Help it help you

The gap

In May the Census Bureau reported that "AI use increased among firms with at least 20 employees but didn't change significantly among firms with fewer than 20 employees." Most small businesses I know are in that second group, and not for lack of interest. These tools were sold as a smarter search engine, and a smarter search engine doesn't change how a business runs.

Source: U.S. Census Bureau, Business Trends and Outlook Survey, May 2026

Help it help you

Search Google for Chick Springs, the old mineral spring in Taylors, and you get Wikipedia and a handful of older pages to piece together. Ask a chatbot and you get the same public material in nicer sentences. But the 1817 survey of that land sits in the county's deed records, handwritten, and no search engine will read it for you.

A search engine answers from what's public. These tools can work from what's yours, and from the sources you connect them to, but only if you set them up to. Three habits do most of that work. Each one is a question you ask the AI about the work, not the work itself.

Most people stop at copy and paste. You export something, upload it to a chat, and get an answer. That works. It's also the slowest version, because you end up fetching whatever it asks for next.

The next step is connecting it. In Claude, you add connectors under Customize, then Connectors. Pick one, sign in, and it can work with that account using your permissions. The + menu in a chat turns each one on or off. QuickBooks is one of them, and I use it to dig through my own books without calling my accountant every five minutes. ChatGPT and Gemini have their own versions.

When there's no connector, ask how to get the data:

"How can we get you this data automatically?"

I'm not a developer. For the Chick Springs history I spent a long time searching the county's deed records by hand and pasting in what I found. When I finally asked that question, it built a small tool that searches the records and pulls up the scanned pages, which it can then read, old handwriting and all. I tested it and told it what was wrong until it worked. Five of the tools I've built that way are public on the tools page, but the point is the question, not the code.

Remember 50 First Dates? Drew Barrymore wakes up every morning with no memory of the day before, so Adam Sandler makes her a tape that catches her up.

Every new chat starts the same way. The model has no idea what you did with it yesterday. The app does its best to fill it in, but you can do much better by telling it what to write down. So before I stop, I ask:

"Before we stop, how do I get you back up to speed as fast as possible next time?"

Over time that turned into a routine. At the end of a session, the AI writes a short note: what we did, where things are, what went wrong, and what I corrected. The next session reads it first. Here's a real one from a day spent checking a client's email series, with names and details taken out.

## 11:29 Email series checked against the approved copy

- Checked the last five issues of a nine-part email series, plus
  five resend drafts, against the approved copy. Subject, preview
  and body match word for word. One difference: the source had a
  typo in a table, and the draft was right.
- Two calls Alex made: the blank recipient name on two issues stays, and
  one wording question was already settled in September.
- Corrected a wrong note of ours: resend drafts do not need to be
  scheduled by hand. I first told Alex they did, repeating a note
  from August. He pushed back, and the send history proved him
  right: 18 resends went out on their own. Fixed the project notes
  and both August notes.
- Marked the team's build task complete.

The third item is the one that matters. It got something wrong, I pushed back, and the correction went into the notes so the next session starts from the right answer.

You don't need a system to start. In a chat app, keep a running notes document in a Project (Claude and ChatGPT both have them), and at the end of a session ask the AI to write the update for you to paste in. In Claude's desktop app you can connect a folder and have it keep the notes file itself. The setup page walks through that.

You've worked on a report for weeks, and it makes perfect sense to you because you know every piece of it. Then a coworker reads it and asks how you got that number.

AI work needs that read more than your own does. When it's missing something, it tends to fill the gap with something plausible instead of telling you. So before anything goes out the door, I ask:

"Before we call this done, have a separate reviewer check it, without our conversation."

In a chat app, that means a new chat in the same tool. Paste in the draft and the sources, and ask it to check every claim against the sources. Give it the draft and the sources, not your explanation of why it's right, or it will just agree with you. Ask for the problems in order of importance, fix them, and run it again until it only finds nitpicks. If you pay for two tools, have the other one do the check. A different model has different blind spots.

Here's what that caught on a weekly ad report for one of my clients. It was one of four reports that morning, and all four changed before they went out.

• A line about money we had held back was false. Spending was on plan.
• A jump in the click rate came mostly from a change in where the ads ran, not from the ads themselves.
• A claim about a competitor held true on only one platform.

It isn't airtight, and it doesn't replace your judgment. I still read most of what ships. But it catches the kind of mistake that costs you a client's trust. There's more in the blind review guide.

One project, start to finish

Here's how the three habits looked on a small job this fall: a first Facebook and Instagram campaign for a small local real estate brokerage.

Access. The AI worked from the transcript of our kickoff call, the photos the owner shared, the brokerage's own website, and the actual advertising rules: federal fair housing law, the state real estate commission's rule on ads, and Meta's rules for housing ads. It read the rules themselves, not someone's summary of them, and that changed the plan early: housing and job ads can't be aimed by age or gender, so part of the targeting in the original brief was off the table before anyone built it.

Write it down. Everything went into one project file: who's who, the budget, what's decided, what's still open, and a dated log. Three days later a fresh session picked it up cold and drafted 20 ads.

Blind review. Before the client saw anything, seven simulated local readers, each a different kind of buyer, seller or agent, reacted to the 20 ads in random order without knowing what we were testing. Their reactions shaped the sets of ads we're testing first.

Seven simulated readers from one model are a sanity check, not a survey. The live test decides.

Start small

The question I get most is how to apply this to a whole process. My answer is older than these tools: don't try to boil the ocean. Find the part of the process that's easiest to automate, start there, and expand.

At the software company I work for, that part was support. When a customer reports a problem, someone has to find where in the code it lives before a developer can fix it. So we built an agent that checks the support inbox every half hour during business hours, reads the relevant code, and drafts a reply from notes it has built up on past tickets. When it writes a problem up for the developers, it has to list several possible causes, not one verdict. One confident answer that happens to fit is usually a guess.

It can read and draft, but it can't send anything to a customer. A person approves every reply. Why that split matters is its own guide: Don't let an LLM gate its own dangerous actions.

What it costs

The paid personal plans that unlock most of this run about $20 a month: Claude Pro, ChatGPT Plus or Google AI Pro. Plans for a team run roughly $20 to $30 per person per month and add admin controls. Prices change often, so check the current pages for Claude, ChatGPT and Gemini.

Try this week

1. 1Hand it one real file. If it needs more, ask: "How can we get you this data automatically?"
2. 2Before you stop: "How do I get you back up to speed as fast as possible next time?"
3. 3Before it goes out the door, open a new chat with the draft and the sources: "Check every claim in this against these sources."