What We Won’t Let AI Do

Post 5 of 5 — The limits, the safeguards, and the rules we will not break

The first four posts in this series have been about what AI is doing inside our firm. This last one is about what it is not doing, and what it will not do.

Because everything I have described in the last four posts only matters if it is done responsibly. And the responsibility piece is not an afterthought. It was the first thing we worked out, before we deployed any of these tools.

Three rules we will not break

There are three rules that govern every AI deployment inside our firm. None of them are negotiable.

1. Every platform meets the same security standard as the rest of our software

Every AI platform we use must meet SOC 2 Type 2 standards — the same bar we apply to our tax software, our accounting platform, and every other system that touches your data. That is not a guideline. It is a baseline.

We do not allow AI models to train on client data. We use read-only integrations where the technology supports them, so the AI can answer questions about your information without having permission to change it. Multi-factor authentication, access controls, automatic logout protocols — the same disciplines that govern your tax file govern every AI tool that comes near it.

2. A human reviews and approves every AI-touched work product

No AI-generated work product reaches you without a human on my team reviewing and approving it. Memos, analyses, emails, deliverables — all of it. AI helps with assembly. People do the approval. There are no exceptions, no shortcuts, no “this one was simple enough to skip the review.”

This rule is the difference between AI as a force multiplier for a thoughtful firm and AI as an excuse for a sloppy one. We are committed to the first version.

3. Some judgment is and will remain a human one

There are categories of work where the answer is and will remain human. I want to name them, because I think you should know the line.

Interpreting your intent on a hard tax position. Sitting with you when something has gone wrong with your business — a bad quarter, a partner dispute, a sudden cash crunch. Advising on a family business decision that has tax implications nobody is supposed to optimize away. The conversation that happens before a major sale or transition. Anything that requires us to know who you are, not just what your numbers are.

AI does not get those moments. The people who know you do.

Transparency, by default

There is one more thing I want to commit to in writing. We will tell you when we are using AI on your work, and what the human review of that work looks like. You should never have to guess.

A piece of research I keep coming back to: clients who understand their firm’s AI process tend to trust it. Clients left to imagine the process tend to imagine the worst. That is a fixable communication problem, not an unfixable trust problem, and it is on us to fix it. This series of posts is part of that fix. So is the standing offer to answer any specific question you have about how a particular deliverable was produced.

Closing the loop

Five posts ago, I told you that the work itself is changing. Then I told you about Chad and how our team is learning to work alongside him. Then about the two kinds of AI we are getting fluent in — generative and semantic. Then about why I am not letting AI turn this profession into a race to the bottom on price. And now this: the rules we will not break, and the human judgment we will not delegate.

Put it all together and what you are seeing is a firm trying to do something genuinely hard: adopt a powerful new technology in a way that makes our work to you better, not cheaper, and not riskier. The technology is new. The standard isn’t. You hired a firm to be careful with your business — that is still our job. We just have better tools to do it with.

Thank you for reading the series. I will be back with more, but probably less of it for a while.

If you want to talk about anything in this series, or about the engagement we have together, email us at [email protected]. We’d love to continue the conversation!