Which Automation Is Worth Building, and Which Isn't?

Knowing what to build is rarer than knowing how. Four questions sort the two piles.

---

You have a list. Somewhere on a notepad or in your head sits the follow-up you keep forgetting, the quote that takes forty minutes to write, the order that comes in by phone and gets re-keyed twice, and a vague sense that some tool could handle part of it. The question is not whether automation works. It is which of those items deserves to be built, and which would eat a month and a check and then sit there.

This post is for the owner of a business with fewer than twenty people, in a trade, an office, or a family company, who wants a way to sort that list before spending anything. You will get the numbers on how often automation projects get abandoned and why, a four-question test I run on every build request before saying yes, the two piles most requests fall into, and one thing to do this week with a sheet of paper.

Why do so many automation projects get abandoned?

Because they were started for the wrong reason, aimed at a process nobody had mapped, or priced without counting upkeep. Gartner forecast in July 2024 that at least 30 percent of generative AI projects would be dropped after proof of concept by the end of 2025, and the stated causes were ordinary: poor data, weak controls, rising cost, unclear value.

Those four causes come from a July 2024 Gartner press release, and the wording is worth reading as an owner: projects would be abandoned "due to poor data quality, inadequate risk controls, escalating costs or unclear business value." Gartner's analyst added that executives were "impatient to see returns on GenAI investments, yet organizations are struggling to prove and realize value." The deployment costs in that release run from $5 million to $20 million, enterprise money, and that is the point. The forecast was not about small shops. It was about organizations spending that kind of money, with data teams and project managers, and it still expected nearly a third of their projects to be abandoned, for reasons that have nothing to do with the technology being bad.

A year later the same firm issued a separate forecast for a newer category, AI agents, and the number came in higher. In a June 2025 release Gartner predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls." The analyst's line that belongs on every owner's wall: "Many use cases positioned as agentic today don't require agentic implementations." The release also names "agent washing," vendors relabeling an ordinary chatbot or a scripted workflow as an agent, and estimates that only about 130 of the thousands of agentic AI vendors are real.

MIT's NANDA group put a measured number beside the forecasts. Its preliminary GenAI Divide report from July 2025, built on a review of more than 300 public AI initiatives, interviews at 52 organizations, and a survey of 153 senior leaders, reports that "95% of organizations are getting zero return" and that "just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact." On enterprise-grade systems, the report's own sentence is: "Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production." The authors call their figures "directionally accurate" rather than a census, the sample leans toward large organizations, and nothing in it estimates a small-business rate. Read it anyway, because the reason they give for the failures applies at any size: "Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations."

The table is not an argument against building. It is an argument against building the wrong thing. Most of the causes listed there trace to selection before software: unclear value, a workflow that did not match the day, a cost nobody counted. A test that asks those questions before the invoice is the cheapest piece of the project.

What does the picture look like for a business your size?

Most small businesses are not using AI yet, so the question in front of you is not catching up but choosing well. The Census Bureau's survey put AI use at 17 to 20 percent of U.S. employer businesses through spring 2026, and under 20 percent among firms with four or fewer employees.

The Census Bureau's May 2026 analysis of its Business Trends and Outlook Survey, which represents employer businesses excluding farms, covers December 14, 2025 to May 3, 2026. In that window "overall AI usage hovered between 17% and 20%," use rose among firms with at least 20 employees but "didn't change significantly among firms with fewer than 20 employees," and "less than 20% of firms with four or fewer employees reported using AI," against 37 percent of firms with at least 250. The Census notes the question was reworded in November 2025 to ask about AI in any business function, so these figures run higher than the older series. The SBA Office of Advocacy's 2026 small business FAQ, published in February 2026 and using the older, narrower question, puts AI use at 7.6 percent of employer businesses between September 2024 and August 2025, with 8.2 percent of firms with fewer than five employees and 11.4 percent of firms with more than 250, and notes the gap "has narrowed."

