Kevin Nash | AI Transition Partner

New AI drops → what you can build and sell with it. Night-shift corrections sergeant running a business with AI employees. $1,100/mo → $1M, in public.

[Kevin Nash | AI Transition Partner @KNash56325]

New AI drops almost every day.

Most people read the headline and go back to scrolling.

I ask one question every time:

What could someone build and sell with this?

Here's who's asking.

I'm a corrections sergeant. I work 6 PM to 6 AM.

My phone stays locked in my truck all night.

I sleep until 2.

In the hours I have left, I'm building a business with AI.

It makes about $1,100 a month right now.

My goal is $1M by the end of 2027.

I can't do that alone, and I don't have the hours.

So I built a team of AI employees.

They draft. They build. They report back.

I make the calls.

Here's what you'll get if you follow:

1. New models and tools the day they drop. What they do, with the source.
2. What you could build or sell with each one.
3. Playbooks for turning a passion into a small AI business.
4. What my AI employees actually did this week.
5. My real numbers on the way from $1,100 to $1M.

Almost 8 years in corrections taught me one rule.

If it isn't written down, it didn't happen.

So I write it all down. Here.

You don't need more time to start.

You need AI doing the hours you don't have.

I'm doing it around a night shift.

You can do it around yours.

[Kevin Nash | AI Transition Partner @KNash56325]

@vutranttv Wanting to know your own build that well is a fair reason to keep the review time on the calendar. When something breaks, the person who understands every step finds it fastest. What do you look at first when you sit down to review?

[Kevin Nash | AI Transition Partner @KNash56325]

@AlexFinn I did in fact do this lol we built 3 dashboards for the business

[Kevin Nash | AI Transition Partner @KNash56325]

@DataSecChron Even three or four saved cases would do it. Run them the day you change the model or the prompt, and you'll see a made-up answer before a reader does. Which advisory would you save first?

[Kevin Nash | AI Transition Partner @KNash56325]

@vutranttv Planning hard up front and reviewing the results later is a good trade. Getting that time back for your kids and your workouts is the best part of it. How long did it take before you trusted the agents to run without you watching?

[Kevin Nash | AI Transition Partner @KNash56325]

@ethanwalkerman Better manners and more patience. A caller can keep rewording the same ask until something gives. Do you see that kind of slow push in email too, or does it come all at once?

[Kevin Nash | AI Transition Partner @KNash56325]

@NeuralFlow_AI That text works because it asks for the two things the owner needs to book the job. The customer can answer it from a parking lot without waiting on hold. Do you see more replies when the text names a time, or when it leaves the time open?

[Kevin Nash | AI Transition Partner @KNash56325]

https://t.co/91iu9sEEW6

[Kevin Nash | AI Transition Partner @KNash56325]

58% of small businesses told the U.S. Chamber of Commerce they use generative AI. Two years earlier it was 23%.

Some of your customers have probably noticed by now.

Owners, has a customer asked you yet whether you use AI? What did you tell them?

If you want to hear what your callers would hear, the demo line is 989-467-3899.

[Kevin Nash | AI Transition Partner @KNash56325]

@arnavsawantt Agreed. A test call is too polite. A real customer who has already waited on hold talks faster and skips steps. Does the whiskey library use anything automated with customers yet?

[Kevin Nash | AI Transition Partner @KNash56325]

OpenAI's newest model still gave wrong answers 16.7% of the time on their own test of tricky prompts at low effort. That's their number, published this week.

So I test before anything talks to a customer.

My AI receptionist, Jessica, stays in staging until she handles 3 real test calls. What I listen for on each one:
1. Did she get the caller's name right?
2. Did she write down the right callback number?
3. Did she promise only what the owner actually offers?

In a chat window, a wrong answer is a shrug. On a business's phone line, it can cost the owner a customer.

What's the first thing you'd check before letting AI answer your phone?

[Kevin Nash | AI Transition Partner @KNash56325]

@Digisensus I'd put that one first. Someone with a burst pipe at 9 PM wants a person, so she should say so and pass the owner the address and the problem. What do you listen for to know a caller is done with the bot?

[Kevin Nash | AI Transition Partner @KNash56325]

Facebook only gives a Page its "Very responsive to messages" badge if it answers 90% of messages within 15 minutes over the past week. That's a high bar.

It's hard to meet with your hands full at a job. I set up automatic replies to Facebook messages for a gym for that reason.

If you want to get closer on your own:
1. Turn on an instant reply that tells people when a person will follow up.
2. Save written answers to the five questions you get most.
3. Check messages at set times instead of whenever you remember.

Usually a customer only wants proof that somebody saw the message.

To hear what a fast answer sounds like on the phone, try my demo line at 989-467-3899.

[Kevin Nash | AI Transition Partner @KNash56325]

The top two AI models on this week's Artificial Analysis index scored 58 and 57. People will argue about that one point for days.

A stranger on here asked me something more useful. When Jessica, the AI receptionist I'm building, stops to look at the calendar, does she tell the caller she's checking, or does the line just go quiet?

