These are notes from work in progress: what building AI inside real businesses actually looks like, written as it happens. No client names - descriptors only. Most notes are posted on Michael's LinkedIn first, and client details are kept general here.

October 5, 2026

Build or buy? Six questions, with our verdicts

  1. Does a mature product already cover this? Buy if yes and the references hold up.
  2. Where does the time actually go? Build if the expensive step is the one products treat as an edge case.
  3. Will it connect to the systems you actually run? Build if the connector does not exist, or exists only as a roadmap item.
  4. What does three years cost, both ways? It depends on volume. Run the numbers before the demo.
  5. Is this close to what makes you good? Build if it is close to your edge. Buy if it is plumbing that happens to be annoying.
  6. Who owns it in eighteen months? Buy if you cannot name that person.

For accounts payable, the same question in a table:

When to buy AP automation software and when to build custom AI
If this is trueBuy a productBuild custom
Invoice formatsMostly consistent, from a stable supplier listA long tail of irregular, scanned, or handwritten documents
Your ERPA mainstream system with a supported connectorOlder, industry-specific, or heavily customized, with no connector
Where the time goesKeying header data: supplier, date, totalLine-item matching, coding judgment, chasing exceptions
Volume and pricingPer-invoice or per-seat cost is comfortable at your volumeSubscription over three years exceeds building it outright
What is special about youYour AP process looks like everyone else'sThe way you do it differently is the thing that needs automating
OwnershipYou are comfortable renting the capabilityYou want it running in your own accounts, on your own rules
October 5, 2026

Our answers, in their words

Our website used to say two things on almost every page: you own the code, and a fixed fee.

While rebuilding it, we went back through our notes from the companies we've talked with since May. None of them said "own the code," and none asked for a fixed fee. Almost every one asked what it would cost, and most asked what happens after we hand it off.

Both lines are still true. They were our answers to those questions, in our words. The new homepage now says it closer to how they asked: one working piece at a time, proven on your own data. Your team runs it without us.

If you've bought AI work this year, what did you ask after the price?

October 5, 2026

Who was really visiting

Before deciding which pages to keep on our new website, we took a close look at who was actually visiting the old one.

About 4 in 10 visits came from towns where the big cloud companies run data centers, or from no town at all. Those visits averaged 0 to 3 seconds. Visits from Madison or Chicago averaged 30 seconds to over a minute.

Most likely, the short visits were email security systems opening links in emails we sent. They accounted for almost half the views of our work page, and they opened our AI chat too.

So we judged every page using the numbers without them, filtering out those towns. Has anyone found a cleaner way to tell people from scanners?

September 21, 2026

Who decides what the company learned

Last Friday I was at the Agentic AI Forum, put on by Matthew Seitz, the AI Hub for Business at the Wisconsin School of Business.

Matt Seitz and his team brought more than a dozen speakers to the Pyle Center, from Hyatt, Epic, Abbott, Mattress Firm, 3M and Intuit to UW faculty, around one practical question: what are AI agents actually, and how can I use them? It was the Hub’s second Ground Truth event this year, and the work behind it showed.

One lightning talk stuck with me, from Siddhartha Agarwal, CEO of JazzX AI. His example was simple. When a loan underwriter disagrees with the AI, JazzX records the correction. Across enough loans, patterns start to show up. But a pattern is not a policy. Someone at the company still has to decide which corrections should apply to every loan going forward. Otherwise, individual habits can quietly become company policy.

At companies we work with, experienced people are correcting AI every day. We have found those corrections usually fall into three buckets. Some are one-offs, so nothing changes. Some become rules, written in plain language with who requested them and where they came from. One client’s rule book has grown to 74 rules, 23 of them coming directly from corrections made to invoices and other documents. And some corrections keep coming back because the problem is not the AI at all. If someone keeps fixing an address that already exists in the company’s own database, another AI rule will not solve it. The work itself has to change.

That third bucket is where Siddhartha ended: the companies that get the most from agents will not be the ones that deploy the most. They will be the ones that redesign the work around what agents make possible.

Joan Bottarini, Hyatt’s CFO, made a similar point from a larger-company perspective. Most people have figured out how to use AI to improve their own work. Changing how the company operates is the harder step, and it takes leadership time.

