Chapters
  1. 0:00The questions this episode answers
  2. 0:43Who is Michael Fauscette
  3. 1:09Why companies buy software and stop using it
  4. 2:01The buyer is far from the user
  5. 2:58Paying for software nobody uses
  6. 3:31Trust is fragile: one bad AI answer
  7. 4:28What AI pilots that reach production have in common
  8. 6:46Pilots fail people first, data second
  9. 7:48Does the owner know he is the problem?
  10. 8:53Train the managers first
  11. 9:42Automating a broken process
  12. 11:30The New York City AI chatbot
  13. 12:48A $10,000 AI budget: what goes first
  14. 14:36Before you hire an AI builder
  15. 15:513 questions to ask on the AI demo call
  16. 17:47The AI tools a research analyst actually uses
  17. 21:04Just because you can does not mean you should
  18. 22:01Where to connect with Michael

You paid for an AI tool. Your team used it for a month. Then everyone went back to the old way.

I asked someone who has watched this happen for years. Michael Fauscette ran enterprise software research at IDC for almost ten years. Then he was Chief Research Officer at G2. Now he runs Arion Research.

Guest card for Michael Fauscette showing his G2 and IDC roles, cropped from his LinkedIn profile

I gave him one question. You have $10,000 for AI. Where does it go?

His first answer: spend nothing.

Why do AI projects fail?

Short answer: the people around the tool fail it first. The tool fits nobody's day. The team does not trust it. And the owner never uses it himself.

Michael put it plainly:

pilots fail people first, data second, technology's, uh, almost the third.
Watch at 6:46

The buyer is far from the user

The person who buys the tool is rarely the person who uses it. The user gets handed something new. It does not fit how they work.

It just frankly makes their work harder, not easier.
Watch at 2:01
A package travels from a block labelled The buyer to a block labelled The user, and the user block turns grey.
3D scene with the buyer and the user standing far apart, from the episode at 2:15

The software you already pay for

This gap is old. It started long before AI.

One bad answer and it is over

AI adds one more problem. The output changes from run to run. So your team waits for it to slip.

trust is fragile, so one bad answer, one bad outcome and, and, and the team doesn't trust the tool anymore.
Watch at 3:31

And the tool itself is often fine.

'Cause often the technology works just fine, the people don't.

Weeks 1 and 2: spend nothing

This is the part I did not expect.

I would spend nothing at first, and in the first two weeks, spend that time documenting the processes that take up the most time
Watch at 12:48

I picked the $10,000 for the question. It is a what-if, not our price.

Do this Monday: Write down the 5 tasks that eat the most hours. Who does each one. How often. How long it takes.

That list costs you nothing. It is also the first thing any good builder will ask for.

If you want a second pair of eyes, see what to automate first in a growing business.

Wrote your list?

Send us the task that eats the most hours

Your first build: internal, repetitive, reversible

Now you spend. But only on one kind of task.

Automate that, you know, internal, repetitive, reversible piece of it. Refuse anything that's customer-facing first, with money attached especially, and, and anything that's custom-built.
So look for off-the-shelf

Three tests for your first task:

  • Internal. Your team sees it. Customers do not.
  • Repetitive. It happens every day or every week.
  • Reversible. A mistake takes one click to undo.

I told him on the call:

Yeah, reversibility is a very important factor.
Two trays from the episode. Automate: internal, repetitive, reversible. Refuse: customer-facing, money attached, custom-built. Caption reads look for off-the-shelf.

What happens when you skip that rule

Michael's warning was about refunds. Tell an AI agent to "make the customer happy" and it will.

Well, that's not an outcome, right?
I could keep your customers really happy by refunding everybody who complains at, at, at machine speed.
Watch at 9:42
Refund tokens leave a box labelled The AI, faster and faster, through an open gate labelled No approval step.

Do this Monday: Take your list of 5 tasks. Cross out anything that touches customers or money. Pick your first build from what is left.

How we do it: One of our builds is a bookkeeping agent. You send a photo of a receipt into a chat. It saves the file in Google Drive. It adds a row to a Google Sheet. It sends back a tick. A wrong row? You delete it. Once it is set up, it runs on your account.

Before you spend, put a number on it.

New York City Mayor's Office release, October 2023: the city launched an AI chatbot to help business owners.
The Markup and THE CITY, March 2024: their test of the NYC business chatbot, with answers that go against the law.

The owner uses it first

This one is on you, the owner.

the owner has the power to keep doing it the old way and no one will push back.
Watch at 6:46

When it works, he said:

the owner picks a task they do personally, they show that it works, they build that trust
3D staircase of the first month. A small task on the first step, bigger projects on the steps above.

Do this Monday: Pick one task you do yourself. Use the tool on it for a week. Do it where your team can see.

How we do it: Our own AI agents write drafts. Nothing goes out until I read it. I use them every day. So the team does too.

Build trust slowly

Do not hand the tool the keys on day one. Start with a person checking everything. Then step back, bit by bit.

I'm gonna step back, and I'm gonna put humans in the lead versus human in the loop.
Watch at 14:36
A dot moves along a path through five stops: Human in the loop, Build trust, Step back, Humans in the lead, Broader deployment. Each stop lights up green.
Path from human in the loop to humans in the lead, from the episode at 15:30

Do this Monday: For the first month, a person checks every output. Count the mistakes each week. When the count stays low, check less.

3 questions to ask before you hire an AI builder

Question 2 of 3 before hiring an AI builder: what happens when you leave, who changes a rule, what maintenance costs, what breaks.

The demo is the easy part. Month four is where it breaks. The builder is gone and nobody can fix it. These three questions catch that early.

How we answer them: Ask us all three on the first call. Hold us to the answers.

Ask us the three questions

Just because you can

Michael uses AI agents for most of his research. He also drops tools fast.

I have fired more note takers than I've used, to be honest.
Watch at 17:47
The Docs Bot chatbot on the Arion Research site, one of the AI tools Michael kept

Then the line I keep coming back to:

just because you can doesn't mean you should.
Watch at 21:04
don't be afraid to walk away from something when it's not working

FAQ

Why do AI projects fail in small businesses?

Michael's answer: people first, data second, technology almost third. The tool does not fit the team's day, one bad answer kills trust, and the owner does not use it himself.

What should a small business automate first with AI?

A task that is internal, repetitive and easy to undo. Leave anything that faces customers or moves money for later.

What should I ask an AI agency before I hire them?

Ask for the process map before the proposal. Ask who owns the prompts, accounts and logins when they leave. Ask which number you will check in 90 days, and what happens if it does not move.

Can I start with AI for less than $10,000?

Yes. Weeks 1 and 2 cost nothing but your time. You write down the tasks that take the most hours, then pick one.

Watch the full conversation

All 22 minutes are on YouTube. More episodes are on The Awais Rafeeq Show. Next read: how to get ChatGPT to recommend your business.

Michael's work is at Arion Research. He is on LinkedIn.

Tell us the one task you would hand off first

About the author. Awais Rafeeq is the founder of AI Data House and the host of The Awais Rafeeq Show. His team builds AI agents and automations for businesses.