The two best AI systems published today were not products. They were built by two staff members automating their own jobs, and they share the same three parts.
The hook, and the answer
The Rundown AI runs a weekly feature where their own staff describe how they use AI at work. Today's edition had two of them, and they were worth more than every product announcement in the newsletter combined.
Zach, their Editor-in-Chief, connected Claude to Beehiiv, the platform their newsletters run on. A scheduled weekly task pulls open rates, clicks, reader ratings, and comments across all three of their newsletters, surfaces themes across issues, and packages it into a page he presents at the team meeting. He also built a filter to weed out bot clicks and fake poll votes.
Then he said something that tells you he understands what he built. "The follow-up questions are where I really feel the value. With the data in front of me, I can keep poking at issues, ask for more, or troubleshoot on the fly in the same chat with the context already there." He also said the connector made him more comfortable diving into an important area he used to dread.
Brooke, their Community Engagement Lead, runs the support inbox alone. She connected Claude to Gmail, Slack, and scheduled tasks. Every Thursday it reads the week's threads, filters out spam and noise, groups the real member questions by theme and then by the team that can fix each one, and drafts one short post per team covering the top issue plus a suggested fix. She reviews and tweaks every post before it goes to that team's Slack channel, ahead of the Friday meeting.
Monthly, it audits the full ticket set for what drove volume, what changed from last month, and what is a product gap versus a one-off. And it checks her saved replies against what members actually asked.
Her stated goal is the thing most support operations never reach: recurring tickets fixed at the source instead of answered over and over.
So here is the direct answer to the headline question. Automate the one task you already do every single week and resent, and build exactly three things around it. A recurring trigger, so it runs whether anyone remembers. A named filter for garbage, because real data has junk in it. And a human review step before anything leaves the building.
Three parts. Everything else is decoration.
One honesty note up front, and I want it in the opening rather than the small print. Neither Zach nor Brooke reported a single outcome number. Brooke says she just finished building hers. These are credible as design patterns. They are not proof of results.
Key Takeaways
- The architecture is a recurring trigger, a named filter for junk, and a human review step before output leaves the building.
- The Federal Reserve found that 46 percent of small employer firms use AI, but just 7 percent of those users have fully integrated it, meaning roughly 3 percent of all small employers have AI genuinely embedded in a process.
- Automated traffic was more than 53 percent of all web traffic in 2025, against 47 percent human, which is why a filter for junk is a load-bearing part of the system rather than a nicety.
- The case for a human review step is accountability and adoption, not accuracy; research actually found human-in-the-loop review increased willingness to use automation while reducing decision accuracy.
- Knight Capital lost 460 million dollars in 45 minutes partly because no written procedure required a second person to check a deployment, and lost 7.5 million separately because automation kept running on weekend test data.
The problem
Almost everybody is chatting. Almost nobody has built a system.
The Federal Reserve Banks' 2026 Report on Employer Firms, drawn from 6,525 responses to the 2025 Small Business Credit Survey, gives you the shape of it. Forty-six percent of small employer firms report that the business or its employees currently use AI. Of those users, about half say they are experimenting, 44 percent have partially integrated AI into business processes, and just 7 percent have fully integrated it.
Do that arithmetic. Roughly 3 percent of all small employer firms have AI genuinely embedded in a business function.
Look at what they use it for, too. Writing or marketing, 83 percent. Individual productivity, 61 percent. Planning or analysis, 51 percent. Notice what is absent from the Fed's top categories entirely: nothing resembling scheduled reporting, inbox triage, or data quality appears. Those are all drafting and thinking tasks. Useful, and all of them one-off.
Note the Fed's caveats, because they matter. This is a convenience sample, not a random one, and the Fed says so itself. The AI questions came from an optional module. And the definition shown to respondents included features of software you already use, which is broad enough that the 46 percent includes people who never deliberately used an AI tool.
Here is why I think the 3 percent happens, and it is not a skills problem.
People try to automate the impressive thing. The thing that would be a story if it worked. And impressive things are complicated, so they take three weekends, and they break in a way nobody can diagnose, and then they get quietly abandoned around week five.
Meanwhile the task you do every Tuesday that you hate, the one that takes forty minutes and produces something three people need, sits there unautomated because it is too boring to be exciting.
That boring task is the one. It recurs, which means the work pays back weekly. It is low stakes, which means an early mistake costs you an apology rather than a client. And you already know the rules, which means you can write them down.
