The 28-tab spreadsheet
The first thing I noticed walking through the warehouse last spring was the second monitor on Karen's desk. Or rather, what lived on it. A single spreadsheet, twenty-eight tabs across the bottom — orders, returns, freight quotes, vendor pricing, a tab she'd labeled "weird stuff." Karen ran operations for a roughly 60-person specialty packaging distributor in the American Midwest. She'd built that spreadsheet over six years. It worked. It also took her four hours every Friday afternoon to keep current, and she was the only person in the company who fully understood it.
That detail — the spreadsheet, the four hours, the single point of failure — is where this story starts. It is also where the series ends.
Across the previous seven installments we worked through what looks like a framework: assessment, the agents-versus-workflows question, building a roadmap, costing it out, measuring ROI, and the patterns of failure that show up across most adoptions. Reading it in sequence makes the work sound tidy. It isn't. The point of all that scaffolding was to give a working leader something to push against when the vendor in front of them is promising the moon and the budget meeting is in three days. A case study is a different animal. Instead of frameworks it offers texture — the order decisions came in, what got harder than expected, what turned out to be easier. So that is what this piece is. One company, composite but real-shaped, before and after.
Before: tribal knowledge at $42M
Karen's company did about $42M in annual revenue, sold custom packaging materials to mid-market manufacturers, and ran four warehouses across two states. Profitable. Growing slowly. The kind of business that doesn't show up in tech press but employs most of the people on a given county road.
The pain points weren't dramatic. They were structural. Quote turnaround averaged 3.5 business days because three different people needed to touch each one. Customer service reps were spending close to 40% of their time looking up information that already existed in five disconnected systems. The Friday spreadsheet ritual cost roughly 200 hours a year of senior-operator time. None of that was bleeding the company. All of it was a ceiling.
When I asked the CFO what kept her up at night, she said this. "If Karen gets hit by a bus we don't know what our margins are for about a week. Maybe two." That is the kind of risk that doesn't show up on a balance sheet until it does.
The leadership team had also done what most leadership teams have done by now. They'd bought a Copilot license. Two of the sales reps had paid out of pocket for ChatGPT. The marketing person had been using Midjourney for proposal covers. None of it was talking to anything else, and none of it had moved a single business metric. "We tried Copilot for a month and just stopped opening it," the operations director told me over coffee. That sentence, almost word for word, is one I have now heard from six different SMB operators on two continents.
The assessment that mattered more than the tool
The first thing we did was not pick a tool. It was sit down with Karen for two hours and have her narrate, click by click, what happened in those four hours every Friday. By minute forty we had a list of seventeen discrete decisions she was making, eleven of which followed rules she could state out loud and six of which she said "depended on the customer."
Strictly speaking the savings we ended up booking later weren't AI savings — they were savings from finally writing down the rules. AI was the excuse to do the writing. That is the unglamorous half of every adoption I've watched succeed. Before any model gets involved, somebody has to interview the human who has been holding the system together and translate what she knows into something explicit. The kind of mismatch we walked through in the failure-patterns piece — picking a tool first, mapping the workflow second — almost always traces back to skipping that conversation.
"I've been doing this so long I forgot it was complicated," Karen said in the second hour, when we got to the part of the spreadsheet that flagged customers on credit hold. "Nobody ever asked me to explain it."
Nobody ever asks. That is the gap, in one sentence.
What we built, and what we didn't
Three things ended up in production over the following four months. The order matters more than the technology.
The first was a quoting assistant. Not autonomous — a retrieval system that pulled vendor pricing, freight rates, and historical margin on similar customers, then drafted a quote the rep reviewed and edited. Average turnaround went from 3.5 days to about 6 hours. The reps still made every final decision. The system just stopped making them hunt for inputs.
The second was a customer-service knowledge layer. The five disconnected systems did not get unified — that would have been a 12-month integration project and the budget didn't exist. Instead the team indexed all five sources into a single search interface a rep could query in natural language. Time-on-information dropped from 40% to roughly 18%. NPS moved from the mid-30s to the high-50s within two quarters, mostly because reps stopped saying "let me check on that and get back to you."
