Because AI made saying yes to new work almost free at the moment you decided, and this is the article that explains what that actually cost you and how to get it back.
The Thing Nobody Priced In
Open the tool you use most and look at the list of things you have set up in it: workflows, agents, automations, scheduled reports, whatever your platform calls them. Give it ninety seconds.
Now count how many of those you read the output of last week.
For most of the business owners I work with, the honest number is between one and three. The list is longer than that, sometimes much longer. And every item on it worked when they built it. That is the part that stings.
Here is the direct answer. You have a pile of half-finished AI automations because AI collapsed the cost of saying yes to new work, and nothing collapsed the cost of owning it afterward. Each yes took ninety seconds and came with an obligation to check, maintain and trust it that you never priced in. The subscription is not what is expensive. Your attention is, and you have been spending it on checking instead of deciding.
Ben Tossell of the Ben's Bites newsletter put a word to this on November 17: "productiv-icky (adj.) – the feeling of doing work you didn't mean to do." I am building on his thinking throughout, and I want to credit him up front, because he named something I had been circling for months.
Candidly, this is the most expensive problem in small business AI right now, and almost nobody discusses it in cost terms. The industry just started to. This week Microsoft shipped spend dashboards for AI at Ignite, and new products launched on Product Hunt with per-action pricing and caps instead of subscriptions. Same admission from two directions: nobody knows what their AI is doing or costing.
Key Takeaways
Renders as a shaded callout box.
- Your AI does not primarily cost you money, it costs you yeses, and every yes carries hidden follow-on work.
- Half-finished automations are a pricing problem, not a discipline problem: the cost of starting dropped to near zero while the cost of owning stayed the same.
- Unfinished tasks measurably degrade performance on your next task, so an automation you stopped checking taxes work you have not begun.
- The market is now shipping meters, caps and spend dashboards, an industry admission that AI usage has outrun AI visibility.
- The fix is a written owner, a written check cadence, and willingness to kill what you will not own.
The Problem Is That Every One of Them Works
Let me be precise about the failure mode, because it is not the one people assume.
It is not that your automations broke. If they broke, you would notice. Something would stop arriving, a client would ask, a number would look wrong, and you would fix it. Broken is loud.
The failure mode is that they all still run, producing output into a folder you stopped opening three weeks ago. Working but unread is silent, and silent is what gets expensive.
Tossell describes the mechanism better than I can. A cold email from a stranger is easy to ignore: "Someone I don't know sends me emails that I'm expected to reply to. I don't know them. I don't work for them. So I ignore them… Nuh-uh, that's not my work, buddy. But it's different if it's my super-intelligent, helpful agent friend…" You can refuse a stranger. You cannot refuse something you built, because refusing it means admitting you wasted the build.
Then he puts the real price on it: "Yeses are easy. They're quick. But every yes has a yes-tax. The hidden extra work you didn't think of. The long tail of regret where you question why you ever said yes."
Read that again. The yes-tax.
[PLACEHOLDER, JONATHAN TO REPLACE WITH A REAL INCIDENT: a specific AI workflow you built, got working, and then abandoned or stopped checking. Include when you built it, what it was supposed to do, how long before you stopped looking at it, and what it cost you in attention or in a missed thing. Suggested shape: "In [month year] I built [workflow]. It worked. By [month] I had stopped reading its output, and I only found out when [consequence]."]
I tell you that so you hear the next part correctly. This is not a lazy person's problem. Lazy people do not build fourteen automations. This is the failure mode of ambitious, capable, curious people who see an opportunity and move on it. The exact trait that got you a business got you the pile. Notice that.
So stop treating this as a willpower issue and treat it as a pricing issue. The price of starting fell through the floor. The price of owning did not move. Everything that follows puts those two numbers back in the same room.
What the Evidence Actually Says
One. Your attention was already fragmented before AI showed up. Gloria Mark, professor emerita of informatics at UC Irvine, has measured how long people stay on a single screen for over twenty years. Her first paper, from a 2003 study, found an average of two and a half minutes. By 2020 it was 47 seconds. In a University of California interview: "most people today can't concentrate on a single screen for more than 47 seconds," while "recovering from an interruption can take almost half an hour." Forty seven seconds of focus. Thirty minutes to recover. That is the budget your automations draw against.
Two. An unfinished task damages the next task, not just the current one. Sophie Leroy's 2009 paper in Organizational Behavior and Human Decision Processes, "Why is it so hard to do my work? The challenge of attention residue when switching between work tasks," found that "it is difficult for people to transition their attention away from an unfinished task and their subsequent task performance suffers." Plain English: the half-built automation you are not thinking about is still taking a cut of the work you are doing right now. It does not need to be open to charge you.
