The Expectation Gap · Part 1 of 8

Everybody uses AI. Almost nobody uses it deeply.

Everybody uses AI. Almost nobody uses it deeply.

Half the country is now talking to a chatbot. Almost none of us are actually running our businesses on one. That gap β€” between "I use AI" and "AI does my work" β€” is the most interesting thing happening in this industry, and nobody selling you a subscription wants to talk about it.

Everyone uses AI, almost nobody uses it deeply

Here are the numbers that started this whole series for me.

Forty-nine percent of US adults used an AI chatbot last year, up from 33% the year before (Pew, Feb 2026). Among lawyers β€” my people β€” it's higher: 71% of solos and 75% of small firms touch AI in some capacity (Clio, 2026). Sounds like a revolution.

Now look at the other column. Only about 3% of users pay for any of it (Menlo, 2025). Only 8% of solo attorneys and 4% of small firms say they use AI "widely or universally" (Clio). Microsoft β€” Microsoft β€” has sold paid Copilot seats to roughly 3–4% of its 450-million commercial base (company figures, Jan–May 2026). The UK government ran a Copilot trial across thousands of civil servants and found 72% satisfaction, "no robust evidence" of a productivity gain, and an average of 1.14 actions per user per day. Just over one.

That's the shape of it. The market is a mile wide and an inch deep.

I don't think that's because the tools are bad. I think it's because of what we're actually asking them to do β€” and where those two things stop lining up.

When you look at what people give an assistant to do, it's not exotic. Find me some information. Draft this email. Summarize this document. For lawyers specifically: drafting correspondence (58%), general research (58%), summarizing (47%) β€” the boring, daily, load-bearing stuff (8am Legal, 2026). Nobody's asking the robot to argue their motion. They're asking it to clear the desk.

Those are exactly the jobs a competent AI should nail. So why the inch-deep usage?

Because the moment those jobs get real, two walls go up.

Wall one: "I can't put this into the tool." The document you actually need summarized is a client's file. The email you actually need drafted quotes a diagnosis, or a settlement number, or someone's tax return. And ninety-six percent of legal professionals say they require confidentiality safeguards before they'll use AI at all (Thomson Reuters, 2026). So the file stays on your desk, and you use the miracle machine to write a LinkedIn caption instead. If you're a doctor, a therapist, an accountant β€” you already know this feeling. The tool is most useful on exactly the material you're not allowed to feed it.

Wall two: "It doesn't know my stuff." Even when you can paste something in, the assistant has no memory of your matters, your calendar, your last call, the brief you wrote in 2023. MIT's big 2025 study found the single most common reason these tools fail in practice is that they "don't retain context or learn." You end up re-explaining your world every single time, which is often slower than just doing the thing yourself.

So we've built a strange situation. The most capable software most of us have ever touched, and we're using it for the least sensitive, least contextual 5% of our work β€” the party trick β€” while the actual job sits untouched two inches away because we can't safely hand it over.

That's the expectation gap. Not "can AI do it?" More like: "can AI do it with my stuff, without my stuff leaving the building?"

For most of the tools you've heard of, the honest answer is still no. And that's not a bug they're rushing to fix β€” it's the foundation their business is built on. Your data going to their cloud is the product.

I've spent the last stretch building the opposite bet: an AI that runs on a box in your office, that knows your files because they never leave, and that answers your phone whether or not OpenAI is having a good day. Over the next few posts I'm going to break down β€” with receipts β€” exactly where the mainstream tools hit these walls, what a local box fixes, and (just as important) what it doesn't. I'm not going to pretend the local option is magic. Part of why I trust it is that I know precisely what it can't do.

Next up, Part 2: the wall that matters most for anyone with a duty of confidentiality β€” "I can't give it my stuff" β€” and the federal case that turned that worry from a feeling into a legal problem.

If you've caught yourself using ChatGPT for the easy 5% and doing the real 95% by hand, that's not you being behind. That's the gap. Tell me I'm wrong.

β€” Banksy AI

Practical AI, done for you. Runs on your hardware. Your data never leaves.


Confidential AI that never leaves your office

The Banksy Box runs on hardware you own. Your clients' data never goes to anyone's cloud.

See the Banksy Box Talk to us
Part 2 → ↑ All 8 parts β€” The Expectation Gap