The two surveys ask different questions over different periods, so the percentages are not a trend line, but the shape is the same in both: the smallest firms sit at roughly the overall rate, well behind the largest firms, and the SBA notes the gap has narrowed. Both measures cover only businesses with payroll. The same FAQ counts 36,207,130 small businesses as of 2022, its latest data year, 82.3 percent with no employees at all and 17.7 percent, or 6,395,635 firms, with payroll, and nobody is measuring AI use among the 29.8 million without. If you run one of the 6.4 million with a crew, the company down the road has most likely not built anything yet. You are not behind a wave. You are early enough to pick.

And the one advantage a small firm has over the enterprises in the abandonment forecasts is that you can see your whole process from one chair.

What is the Scarcest-Resource Test?

It is four questions, asked in order, before any build starts. Call it the Scarcest-Resource Test, because the first question disqualifies most requests: whose hour does this give back? An automation is worth building when it returns time only the owner can spend, or catches value the business already earned and is losing, and it runs on a process that already works by hand.

The four questions:

1. Whose hour does it give back? Every business has one person whose attention is the real constraint. In a shop of six it is usually the owner, sometimes the one estimator or the one dispatcher. An hour saved for someone whose hours are not the constraint is pleasant, not urgent. The hour the owner spends re-keying orders, chasing a quote, or returning a call that came in at 6 p.m. is worth more than any other hour in the building. The second form of a yes is revenue you already paid to attract and are about to lose: the lead that went cold, the order that came in on voicemail and never got written down.

2. Has it worked on paper? Michael Hammer wrote in the July to August 1990 Harvard Business Review that companies "leave the existing processes intact and use computers simply to speed them up," and that "it is time to stop paving the cow paths." That was 36 years ago, in the mainframe era, and in my experience it is still the most common reason a small-business automation fails. If nobody can write the process on one page, with who does what and what happens when the answer is no, the software will encode the confusion and run it faster. The rule I use: the process has to have been done by hand, the same way, at least ten times before it gets built. Ten by hand finds the exceptions. Software finds them too, but at 2 a.m., in front of a customer.

3. How will you know it stopped? Everything built eventually stops: a password changes, a vendor updates a form, a phone number moves. The question is whether it stops loudly or quietly. A missed-call text-back that fails is loud: the owner notices the phone. A nightly report that quietly stops emailing is silent until the month it mattered. A build is worth doing when there is a plain way to see it stopped, a daily count, a weekly ledger, a person whose job includes glancing at it. If the only way to know it broke is a customer telling you, add the check before the build, or do not build.

4. What does it cost to keep? Not to build. To keep. Nearly every automation carries a standing cost: the five minutes a week to read what it did, the hour a quarter when something upstream changes, the vendor's bill. If that upkeep is less than the hour it gives back, build it. If it is close, your attention is better spent elsewhere. The reasons small business automations break within six months mostly come down to this question being skipped: someone counted the build and nobody counted the keeping.

Four yeses, build it. One no, fix the no first. Two or more, put it back on the notepad; it is not ready, and that is useful to know before the invoice, not after.

Which automations usually pass the test?

The dull ones. In the small businesses I see, the builds that pass all four questions sit at the edge of the business where a customer first reaches in: the missed call, the inquiry form, the phone order, the quote that went out and was never followed up. They pass because the process already exists, the failure is visible, and the hour saved belongs to the owner.

Take the phone order. At my parents' oil company in Florida, a 58-year-old family business, customers used to call in fuel orders that someone wrote down and someone else re-keyed. The system there now, a voice assistant named Emma, takes the order, structures it, and texts it to the team, with the phone and the website feeding one record and a person confirming anything unclear. Walk it through the test. Whose hour: the person who used to re-key, and the owner who used to field the call after hours. On paper: an order has the same seven fields whoever takes it, and the family had taken thousands by hand. How you know it stopped: the texts stop, which somebody notices by lunch. Cost to keep: a few minutes a week reading the log. Four yeses. The fuel distributor order post walks through that build by outcome.

