AI data access for accountants is the leadership decision most CAS firms have already made — by never deciding. A few weeks ago I needed to track down a billing dispute. A client had pushed back on a fee, and I needed to find what we'd actually told them and when. Without AI, that's a 30-minute hunt through email, our practice management system, and billing exports. With AI that had access, it took seconds. The agent just had to be allowed to look.
The YC moment that matters
In late May, Y Combinator's general partner Pete Koomen and CEO Garry Tan walked through, on camera, the internal AI stack YC uses to run itself. One detail explained the rest. Early in the project, an engineer — almost apologetically — gave the agent direct read access to YC's production database. "Surreptitiously, late at night," Tan put it. Their realization: "the thing that was hampering the world was being worried about security and all the things that could go wrong. And when you worry a bit less, these things are unbelievably powerful."
That's the actual lesson from the YC stack. Not the tooling. Not the model. Access. Monday's roundup framed the week as AI getting good, getting expensive, and demanding leadership — this is the leadership thread. The decisions that separate firms over the next five years aren't tool selections. They're access decisions.
The same dynamic is alive in your firm
Most CAS firms have AI locked away from the data that would make it useful. We worry about security and confidentiality — and so the AI gets nothing. Then we wonder why it isn't very smart.
Ask yourself which of these questions your AI could answer today:
- Find every email exchange with this client about pricing and tell me where we've made promises we haven't yet delivered on
- Which clients haven't filed in three years, and what's the state of their general ledger?
- What was the actual revenue mix across my book last quarter — by service line, by industry, by client size?
None of these are practice management system questions. They cross email, tax filing systems, billing, and individual client ledgers. In most firms, AI can answer none of them — not because AI can't, but because AI isn't allowed to look. Scoped AI access for accounting firms is the bridge that closes that gap.
You're asking the wrong question
The default question in most firms is "should we give our AI access to our data?" That question has one safe answer, and it guarantees weak results. The real question — the AI access policy every firm needs to answer — is three-part: how much access (read-only or read-write), of what scope (firm-internal or client data), and with what cross-client controls.
The cross-client piece is subtler than it looks. As a partner, I want cross-client AI analysis — revenue mix, capacity, who's at risk — and that's essential management information, not a confidentiality violation. The codes-of-ethics line isn't "AI must not see across clients." It's "one client's data must not appear in another client's deliverable." Two different things.
Three access patterns — and only one is honest for accounting
- Locked down. AI sees nothing. You stay safe and extract no value. Where most firms are now.
- Wide open. AI sees everything. You get YC-style power, and you violate your professional obligation. Where no accounting firm should be.
- Scoped. AI sees what it needs, when it needs it. Per-client containers structurally isolate client data so cross-client leakage is impossible — not because policy says no, because the architecture prevents it. Firm-internal methodology stays more open, so it compounds across the firm.
That third pattern is the only honest answer for a CAS firm — and almost nobody is building it today.
The leadership question
Who in your firm decides what access AI gets? In most firms, nobody does. The most cautious voice wins by default, AI gets nothing, and the firm extracts a fraction of the value it's already paying for. The Champion, the IT lead, the managing partner — somebody has to own this question and answer it deliberately, not by accident.
Friday's piece on the OpenAI/Thrive tax agent named the Champion as the firm's forward-deployed engineer — the person who operates the encoding loop. The access question lands at the same desk. The Champion who decides what gets encoded is also the Champion who decides what AI is allowed to see in order to encode it. Same role, same person, two sides of the same job.
The compounding effect nobody talks about
When asking questions of your firm's data gets cheap, you don't just ask the same questions faster — you ask more of them, and bigger ones. Questions like "what's the early-warning signature of a client about to churn?" or "are any of my staff showing the work patterns of people who left last year?" silently die today because the cost of answering them is too high. With AI access in place, those questions get asked routinely, and the leaders make different decisions. The partners who win the next five years will be the ones operating with cheap, AI-powered visibility into their own firm. The ones who lose will be the ones still asking the questions they could afford to ask in 2024.
What to do this Monday
Three moves. First, audit what your AI currently has access to — in most firms, the answer is "almost nothing." Second, list three questions you wish you could ask of your firm right now and trace why your AI can't answer them; the answer is almost always access. Third, decide who owns the access question. If nobody owns it, the default-of-no will continue, and you will keep paying for AI that delivers a fraction of its promise.
AI's value compounds with the access you give it. Locking it out is a choice — usually an unconscious one. The firms that design access deliberately will look meaningfully different from the firms that don't.
Working through questions like this is what the AI Practice Transformation Program is for. If your firm hasn't named who owns the access question, it's worth a conversation. Visit theaiaccountant.ai/transformation.