The tool may be the easier part. Every time the AI learns something new, someone still has to decide: is this a one-off, a new rule, or a sign the work itself has to change?

Who gets to decide whether the company learned it too?

Thank you, Matthew Seitz and the AI Hub team.

September 1, 2026

This call is recorded

We built a phone assistant that answers calls for a business.

A few weeks ago, we added one line to the first thing she says: this call is recorded.

It sounded simple. It took three tries.

The first version used the familiar phrase, “for quality and training purposes.” We cut it. “Training” means something different now. In 2026, a caller could reasonably hear that and think their conversation is being used to train an AI model. That is not what happens.

So we gave the real reason instead: the call is recorded so she can get the caller’s details right. Then we found a different problem. The settings were correct, but when we actually called the line, the word “recorded” was missing. Everything else came through. The one word that mattered most did not. It was buried in a longer sentence. We gave it its own short sentence, called again, and this time it came through clearly.

That was a useful reminder for me. Reading the settings tells you what the system is supposed to say. Calling the number tells you what the customer actually hears.

She also says she is an AI assistant in that same opening. That part was easy for us. It takes about two seconds, and the caller knows who they are talking to from the start.

I wonder how others are handling disclosure with voice AI right now. There still seems to be a lot of variation in what gets disclosed and when.

August 27, 2026

The 5 questions IT keeps asking

Sooner or later, every AI project meets the IT team. The meetings I've been in this year keep coming back to the same 5 questions:

  • Where does it run?
  • Does anything leave our walls?
  • Does our data train the model?
  • Who supports it when it breaks?
  • Is this just a wrapper around a big AI model?

These are the right questions, and the meeting goes well when the answers are plain.

It moves into your cloud, on your contracts, behind your sign-on.

Your data trains nobody's model.

Your IT team owns the hosting and the infrastructure, and we own the logic and the tuning. If the tool is ever down, your team does it the old way that day. Nobody gets paged at 3 AM.

The wrapper question deserves a straight answer too. The AI model is the engine, but the value is what gets built around it. Your rules and your checks, running on your data. A competitor can't buy that off a shelf.

The projects that go well seem to be the ones where IT is in the conversation early. They own it once it's running, and the good ones ask exactly these questions.

If you're on the IT side of these meetings, does this sound right? What else do you ask?

Update, October 2026: Today the answer to who supports it is your team, without us. The handover is part of the work.

August 20, 2026

Eight brands, one marketing team

I've spent a good part of this year working with the marketing team at a national catalog retailer on a problem I hadn't really thought much about before.

They have eight brands. One marketing team. And every campaign still needs to sound like it came from the right brand.

We've been building an AI brand agent at WiscAI that learns each brand from the company's own catalogs and past emails. One thing that surprised me was how useful the older material was. In a few cases, going back to the years when the brand voice was strongest gave us much better material to work from than the newer stuff.

The team can give it a short campaign brief and get back drafts for email, display ads, social and SMS. People still make the decisions and approve the work. AI just helps get a lot of the first-pass production out of the way. The other thing I confirmed is that writing wasn't really the biggest problem. Review was. We've spent a lot of time teaching the agent to check its own work against each brand's rules before it gets to a person. That has probably been one of the more useful parts of the whole project.

Still learning a lot as we build it.

I'm curious how other companies with multiple brands handle this. How do you keep each one sounding like itself when the same team is working across all of them?

August 11, 2026

The retiring expert

At a Wisconsin manufacturer we work with, the billing rules live in the heads of a few long-tenured people. A lot of what they know isn't written down anywhere. Which surcharge applies. How packaging bills differently when an order drop-ships. Which freight lines stay off the invoice. Dozens of rules like these, learned one job at a time. When those people retire, the rules go with them.

Since early this year we've been building AI that does the invoicing with the billing team. When the tool gets an invoice wrong, they correct it. We figure out why, write down the rule, and the tool knows how to handle it the next time. We started out automating invoices. Along the way, we've also been writing down decades of knowledge the company never had written down before.

We're hearing versions of this at other companies too. A food manufacturer recently told us about a long-tenured employee. Their concern is that if that person is ever out for an extended period, nobody is really ready to step in and do what they do.

I'm starting to think this is one of the more practical uses of AI for a lot of established businesses: capture what your most experienced people know while they're still there.

Curious how many other companies have started to deal with the same problem.