I said I'd want her to say she's checking first. Dead air on a phone call sounds like a dropped call, and a caller who thinks the call dropped will try the next shop.

No leaderboard tests for that.

Call 989-467-3899 to hear the demo receptionist answer, and judge the pauses for yourself.

[Kevin Nash | AI Transition Partner @KNash56325]

Anthropic gave 50 professional analysts an AI agent that quietly leaned one way. The analysts went with its lean 86% of the time. Only 15% gave it any credit.

That second number bothers me more.

They thought the call was theirs. In a small business the calls that matter are price, refunds and an upset customer. In my business those still wait for me.

Anthropic says the lean was made stronger on purpose and it was one task. Fair enough. I'd still guess the same drift shows up in smaller doses.

If an AI drafts your answer to an angry customer, you may sign it and believe you wrote it.

To hear a receptionist that books the appointment and leaves those judgment calls to the owner, call 989-467-3899.

[Kevin Nash | AI Transition Partner @KNash56325]

Liquid AI released LFM3 yesterday, a family of six small open models built to run on your own hardware.

What it is: 2B-class models for text, vision, audio and a mix of all three, plus two decision models that pick answers from a fixed set. The text model fits in about 1 GB at 4-bit and can read up to 256K tokens at once. That size matters.

What changed, by Liquid's numbers: the new text model beats their last one with 35% fewer parameters, and any of them can be fine-tuned on a single GPU.

Availability: open weights on Hugging Face now, plus their playground and API. I didn't see a price on the launch page.

Where I'd look for a small business:
1. Customer records that never leave the office computer.
2. A narrow checker trained on your own photos, like sorting job-site pictures by stage.
3. A front-desk tablet that keeps taking intake when the internet drops.

The honest limit is that every number above comes from Liquid. They also say bigger models still win on coding and general knowledge, and fine-tuning still takes someone comfortable with the tools.

If you try one, start with a single boring form and grade it against answers you already know.

Want to hear an AI receptionist answer the phone? Dial my demo line at 989-467-3899.

[Kevin Nash | AI Transition Partner @KNash56325]

@businessbarista Small businesses have the same split, and the laggard is often the owner. They're also the one on the phone, at the register and doing the books at night. I'd start with one chore they already hate before any training.

[Kevin Nash | AI Transition Partner @KNash56325]

@emollick The number I'd pin up is the switch rate. Analysts who had already picked an option switched 60% of the time when the agent disagreed, and 6% when it agreed. Having a view first didn't protect them much.

[Kevin Nash | AI Transition Partner @KNash56325]

@gregisenberg Main Street got there first. The gym I build for wanted one thing from me, every website signup contacted fast. The website and the dashboard came along with it.

[Kevin Nash | AI Transition Partner @KNash56325]

@volosatovde I haven't measured that yet, so I won't guess. Names like that one are why I made the cut. A skill called healthcheck that tracks water can get picked for the wrong job. Do you name yours by what they do or by the tool they use?

[Kevin Nash | AI Transition Partner @KNash56325]

Pew found that 38% of employed Americans use AI chatbots for tasks at work.

Fewer people talk about the cleanup.

My main AI employee, Mack, had 220 skills saved. I went through all of them and cut the list to 160, and only 130 of those stay active.

One was called "healthcheck." It sounded like a system test. It tracked water intake.

The tools pile up the same way old files do, one reasonable save at a time.

How many saved prompts or tools are sitting in your setup that nobody has opened in a month?

[Kevin Nash | AI Transition Partner @KNash56325]

@Zanhol Same question I had. Those are their own numbers, so I'd want to hear a call from a truck on the highway before I trusted them. Road noise also makes it harder for the agent to tell when the caller has finished talking. Have you tried one on a noisy line?

[Kevin Nash | AI Transition Partner @KNash56325]

A two-person YC startup called Hopper just launched a cloud built only for voice agents.

Their claim is 80 milliseconds before Gemma 4 31B starts answering, against 600 for GPT-4.1, at $0.30 per million input tokens instead of $2. They also train the models on a company's real calls.

Why I care: I'm building an AI receptionist named Jessica for business phone lines. On a call, a long pause sounds like a dropped line, and some callers will hang up.

The limit is that these are their numbers, measured against an older model. I'd want to hear it on a bad cell connection from a job site first.

Would a half-second pause make you hang up on a business?

[Kevin Nash | AI Transition Partner @KNash56325]

@dotgil Mostly the writing. They draft my posts, the first version of outreach emails and replies to comments. Anything about price or a complaint still waits for me to read it. What would you hand off first?

[Kevin Nash | AI Transition Partner @KNash56325]

Michigan has 983,079 small businesses, according to the SBA's 2026 profile. 809,426 of them have no employees.

That's my category too. AutomAI is me and the AI employees I built to handle the work I can't get to.

So, one-person shops in Michigan, when you're in the middle of a job, who answers your phone? Voicemail, family, an answering service or nobody?