Three. The unexpected AI bill is now the normal AI bill. Zylo's 2026 SaaS Management Index, published January 29, 2026 and built on more than 40 million SaaS licenses and $75 billion in spend, surveyed 218 IT leaders. In the previous twelve months, 78 percent reported unexpected charges tied to consumption-based or AI pricing models, and 61 percent were forced to cut projects because of unplanned SaaS cost increases. Sit with the 61 percent. Six in ten cut real projects because of software costs they did not see coming.
Four. The companies that fix this fix it by subtracting. Kameron Rezai, Chief Financial Officer at Verkada, wrote in TechRadar Pro in June 2025 about putting the company on a "software diet." They paid employees a $1,000 bonus for helping cancel a contract or switch to something cheaper, with rules: the contract had to be active at least six months and worth $10,000 or more annually. Since late 2024 they shed 12 tools entirely, replaced 4 with cheaper alternatives, saved $1.3 million, and paid out $36,000 in bonuses. They did not buy a tool to manage their tools. They paid people to kill tools and wrote down the rules.
And here is what confirms the diagnosis is industry-wide rather than personal. At Ignite on November 17, Microsoft announced that "Agent 365 turns AI spend into business value with FinOps for AI," because "AI spending is growing fast and moving beyond IT, and leaders need to see what it costs and delivers." On the same day, RankSpot relaunched on Product Hunt with no subscription at all: 3 dollars per article, 0.004 dollars per prompt check, and a spend cap you set yourself. Its co-founder publicly reversed his own automation promise, writing that he "truly believed SEO should be 100% automated… No human needed… And you gave us Product of the Day for it." The new pitch: "Type 'Run this week's fixes for every client' and approve what goes out."
Microsoft shipped you a meter. A founder who won Product of the Day for full automation shipped you an approval step. In twelve months, caps and approval gates will be standard rather than differentiators. The story said AI would get cheaper and more autonomous. What shipped is more supervised.
The Framework: One Hole, Dug Properly
Tossell's metaphor is the best thinking tool I have seen on this, so I am using it with credit. He writes: "So you grab a shovel and start digging. A little hole over there to triage your inbox, one here to draft your tweets, one to check on those. Then you spend your day checking the holes are being dug right. Before long, you're surrounded by holes and a mountain of dirt you gotta do something with."
Then the alternative, which I would put on a wall: "Now as a man, holes are hardwired into my DNA. A few beers, a beach and a shovel, I'm digging a hole. I'd take pride in it. I'd make the best one, impressively round, deep enough to see sea water seep through the bottom. But I'd only dig one. There's a whole beach but I chose this spot."
He also explains why we all over-dug: "We started measuring shovels instead of the hole." Token counts, seat counts, automation counts, tool counts. All shovel metrics. None tell you whether a hole got dug.
So here is the framework. Three rules, deliberately boring.
Rule one: every automation gets a written owner and a written check cadence, or it gets deleted. Not "I'll keep an eye on it." A name and a frequency, written where you will see it. If you cannot commit to a cadence in writing, you do not want the automation, you wanted the feeling of building it. Delete it. That is the decision you were avoiding.
Rule two: cap your work-in-progress at three. Three live automations you actively own. Not three per category. Three. A fourth requires retiring one. This is the rule people fight me on and the one that works, because it forces the yes to cost something at the moment of decision.
Rule three: measure the hole, not the shovel. For each automation, write one sentence naming the outcome it changes and how you would know. "Inbox triage" is a shovel. "I touch client email once a day instead of six times, and I can see that in my sent folder" is a hole. If you cannot write the second sentence, the automation is decoration.
A word on token FOMO, because it drives more of this than people admit. Tossell again: "But now we feel like we should use all our tokens before they run out. Why? Because everyone else is." Your plan is not a quota to hit. Unused capacity is not waste.
One last signal that this is what people actually want. On Ballet's Product Hunt launch this week, a user named Surya asked for one thing: "weekly client report for me. takes an hour every monday. does it work with google sheets??" Nobody asked for autonomy. They asked for one hole, dug properly.
Seven Steps to Dig Out This Week
1. Inventory everything in one sitting, before judging it. Open every AI tool you pay for and list every automation, agent, workflow and scheduled job. Do not evaluate, just list. Twenty minutes, one document. The list being longer than you expected is the point.
2. Mark each one with the last date you actually read its output. Not the last date it ran. The last date you read it. If you cannot remember, write "unknown," which counts as a no. This one column does most of the audit.
3. Delete everything in the "unknown" column today. Not pause. Delete. A paused automation still occupies attention because you keep wondering about it. If you need it later you can rebuild it in ninety seconds, which is both the problem that got you here and the thing that makes deleting safe.
