November 17, 2026
The hard size for AI is the middle
Issue 10 · November 17, 2026 · 5 min read
The hard size for AI is the middle
Issue 10 · November 17, 2026 · 5 min read

Last issue I told owners who say they can build it themselves that they're right. I said the line was about who answers for the thing once other people depend on it, and that it wasn't about how many people work there. I'd take that last part back. The size of the company decides how much there is to answer for.
At a small company, the owner who builds the tool is usually the person who uses it and the person who decides who sees what. HummingAgent still runs that way on the inside. I also spent 13 years at Comcast Labs building tools that 40,000 technicians used, and a company that size can put a team on a problem and keep it there.
Most of the leaders I talk to now run something in between, a hundred people or five hundred. They have a lot of what makes this hard at a large company and they don't have a team to spare for it. I think that's the hardest size to be with AI right now.
FROM THE FIELD
I've done this at both ends
Inside HummingAgent we moved off software we used to subscribe to and onto tools we built ourselves. Since this fall our proposals go out of a system we wrote.
That worked for reasons that have little to do with how well we build. The people who use those tools are the people who decided how they should work. There's one way we write a proposal, because I decided it. I know who can see the books because I set it up. When something breaks, I'm the one they tell. A small company starts with all of that already settled, and it's easy not to notice.
At Comcast none of that was settled in advance. I built one of those tools myself, and it ended up in the hands of 40,000 technicians. Writing it was one person's work. The hard part was getting it in front of people who had a job to do, inside systems the company already owned, and keeping it alive once it was theirs. Comcast had people for that.
THE PATTERN
What changes between ten people and three hundred
Take a company of three hundred. It isn't Comcast, but it usually has a shorter version of the same list. It runs on several systems that were bought years apart and were never connected. The same job gets done a little differently in each office, because each office manager worked out a way that fit their customers. There are pay records and customer files that some employees should see and others shouldn't. And the people who know the business well enough to own a new tool already have a full day.
In a good company, experienced people handle all of this without thinking about it. An office manager who has been there for years knows which system to believe when two of them disagree, and which numbers aren't theirs to read out.
An AI tool has none of that judgment unless somebody builds it in. It works from the systems it was connected to and the instructions it was given. If it's shown one office's way of doing the job, that's the way it will do the job in every office until someone tells it otherwise. If it can reach a file the person asking shouldn't see, nothing in the model stops it from using that file. The limit has to be built, and somebody has to decide what it is. When it's wrong, somebody has to already be the person who hears about it.
That's where I'd expect a build that works fine at ten people to get stuck at three hundred. It's also why a license for every employee, by itself, doesn't finish the job. An assistant for each person is useful. It doesn't settle which office is doing the job the right way, or who hears about it when it's wrong.
A ten person shop can rent the same models a large company can. What's different at three hundred is that there's more to do around the model and no team free to do it. That work is most of what we do at HummingAgent. The tool is the smaller part. Most of the time goes to connecting the systems the company already has and getting the people who do the job to agree on one way to do it. Before we're done, somebody inside the company is the person who hears about it when it's wrong.
TRY THIS
Have two people show you the same job
Pick the AI project you're most likely to pay for next year. Before anyone prices it, find two people who do that job today in different offices or on different shifts. Sit with each of them separately and have them show you how they do it, start to finish.
While you watch, count the places they log in. Before you leave, ask who they'd call if the tool gave a customer a wrong answer.
If both of them do the job the same way and give you the same name, that's a project I'd expect your own people to handle, and what I wrote last issue applies to you. If the login count was higher than you expected, that's the connecting work from the section above, and it belongs in the plan before the tool does. If you watched two different jobs, or neither of them could tell you who they'd call, the tool isn't the first thing to buy. That project needs an owner, and it needs someone to decide which way of doing the job is the right one.
ONE ASK
If you run a company that size
If AI is on your list for next year, I'd like to hear which project you're looking at. Book time with me and we'll run the check above on it together, and I'll tell you where I think it gets stuck.
Until next time,
Shawn Kercher
CEO, HummingAgent AI
P.S. The office manager in this issue isn't anyone in particular. If you read that paragraph and thought of someone at your company, that's the first person to sit with.
Get the next issue in your inbox
Practical AI automation ideas for business owners. No hype.