Finish One

I want you to count something before you finish your coffee.

How many AI projects have you started this year that aren't finished?

I don't mean the ones that failed. I mean the ones sitting at 60 percent. The custom GPT you built in March that nobody uses. The Make scenario that works except for that one edge case. The voice agent you set up on a Saturday and never connected to the calendar. The three subscriptions you signed up for after reading a newsletter (yes, possibly this one) and haven't opened since.

If your number is zero, you can go back to your weekend. Most people I ask land somewhere between four and nine. And they say it with the same face people make when they tell you how many unread emails they have.

You're in good company, which is the bad news.

Martech Weekly's Enterprise Martech Outlook for 2026 found that 95.3 percent of organizations include AI agents in their plans. Only 10.8 percent of agent initiatives have reached full scale production. And 40.2 percent of marketing technology leaders can't show a clear financial contribution from what they've spent. Two thirds of those organizations have a dedicated AI budget.

Read those together. Almost everybody is planning. About one project in ten gets finished. And four in ten of the people in charge can't point to the money.

A separate survey from WSJ Intelligence and the agency Code and Theory, covering 801 senior executives, found large US companies run an average of 183 AI tools. Eighty five percent described their AI systems as disconnected. About a quarter said nobody has clear responsibility for coordinating AI across departments.

One hundred eighty three tools. I can't name 183 of anything.

Then there's Accenture. According to an internal memo reported by 404 Media, starting October 1 the company will centrally fund $29 per employee per month for Claude. Some staff had been spending as much as $1,000 a month, mostly on agent coding work. Anything over the $29 now has to be covered by a specific client project budget. One employee called the change disappointing. Accenture said it's about giving people the right tool for the right work.

You can read the Accenture move as a cost cut. I read it as a company that looked at its AI spending and asked the question every business eventually asks: what did we actually finish with this money?

Big companies have their own reasons. Committees, procurement, security reviews. Small businesses don't have those excuses, and we still end up with a graveyard of half built things. In my experience it comes down to five causes.

There's no definition of done. "Set up an AI assistant for support" has no finish line, so it'll never be done. There's always one more thing to tweak, so it never ships, so it never pays.

Nobody owns it. The owner started it in a burst of inspiration, then got pulled back into running the business. Nobody else knows it exists or how it works.

There's no budget line. It runs on a credit card that also pays for six other tools. Nobody can say what it costs or what it returns, so nobody fights for it.

Too many at once. Five projects at 60 percent deliver exactly zero value. One project at 100 percent delivers all of it. Until it's finished, you're still doing the work by hand and paying for the tool on top.

The next shiny thing. Every week brings a new model, a new tool, a new feature. Each one is a reason to start something new instead of finishing something old. I'm aware of the irony of writing that in a daily AI newsletter. Read us for ideas. Don't start one new project per issue.

Here's the system I'd run for the next ninety days. It's simple enough that you'll actually do it.

Make a list of every AI project, tool and subscription you've touched this year. Be thorough. Check your credit card statements for the last six months. Search your email for "welcome to." Look at the connected apps in your Google or Microsoft account. Ask your team.

For each one, write down:

What it was supposed to do, in one sentence.

How finished it is, as a rough percentage.

What's blocking it, honestly. Usually it's one specific thing: an integration that didn't work, a decision nobody made, a test nobody ran.

What it costs per month.

What it would save or earn if it were done, in hours or dollars per month. Make your best guess, then cut it in half, because everyone overestimates.

Sort the list by what it would save or earn, divided by how much work is left. The project near the top, with high value and not much left to do, is your one.

It's rarely the most exciting project. It's usually something boring, like the follow up automation that's 80 percent done and just needs the last step connected. I'll take the boring one that's nearly done over the exciting one that's half built.

This is the hard part. Everything that isn't your one goes into one of three piles.

Kill. Cancel the subscription, archive the project, write one line about why. If you weren't using it, you won't miss it. If you're nervous, export whatever data's in it first.

Freeze. Valuable, but not now. Write down exactly where you left off and what the blocker is, so you can pick it up in two minutes later instead of two hours.

Keep running. Things that are finished and working. Leave them alone.

Most people find they can kill a third of the list on the spot. Those subscriptions add up. Put the savings toward finishing your one.

Before you touch the project, write what done means. Here's the checklist I use. A project is finished when all six are true:

All six. "It works on my laptop" doesn't count, and neither does "it worked in the demo."

Here's where Accenture's move is worth stealing, scaled down.

Give the project its own budget line. Its own card, or at least its own row in your books, covering every tool and model it uses. If you build in Make.com, give it its own team or folder so its usage is visible on its own. Now you can see exactly what it costs, which means you can see exactly what it returns.

Give people an allowance. Pick a monthly AI allowance per person for general use. Something like $20 to $50 depending on the role. Anything above that needs a reason tied to a specific project with a specific goal. The point is making sure the big money goes to things that actually get finished.

That's what I'd take from the Accenture memo. Spend on AI on purpose, and know what each dollar is for.

The fastest way to kill a finished project is to never prove it worked. Within a month, someone asks "is this thing even helping?" and nobody has an answer.

Measure before and after. For time savings, Rize tracks where hours go by app and project without anyone filling in a timesheet, so you can compare the month before the project went live with the month after. For revenue projects, track the one number from your definition of done, weekly, in a sheet that takes thirty seconds to update.

Then tell your team. "The follow up system saved us eleven hours last month and booked four extra appointments" does more for AI adoption in your business than any training session.

When the six checks pass, the project moves to "keep running." Go back to your list. Sort it again. Pick the next one.

One finished project a month is twelve a year. That puts you ahead of the roughly nine in ten big company agent projects that haven't reached full production. A small business that finishes twelve real automations in a year is a different business by December.

Picture a landscaping company with eight employees. The owner digs up the graveyard and finds seven items. A chatbot on the website that answers a few questions but can't book. An estimate generator that's 70 percent done. A review request automation that's 90 percent done but never got connected to the job completion step in the CRM. Two AI writing tools nobody's opened in four months. A voice agent trial that expired. And a custom GPT for writing job descriptions.

Kill: both writing tools and the expired voice trial. That's about $70 a month back.

Freeze: the chatbot and the estimate generator, with notes on exactly where each one stopped.

Keep running: the job description GPT, which works fine and costs nothing extra.

Pick one: the review request automation. It's 90 percent done, and more reviews mean more calls. The blocker turns out to be one trigger. When a job gets marked complete in Go High Level, it needs to start the review request workflow. Two hours to finish, including testing.

Two weeks later, all six checks pass. The office manager owns it. The off switch is documented. The number is reviews per month, and it goes from four to fourteen in the first month.

Next month: the estimate generator. It's suddenly a lot easier to justify, because the business just watched one AI project actually pay off.

Having the most tools doesn't put a company ahead with AI. Big US companies average 183 of them, and roughly nine in ten agent projects still haven't reached full scale.

Whoever finishes wins. That's always been true in business. AI just makes it obvious, because starting something new now takes about ten minutes.

So here's your Sunday assignment. Skip the new tools this week. Take one thing you already started and finish it.

Dig up the graveyard. Credit card statements, email search for "welcome to," connected apps, and ask your team.

Fill in the five columns for every item on the list.

Kill a third of it today. Cancel the subscriptions and export anything you need first.

Pick your one using value divided by work left.

Write the six check definition of done at the top of a doc and put a name next to it.

Give it its own budget line, then block two hours on your calendar this week to finish it.

Jordan

The AI Newsroom | Practical AI for people with a business to run.