4. Pick your three and write the owner and cadence beside each. From what survives, choose at most three to actively own, with your name and a frequency next to each. Put that document where you will physically see it. Gloria Mark's advice on goals applies: keep it in your field of view, because goals slip out of mind fast.
5. Find the real number you spend on AI each month. Pull twelve months of card and bank statements and highlight every recurring AI or software charge, including personal cards. Give AI its own budget line. If your tools offer spend caps, turn them on and set them below what you think you need.
6. Run a one-week kill test on anything you are unsure about. Turn it off for seven days, write down what you expect to break, then compare. Most of the time nothing breaks, and that answer beats any amount of deliberating. Same logic Verkada used: make subtraction a process with rules, not a mood.
7. Use a structured prompt to pressure-test each survivor. Do not ask AI for a generic audit, it will flatter you. Use this, filling in the brackets:
[The Job]
Pressure-test one AI automation I currently run and tell me honestly whether I should keep it, simplify it, or kill it.
[The Background]
I run [BUSINESS TYPE] with [NUMBER] people. The automation is called [NAME]. It was built to [ORIGINAL PURPOSE]. It currently [WHAT IT ACTUALLY PRODUCES AND HOW OFTEN]. The last time I read its output was [DATE OR "UNKNOWN"]. If it quietly produced wrong output for a month, the consequence would be [CONSEQUENCE]. I check it [FREQUENCY OR "I DO NOT"].
[The Deliverable]
A one-page verdict with: keep, simplify or kill plus the single reason; one sentence describing the outcome this automation changes and how I would know; the smallest check I could run in under five minutes a week; and the failure mode most likely to go unnoticed. Written from the perspective of an operations advisor speaking to a non-technical business owner who is already overcommitted.
[The Questions]
Ask me any questions you have.
Frequently Asked Questions
How do I find out what I am actually spending on AI every month?
Pull twelve months of card and bank statements, including personal cards, and highlight every recurring AI or software charge. Give AI its own budget line instead of lumping it into software, then turn on any spend caps your tools offer. Zylo's 2026 index found 78 percent of IT leaders hit unexpected AI or consumption charges last year, so assume your number is higher than you think.
Should I cancel AI tools I am not using, or will I need them later?
Cancel them. Rebuilding an AI workflow now takes minutes, which is precisely why you accumulated them. Keeping a tool "in case" means paying a subscription and an attention tax for an option you can recreate cheaply. Verkada shed 12 tools and saved $1.3 million applying that logic.
How do I know if an automation is working if I stopped checking it?
You do not, and that is the answer. Working but unread is functionally identical to broken, except it costs more because you still trust it. Run a one-week kill test: turn it off, write down what you expect to break, then compare. If nothing breaks, you have your decision.
Is per-action pricing cheaper than a subscription for a small business?
Not necessarily cheaper, but far more visible. Metered pricing with a self-set cap, which several products launched with this week, ties cost to actual use instead of a flat fee you forget. That visibility is the benefit. If your usage is steady and heavy, a subscription may still win.
How many AI automations should a small business actually run?
Three you genuinely own, as a working cap. Not three per category. Three total, each with a named owner and a written check cadence, and a fourth only ever replaces an existing one. The constraint is the point: it restores a real cost to saying yes, which is what AI removed.
Do You Really Want Another Hole?
Go back to the beach.
You have a whole beach, a shovel and an afternoon. The version of you that gets the good result does not dig nine holes. That version picks one spot, commits, and digs until seawater seeps through the bottom. Then stands there, filthy and satisfied, looking at one impressively round hole.
The version of you that got buried dug nine, checked all nine, finished none, and ended the day in a mountain of dirt wondering why a productive day felt so bad.
You are not behind and you are not undisciplined. You are an ambitious person handed a shovel that costs nothing to swing, and you swung it, because that is the same instinct that built your business. So do the ninety second count. Delete what you have not read. Keep three. Write down who owns them and when you will look.
Tossell ends by asking: "Next time the needy little nerd pings you: do you really want another hole?" Then he adds a joke that is actually a warning: "Also, don't dig big holes on the beach. They collapse."
Dig one hole. Dig it deep enough to see water. Then put the shovel down and go home.
Ready to stop managing shovels? This audit is the entry point to what we teach inside the White Beard Strategies membership: AI systems you actually own, built around decisions rather than tool counts. Grab the training replays to work through the full audit and the Perfect Prompt Framework 2.0, or join the membership and bring your tool list to a live session.
About the author
Jonathan Mast is the founder of White Beard Strategies, an AI coaching and mentorship company for entrepreneurs, and the creator of the Perfect Prompt Framework 2.0. He teaches non-technical business owners to build AI systems they can actually maintain, speaks on practical AI implementation, and spends an unreasonable amount of time deleting his own automations so he can tell you honestly which three are worth keeping.