But what if the reason your last automation attempt died was not that AI could not do it, and not that you lacked the technical skill, but that you built one part of a three-part system and the other two were never there?
The evidence
Let me take the three parts in turn, because each one has real research behind it, and in one case the research says something different from what everyone repeats.
On starting narrow. The most cited study in this field is MIT NANDA's "The GenAI Divide: State of AI in Business 2025," famous for the claim that 95 percent of AI pilots fail. I read it, and I want to tell you what it actually measured, because the headline is being used dishonestly by people who have never opened it.
The report labels itself Preliminary Findings. Its method was a review of over 300 publicly disclosed initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders recruited at four industry conferences. That is a convenience sample. The 95 percent does not describe AI projects generally: it is the complement of a 5 percent figure that applies only to custom, task-specific tools. In the same report, general-purpose tools like ChatGPT and Copilot run 80 percent investigated, 50 percent piloted, 40 percent implemented. Success was defined as users or executives having remarked on a marked and sustained impact, within a six-month window, which the authors themselves concede may be too short. And all four authors were developing the agentic framework the report names as a path forward, with no conflict disclosure anywhere in the document.
So treat the 95 percent with suspicion. But do not throw out the report, because its findings on what works are specific and they support exactly what Zach and Brooke built.
Page 14: "The best startups crossing the divide focus on narrow but high-value use cases, integrate deeply into workflows, and scale through continuous learning rather than broad feature sets." Page 16: "Tools that succeeded shared two traits: low configuration burden and immediate, visible value. In contrast, tools requiring extensive enterprise customization often stalled at pilot stage." Page 17 maps scope against complexity, and narrow plus simple is labeled fast wins while broad plus complex is labeled fails.
And then page 8, which is the finding that made me sit up. Only 40 percent of companies said they had purchased an official AI subscription, while workers from over 90 percent of surveyed companies reported regularly using personal AI tools for work tasks. The report calls it a shadow AI economy and notes that many of the strongest enterprise deployments began with power users who became champions.
The best systems are being built by the person who does the job. That is in the data.
I will be straight with you about where that report disagrees with me, because you deserve to know. Its headline prescription on page 23 is to buy rather than build, and page 19 reports external partnerships reaching deployment around 67 percent of the time against roughly 33 percent for internally built tools. Two things about that. The report's own caveat on the same page says those differences may reflect organizational capabilities rather than approach, and that correlation does not prove causation. And build in that report means an enterprise software project with engineers and procurement. It does not mean an editor wiring a scheduled task to his newsletter platform. That is the report's own prosumer finding, which it endorses.
On the filter for junk. This is the part people skip, and it is the part that quietly ruins the output.
Imperva's 2026 Bad Bot Report found automated traffic accounted for more than 53 percent of all web traffic in 2025, with human activity down to 47 percent and still falling. Imperva sells bot protection, so flag the interest, and the measurement base is traffic crossing their own network. But their own conclusion is the sentence you need: "If a growing share of traffic is automated, then traditional metrics such as user engagement, conversion rates, and demand signals become harder to interpret. A spike in traffic may not indicate customer interest."
That is why Zach's bot-click filter is not fussiness. It is the difference between a report and a hallucination made of real numbers.
On data quality generally, I want to warn you off the two most quoted figures in business, because both fall apart on inspection. The claim that bad data costs the US economy 3.1 trillion dollars a year, usually attributed to IBM, traces to a 2011 magazine column whose own author wrote "no matter how far off my estimate is." There is no national estimate with a published method. And the Gartner figure of 12.9 million dollars per company is an estimate supplied by 154 reference customers of data-quality vendors, in 2020, who were already shopping for data-quality software.
Use this instead. Nagle, Redman and Sammon, publishing in Business Horizons in 2020, found that across 75 data quality assessments, an average of 47 percent of newly created data records contained at least one critical error. Only 3 percent of those assessments scored above 97 percent.
Half your new records have something wrong in them. A system that cannot name its junk will average that into the answer.
On the human review step, where the research will surprise you. I was ready to tell you that a human reviewer makes the output more accurate. The research does not say that.
Sele and Chugunova, publishing in PLOS ONE in 2024 with 292 participants, titled their paper "Putting a human in the loop: Increasing uptake, but decreasing accuracy of automated decision-making." Adding a human in the loop increased people's willingness to use the automation and decreased the accuracy of the resulting decisions. A related 2024 study found that incorrect algorithmic support degrades human judgment particularly when the human sees the AI output before forming their own judgment.
