The AI Lab Weekly

Welcome back to The AI Lab Weekly!

Before we meet:

The whole system comes down to one five-word loop: fork → branch → commit → pull request → merge. Understand that sentence and you understand 80% of GitHub. The terms to walk in knowing:

Michael will cover the rest on tomorrow’s call.

"Is it safe to put client data in this?" is the question that stops most firm owners cold. So we answered it — for 228 common tax and accounting tools.

Here's the number that should change how you think about your stack: only 42 are safe to use out of the box. 161 are safe only after you get one specific setting or tier right. Safe-by-default is the exception, not the rule — and until now, nobody was telling you which was which.

We read the actual terms of service for all 228 and recorded where each lands for client data. Every entry dated and cited to primary source. Each tool gets one of four statuses: Supported (42), Conditional (161), Not supported (17), or Undisclosed (8). Behind each sits a five-check scorecard: no-training, DPA, certifications, US residency, retention.

And so you know exactly where we stand: we have no financial stake in any tool on this list and we're not sponsored by any of them. Proof of that is in the index itself — we ran the platform we build on through the same five checks as everything else, and its consumer tiers came back Conditional. No pass because it's ours. This exists to protect client data, nothing else.

It's a group project, so it needs you in it. Open any tool and hit Flag for review if you think it deserves a different rating. Every flag gets re-verified against primary source. Vendors change their terms without notice. The index only stays accurate if the people using these tools every day keep us honest.

See the full index and how to use it

Charlie's been building a federal tax research tool for the Lab, and it's ready for a small group of testers. It's called Ask ALFA.

What makes it different from pasting a question into ChatGPT: it's built to answer only from primary authority (the Code, Treasury regs, IRS guidance, court opinions), to cite every point, to check every figure and quote against the source before you see it, and to refuse when the authority isn't there instead of guessing. A refusal is the tool working, not failing.

Those are the design goals — and your job is to find where they don't hold yet. Test it on real research questions from your practice: anonymized facts only for now, federal law only for now, and you stay the reviewer of record. Try to break it. Tell us what's wrong, missing, or refused when it shouldn't be.

(This is the same rule we just published for every other tool, applied to our own first: anonymized in, you review before anything leaves.)

A few spots, first come first served. Reply to this email or DM Charlie in Slack.

We're finishing the first paid build Labs and if you haven't answered our questions yet, there's still time. Your input decides what we build first, so tell us where you're stuck and what you actually want to be able to do. These are build Labs: you'll leave with something working, not a page of notes.

Answer the questions here

Getting good with AI yourself is step one. Getting your team good with it is where the hours actually come back. Our Ops/Admin path lands mid-August, built for the people who run your firm's day-to-day. Start thinking now about who should be in the room and let us do the guiding. You don't have to be the one who teaches this!

This Week’s Live Call:

Recent Recordings:

Wednesday, July 15, 2026 - Member Builds: Collin Kane