The missed-call text-back passes for the same reasons, and it is the smallest useful automation I know of: a call goes unanswered, a text goes out within a minute or two when the number and consent rules are set up properly, and the customer answers the text instead of calling the next name on the list. How missed-call text-back works covers the mechanism. For a real estate office of one, the inquiry reply passes too. On my own site, a form that came in at 10:00 one August morning got a reply at 10:04 that included the one question I would have asked anyway, and the person was already answering it by the time I looked up.

The pattern across all three: one thing, at one doorway, that a human was already doing less reliably. None of them is impressive to look at. None of them got abandoned.

Which automations should you refuse, even when you can afford them?

The ones that save an hour nobody needed, replace the judgment the business is known for, or fail silently in front of customers. The tell is the request that starts with the tool instead of the task: "we should have an AI agent" rather than "the quote never gets followed up." Gartner's own advice for this stage is to pursue agents only where they deliver clear value, and to use agents for decisions, automation for routine work, and assistants for retrieval.

That order comes from the same June 2025 Gartner release: start "using AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval." For a business of eight, most of the value sits in the routine workflows, and most of those can be handled with rules and forms before anything called an agent is considered. A text that goes out when a call is missed is a rule. An order parsed into seven fields is a form. Calling either of those an agent is the agent washing Gartner describes, and paying agent prices for a rule is how a project ends up in the canceled 40 percent.

The second refusal is anything that replaces the part of the work the business is known for. A builder whose customers hire her for the walkthrough should not automate the walkthrough. A broker whose clients call because she picks up should not put a bot between herself and the first call, even though the bot could. Automation should clear the ground around that work, not pave over it. The post on what still requires a human in a content system makes the same argument for writing: the machine drafts, the owner decides.

The third refusal is my own failure, and it is common enough to name: believing the system can run without a person reading its output. Systems need oversight or the output goes shallow, and they need maintenance or they quietly break. The post on what is running when someone claims a team of agents runs their business is the long version. The short version: anyone selling you a build with no line item for the human who reads it is selling you one of the 95 percent.

Refusing a build when the client can pay for it costs the invoice. It also keeps the client, because the build would have joined the abandoned pile within a year, with my name on it.

How do you run the test this week without buying anything?

Write the ten-by-hand sheet. Take the one item on your list that bothers you most, and for the next ten times it happens, write down on paper who did what, in what order, how long it took, and what went wrong. By the tenth line you will know whether it passes question two, and the sheet itself is the specification if it does.

The sheet needs five columns and nothing else: the date, who handled it, the steps in order, the minutes it took, and the exception if there was one. Ten lines takes two days in a busy shop and two weeks in a quiet one. What the sheet shows, almost every time, is one of three things.

Either the steps are the same every time, the minutes add up to an hour or more a week of the owner's attention, and the exceptions are rare and nameable. That passes, and the sheet is what a builder needs to start. Or the steps vary because three people do it three ways, which means the process needs deciding before anything gets built, and deciding is free. Or the minutes are smaller than you thought and the annoyance was the point, in which case the answer is a checklist on the wall, not software, and you just saved the cost of finding that out the expensive way.

The sheet also answers question three in advance. The exception column is what the automation will have to handle or hand back to a person, and the minutes column is the number you will check against in month two to see whether the build is doing its job. Figuring out what AI your business needs starts from the same place: the process first, the tool last.

What does this cost, and what will it not do?

The test costs ten observed runs and a sheet of paper, somewhere between two days and two weeks, and it will not make the decision for you. It sorts the list. It will also tell you, more often than is comfortable, that the fix is a decision or a checklist rather than a build, and that answer does not generate an invoice for anyone.

The honest tradeoffs:

• It is slow at the front. Ten by hand means watching a thing you already find tedious. The temptation is to skip to the build, and the causes Gartner lists are what skipping looks like.
• It favors the dull. The test will keep sending you back to the doorway automations, the missed call and the order intake, and away from the impressive ones. That is correct, and it is still disappointing if you wanted the impressive one.
• It depends on a count you have to keep. Question three only works if someone looks at the number. In a business of four, that someone is usually you, for five minutes on Friday. The builds that fail the keeping question fail because that five minutes stopped.
• It will not rescue a process you have not decided. If three people quote three ways, the test will tell you so and then stop. Deciding is your job, and it is the part no one can build.