So be honest about why the review step belongs there. It is not because you are a better classifier than the model. It is accountability, it is adoption, and it is the fact that you are the one whose name is on the output. Those are sufficient reasons.
And there is a different kind of review that the evidence supports very strongly, which is procedural rather than cognitive. A second person checking that the thing was deployed correctly, rather than a person second-guessing a judgment.
Which brings me to the best-documented automation failure in American business.
In the days before August 1, 2012, beginning July 27, Knight Capital deployed new order-router code across eight servers in stages. According to the SEC's administrative order, a technician did not copy the new code to one of the eight, and no second technician reviewed the deployment, because Knight had no written procedure requiring one. Dormant code from 2003 on that eighth server was still callable, and the new deployment reused its activation flag.
Two hundred and twelve incoming orders produced four million executions in 154 stocks, more than 397 million shares, in approximately 45 minutes. Knight realized a 460 million dollar loss. The SEC penalty was 12 million.
Here is the detail I cannot get past. Starting around 8:01 that morning, before the market opened, an internal system fired 97 automated warning emails. The SEC's order states it plainly: Knight did not design these types of messages to be system alerts, and Knight personnel generally did not review them.
The warnings existed. Nobody read them.
The SEC wrote the moral itself: a written procedure requiring a simple double-check of the deployment could have identified that a server had been missed and averted the events of August 1.
And in the same order, the smaller, more relatable loss. In October 2011, Knight used test data for a weekend disaster recovery test. Afterward, the desk mistakenly continued using that test data to generate automated quotes when trading opened on Monday. Nearly 7.5 million dollars gone. The SEC's finding: Knight did not have a mechanism to test whether their systems were relying on stale data.
Automation ran on fake data because nothing checked whether the input was real. That is the filter argument, with a regulator's signature on it.
The three parts, built
Now look at what Zach's and Brooke's systems have in common, because it is the architecture, and it is the same every time.
A recurring trigger. His runs weekly. Hers runs every Thursday, with a monthly deeper pass. Nav, a reader whose workflow The Rundown published in the same edition, runs his on Fridays: he records his sons' tutoring sessions, transcribes with Otter.ai, has ChatGPT summarize each lesson and identify weaknesses, then holds a Friday progress checkpoint. His line was perfect. "I'm no longer just paying for tutoring and hoping it's working."
Three systems, three fixed days. Nothing that depends on remembering.
A named filter for junk. His weeds out bot clicks and fake poll votes. Hers filters spam and noise before anything gets grouped. And this is the part to get right, because the instinct is to make the filter aggressive. Do not. Build three buckets, not two: genuine, definitely not relevant, and needs my eyes. Anything the rule cannot classify confidently goes in the third pile rather than being guessed at.
A human review step. She reviews and tweaks every post before it goes to a team's Slack channel. Nothing reaches another human unreviewed.
Then there is a fourth thing in Brooke's system that is not part of the pattern but is the best idea in either of them, and I want you to steal it specifically.
Her system does not answer tickets faster. It routes findings to the team that can stop causing them. It groups member questions by theme, and then by who can fix each theme, and drafts a note to that person. Monthly, it tells her what is a product gap versus a one-off, and it checks her saved replies against what members actually asked.
Most people point AI at replying faster. She pointed it at the cause.
That difference is worth more than any tool you will buy this year, and it is also why I believe her instinct even without numbers. Nobody publishes before-and-after data on recurring ticket classes that stopped recurring, because the industry rewards deflection metrics instead. She is chasing a number her own industry does not measure.
One more observation, and it should settle the build-versus-buy question for you better than any study. readywhen launched on Product Hunt today: an AI assistant that is nothing but an email address, which chases your follow-ups and sends nothing until you reply with the word send. Its commenters praised that approval gate specifically. It is a good product.
It is also one narrow slice of Brooke's job, sold as a product, while she built the whole thing with tools her employer already paid for.
The workflows people build for themselves are the product roadmap. You can be eighteen months ahead of it.
Practical steps
1. Pick the boring recurring task, not the impressive one. List what you do every week with rough time attached, then circle the one you resent where a mistake is cheap to fix. That is your first build. The impressive automation is your fourth.
2. Set the trigger before you build anything else. Pick a day and time, and pick one that is not your busiest. Decide now what should happen when the run fails or the data is not ready, because that is the moment that kills most systems silently.
