- The documents get read. W-2s, 1099s, K-1s, brokerage statements, and 1098s are parsed into structured data. The messy scans and the complicated K-1s get flagged for me instead of guessed at.
- The workpaper assembles itself. Income is sorted, indexed, and tied out into a review-ready set, with open items left open.
- I open a draft, not a folder of PDFs. The first pass exists before I start. My time goes to the judgment calls, the planning conversation, and the review.
- I sign it. The responsibility never moved. That is the point of the whole design.
The gain is not that the return takes ten minutes, because it does not. The gain is that the hours go to the parts a client is actually paying a CPA for, and that the same person can carry more clients without the quality sliding.
The four layers underneath it
Firms that make this work tend to have four things in place. Firms that struggle are usually missing the third or fourth.
- Structure. Consistent folders, consistent file naming, and a written record of what each client is and what they need. AI is very good at working inside a structure and very bad at inventing one you never decided on.
- Context. Somewhere the firm has written down how it works: the standards it follows, the conventions it uses, the things that are always true. Without that, every session starts from zero.
- Guardrails. An AI use policy, a security plan, client disclosure, and vendor diligence, so the answer to what may go into which tool is already decided rather than improvised per person. I give those away as free templates.
- A review point. A named human who reviews before anything reaches a client or a return. Not a policy sentence, an actual step in the workflow that cannot be skipped.
What AI-native does not mean
- It does not mean no people. It means people spend their time on review, judgment, and clients instead of on retyping a W-2.
- It does not mean the AI signs anything. Professional responsibility for the work does not move, and no tool changes who is accountable under Circular 230 or to a state board.
- It does not mean client data goes anywhere convenient. The tier of tool you use has to match the sensitivity of the data, and a consumer chatbot that trains on its inputs is not the same product as a business plan with no-training terms.
- It does not mean buying one platform. No vendor sells AI-native. It is a property of how your firm works, so it is something you build, in pieces, on the workflows you actually have.
How a firm actually gets there
Not by planning for a year, and not by buying the tool at the top of a conference sponsor list. The firms that get traction pick one workflow that is genuinely painful, rebuild that one properly, and let the second one be easier because the structure and the guardrails already exist.
A reasonable order of operations: put the guardrails in first, because they take a week and they unblock everything after. Then pick the bottleneck that costs you the most hours in your worst month. Then build that one, with a review step, and use it for a full cycle before you touch anything else.
If you want the version of this written as a setup walkthrough rather than a definition, the Claude Code guide for CPAs is the one I hand people, and the engagements page is how I do it with firms directly.