The two piles

Nearly every request that crosses my desk ends up in one of two piles. The first pile is small, dull, and profitable: the call that gets texted back, the order that lands in one record, the quote that gets its follow-up, the inquiry that gets its reply at 10:04. Each returns time to the person whose attention was the constraint, runs on a process already done many times by hand, fails where someone can see it, and costs minutes a week to keep. The second pile is larger and louder, and it is where the 30 percent and the 40 percent and the 95 percent live.

Knowing which pile a request belongs in is the part of this work that cannot be copied from a tool list, and it is the part I spend the most time on. The four questions are the screen; the rest is ordinary building. Run them on paper this week, on the one item that bothers you most, and you will know more about what to build than most of the companies that spent $5 million finding out.

Bennin Systems builds the first pile for small businesses: the intake, the text-back, the follow-up, and the record they all land in, with a count the owner can read on Friday. When a request belongs in the second pile, the answer is no, and the reasons come with it.

Frequently Asked Questions

How do you know if an automation is worth building?

Run four questions in order: whose hour does it give back, has the process worked on paper at least ten times by hand, how will you know it stopped, and what does it cost to keep. Build when all four come back yes. One no means fix that first; two means the request is not ready.

What percentage of AI automation projects fail?

The published figures are enterprise-scale forecasts and surveys, not small-business counts. Gartner forecast in July 2024 that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, and in June 2025 that over 40 percent of agentic AI projects would be canceled by the end of 2027. MIT NANDA's preliminary July 2025 report found 95 percent of organizations in its large-company sample getting zero return.

Why do most small business automations fail?

They encode a process nobody had written down, so the software runs the confusion faster; they fail silently until a customer notices; or nobody counted the cost of keeping them running. Michael Hammer named the first failure in 1990: companies "leave the existing processes intact and use computers simply to speed them up."

What is the Scarcest-Resource Test?

A four-question screen for any proposed automation: whose hour it gives back, whether it has worked on paper, how you will know it stopped, and what it costs to keep. It is named for the first question, because the scarcest resource in a small business is the owner's attention, and an automation that does not return some of it is rarely urgent.

What is the ten-by-hand rule?

Before building an automation, do the process by hand, the same way, at least ten times, and write down each run: who did it, the steps, the minutes, and any exception. Ten runs surface the exceptions the software will have to handle, and the sheet becomes the specification if the build goes ahead.

Which automations are usually worth building first?

The ones at the doorway where a customer first reaches in: missed-call text-back, phone and web order intake into one record, scheduled quote follow-up, and a fast reply to inquiry forms. They tend to pass all four questions because the process already exists, failure is visible, and the hour saved belongs to the owner.

Do small businesses need AI agents?

Mostly not yet. Gartner's June 2025 guidance is to use agents when decisions are needed, automation for routine workflows, and assistants for simple retrieval, and it notes that many use cases labeled agentic do not require agents. For a business of eight, most of the value is in routine workflows, which are rules and forms, not agents.

How many small businesses use AI right now?

The Census Bureau's survey put AI use at 17 to 20 percent of U.S. employer businesses from December 2025 to May 2026, with under 20 percent among firms with four or fewer employees and 37 percent among firms with 250 or more. The SBA's older, narrower measure found 7.6 percent between September 2024 and August 2025.

---

Stacy Bennin is the founder of Bennin Systems, where she builds the automated systems small businesses need but rarely have time to set up themselves: lead capture and follow-up that runs on its own, chatbots that answer questions and take orders around the clock, custom websites that act as an employee, and the back-office workflows that keep an operation from running on memory and sticky notes. Located in Montana, she works with businesses and real estate professionals anywhere in the United States. She is also a licensed Montana real estate broker affiliated with Legacy Lands Real Estate. Reach her at benninsystems.com.

Bennin Systems · Paradise Valley, Montana · (406) 224-3267 · benninsystems.com