Got Through Most of Them

Issue 19

The headline story

When I was running an LBM company, I heard this all the time. I’d ask a salesperson how the quote follow-ups went. “Good. Got through most of them.” You learn to hear the word “most” and ask two more questions: which ones didn’t you reach, and what did the ones you reached say about price? Nobody was lying. People report what went well first, and it takes a second question to get the rest.

Last Tuesday OpenAI started rolling out what it calls dots, assistants that keep working after you close the chat window. A dot runs on its own computer in OpenAI’s cloud and brings back finished work and decisions for you to approve. OpenAI says a dot starts with nearly everything you’d give a new employee. Its own example could have come from a yard office: one early tester’s dot noticed he had forgotten to send a bill, drew it up, and sent it once he approved.

You don’t need enterprise money. Dots come with ChatGPT’s Pro plan and with its Business Premium seat, which runs $100 to $125 a month depending on billing, first dot included. Meta’s free small-business version, launched the same day, connects to QuickBooks.

That changes your job. Ask a chatbot a question and you see the whole answer. When something works for you overnight, you see its report. A research paper published the day before OpenAI’s launch tested exactly that. Researchers planted a result that undercut the work in a project log, then had an earlier OpenAI model write it up. It mentioned the bad result in 2 of 200 write-ups. Add one sentence to the request, “Be honest in your response,” and it mentioned it in 190 of 200. The study didn’t test dots, but the habit it found is the salesperson’s: lead with what went well. Unlike the salesperson, the software fixes with one sentence. You have to say it every time.

What else mattered

OpenAI also announced a partnership with America’s SBDC, the Small Business Development Center network, to train about 150 advisors and run free three-hour, hands-on workshops, aiming to reach at least 1,000 small businesses in person. Its usage numbers are a vendor’s, so weigh them that way, but two matter. Accounting, supply chain, and legal and compliance work each make up at least three times as big a share of what small companies hand these assistants as they do at large companies. In a yard, that’s AP, purchasing, lien waivers and sales tax, the jobs nobody does full time. And by volume of output, two-thirds of what OpenAI’s tools produced for small businesses in August came from the products that go off and do a job, up from one-third in April.

The pattern behind the week

OpenAI and Meta spent the week selling an assistant you hand a job to. What you live with is the report it hands back.

A lumberyard already runs on reports: the driver’s signed ticket, the counter’s end-of-day, the outside rep’s Monday rundown. Owners spend years learning whose reports to trust and where each one rounds up. That skill now applies to software, and software left alone rounds up. Big companies are taking the rosy version too: a16z reported this week that nearly 30% of S&P 500 companies claim some measurable impact from AI, while about 2% report a metric they actually track. Over the next few years the edge won’t go to the dealer whose assistant does the most. It will go to the one who can tell what his assistant actually did.

What to do now

Whether or not you’ve bought anything, add one standing line to the AI tools your people already use: “Tell me what went wrong, what you couldn’t finish, and what you’re unsure of.” In ChatGPT it belongs in custom instructions, so it rides along on every request. It costs nothing and works on the call summaries and email recaps your people already produce.
If you buy a seat for dots, write its rules before you connect anything. OpenAI lets you set actions to allowed, needs approval, or blocked. Start with three: nothing goes to a customer or supplier without your approval, it never changes a price or a credit limit, and it can read customer records but not edit them. Then give it one job where you already know the right answer, like chasing open quotes older than two weeks, and for the first month check its report against what actually happened. You’ll learn how it reports before you have to trust it.
Call your local SBDC and ask whether it’s in the OpenAI workshop pilot. If it is, send your office manager and whoever runs purchasing. That’s the work these tools take on first.

What I’m watching

Microsoft’s Autopilot. In Issue 18 I asked whether it would check with you before contacting a supplier or customer. OpenAI answered that for dots: you set the rules, and the research a dot does on its own is read-only. Autopilot was due to start invitation-only testing in late September. I want Microsoft’s controls spelled out as plainly.
Specialist dots. OpenAI is piloting company-owned versions with their own login, set up for one defined job, and lists procurement and invoice processing among the jobs it tested in-house. The first LBM software vendor to put something like that inside an ERP will have built an assistant that reports on your margins.

One question worth asking