The third — the one nobody saw coming — was a Friday-spreadsheet replacement that wasn't a replacement. Karen still wanted her spreadsheet. What we built was a pipeline that pre-filled twenty of the twenty-eight tabs automatically every Thursday night, leaving her with the eight that genuinely required her judgment. Her Friday went from four hours to about forty-five minutes. She used the recovered time to start documenting the rules for the remaining eight tabs, which means in another six months that ritual probably disappears entirely.
Notice what is not on that list. No autonomous agents. No replacement of humans. No "AI transformation." Three narrow, boring, retrieval-and-drafting systems wired into workflows that already existed. Total spend including consulting, tooling, and internal time over the first year ran roughly $180K against a measured first-year benefit of around $640K. The 3-to-1 return middle-market surveys keep reporting isn't mysterious. This is what it looks like up close.
What carried over from earlier installments
The roadmap piece argued for twelve months of small wins instead of one big transformation. The case above is, almost embarrassingly, exactly that. Quote assistant in month two, knowledge layer in month four, spreadsheet pipeline in month six. None of it was the headline project the original vendor pitched, which was a customer-facing chatbot the sales team did not want and the customers had not asked for.
The costing piece pushed back on vendor math. In Karen's case, the original quote from the chatbot vendor was $310K in year one with a 14-month payback. The three things we actually shipped came in at $180K with a payback under five months. The framework didn't predict that gap. It just made the gap visible during procurement.
Eighteen months later, and what closes the series
Here is what is different at Karen's company eighteen months in. Quote-to-close cycle is meaningfully shorter, which has shown up as roughly 9% revenue growth without adding headcount. Customer service quality is up by every internal metric the company tracks. Karen took an actual vacation — first one in three years — and the company's margin reporting continued without her. Two of the warehouse staff who were skeptical at the start are still skeptical, though they use the quote tool every day.
What didn't change is also worth naming. The company still runs on the same five core systems. The org chart looks identical. The culture didn't get reinvented. AI didn't transform anything. It removed friction from the things the business was already trying to do.
In conversations with operators across Korean and Japanese SMBs, the same pattern keeps showing up. The companies that get returns aren't the ones with the most ambitious AI strategy. They are the ones who can point at a specific Karen, a specific four-hour ritual, a specific 28-tab spreadsheet. The framing changes by market — Japanese mid-market firms tend to call it "removing 属人化," Korean ones "process standardization," American ones "scaling without hiring" — but the underlying move is the same. Find the person holding the system together. Make her tacit knowledge explicit. Use AI to industrialize the explicit version. Companies that skip the first two steps and try to industrialize the implicit one end up with the Copilot story. A tool nobody opens after the first month.
Eight pieces. Worth a sentence each before we close. The assessment piece was about asking the right question before buying anything. Agents-versus-workflows was about not confusing autonomy with usefulness. The roadmap installment argued that twelve months of small wins beats one big transformation. The costing piece pushed back on vendor math. The ROI piece was about measuring the right thing, not the easy thing. The failure-patterns piece named the five places adoptions die. And this one was the attempt to show what it actually looks like when the framework meets a real shop floor.
If there is a single thread, it is that AI adoption in small and mid-sized companies is less about technology than it looks. It is about whether you can describe your own business clearly enough for a system to help with it. Most companies cannot, yet. The ones that learn to are the ones who will quietly compound an advantage over the next three years.
The series ends here, but the work doesn't. If you have read straight through, you have, at minimum, the vocabulary to push back on whatever is being pitched to you next quarter. The harder thing is starting. Pick the Karen in your own company. Sit with her for two hours. Watch what she clicks. The pilot you should run next is whatever that conversation tells you to run — not whatever was on the slide deck last week.
At 5years+ we spend most of our time on exactly this kind of work with SMBs across Korea and Japan, and the pattern in the case above is closer to typical than exceptional. If you'd like a second pair of eyes on where to start, the contact form on the site goes straight to a real person. But honestly, you can probably begin on Monday morning without us. The spreadsheet is already there. Somebody on your team has been opening it for years.