3. Write the junk rule down, with examples on both sides. Give at least three real examples of junk and two examples of real data that superficially resembles junk. Then test the rule against your own counter-examples before you trust it. Three buckets, with a review pile for anything uncertain.
4. Size the review step for under three minutes. Four checks maximum, in priority order, with one marked as never skippable. A twenty-minute review gets skipped by week four, and then you have an unsupervised system you believe is supervised.
5. Run it manually once, following your own instructions exactly. Do not help it. Note every point where your instructions were ambiguous and what got assumed. This is the step that finds the holes, and it is the step everyone skips because the system already feels built.
6. Add the monthly layer, which answers a different question. The weekly run tells you what happened. The monthly run should tell you what to fix at the source: what drove volume, what changed, and what is a root cause versus a one-off. Require it to recommend one thing to fix upstream. That is where the compounding lives.
7. Write the pattern down for your team before you build the second one. Three parts, one worked example from your own business, and one rule that is not negotiable: nothing built this way sends anything externally without a human review. If only you can build these, your business has a bottleneck shaped exactly like you.
Frequently Asked Questions
What should I automate first?
The task you already do every week and resent, where a mistake is cheap to fix. MIT NANDA's research found narrow and simple use cases produce fast wins while broad and complex ones fail. The Fed's data shows most small-business AI use is writing and marketing, which leaves recurring operational reporting almost entirely unoccupied.
Should I build it myself or buy a tool?
Build first if the workflow is specific to how you work, because that specificity is exactly what a general product cannot sell you. Buy when the job is generic and the maintenance would fall on you. Count your own time at a real rate, and count maintenance, which is the cost everyone forgets.
Why does my data need a filter?
Because over 53 percent of web traffic is now automated, and because research across 75 data quality assessments found an average 47 percent of new records contained at least one critical error. Knight Capital lost nearly 7.5 million dollars because automation kept running on weekend test data and nothing checked whether the input was real.
How do I stop AI automation from making mistakes?
Write the procedure down, including a double-check. The SEC concluded that a written procedure requiring a simple double-check of a deployment could have averted Knight Capital's 460 million dollar loss. Also form your own judgment before you look at the AI's answer, since research shows reviewers anchor badly when they see the machine's output first.
Do I need new software to do this?
Usually not. Both workflows described here were built by connecting an AI subscription the employer already paid for to tools they already used, specifically a newsletter platform, Gmail, and Slack, plus scheduled tasks. Start with what you own and find out where it genuinely stops before you add a bill.
The close
Three parts. A day it runs, a rule for what to throw out, and your own eyes before it ships.
I have watched a lot of people spend a lot of money looking for the AI tool that fixes their week, and I understand the appeal, because a purchase feels like progress and a Tuesday afternoon spent writing down a filter rule does not.
But look at what actually happened today. Two people at a newsletter, neither of them engineers, built systems that do real recurring work, using a subscription their employer already had. A parent built a system that tells him whether the tutoring he pays for is working. None of them bought a platform. None of them ran a pilot. All three picked one recurring job and wired three parts around it.
And candidly, that is the part of this I find genuinely hopeful, more than any model release this year. The gap between what AI can do and what small businesses are getting from it is not a capability gap. The Fed says 46 percent of us are using AI, and of those, about half are still experimenting, 44 percent have partially integrated it, and 7 percent have fully integrated it. That is not a wall. That is an empty room.
The thing standing between you and the 3 percent is not a tool you have not found, a budget you do not have, or a technical skill you never learned. It is one boring task, picked on purpose, with a day attached to it.
So pick the task you resent. Give it a day. Name the junk. Keep your eyes on the output.
Then do it again next month with a different task, and in a year you will not have an AI strategy. You will have a business that runs differently, built out of six boring Tuesdays.
P.S. If you build one of these, build the monthly layer too. The weekly run saves you time, which is nice. The monthly run tells you what to fix so the work stops arriving, which is the whole point, and it is the part Brooke built that almost nobody else does.
About the author
Jonathan Mast is the founder and CEO of White Beard Strategies, where he teaches non-technical business owners how to use AI to amplify the skill and experience they already have. He runs a 500,000-member community of entrepreneurs learning AI in plain language, created the Perfect Prompt Framework 2.0, and speaks regularly on practical AI adoption for small business. He thinks in systems to a degree his team finds mildly exhausting, which is the one personality trait that turned out to be genuinely useful in the AI era.