The wrong question, asked more politely
A few weeks after the boardroom scene I described last time, the same CFO called me back. He had clearly been thinking. "Okay," he said, "so we agreed we're not behind. But where do we actually start?" It was a better question than the first one. It was still the wrong question.
The right question isn't where do we start. It's which part of our company is already screaming, and we've been too polite to listen. Every mid-market company I've walked through has at least one function that has been quietly absorbing pain for years — a Friday-afternoon report someone dreads, a stack of paper invoices that nobody volunteers to touch, a customer-service queue that grows on weekends and gets blamed on Monday. AI doesn't help you find new work. It helps you stop doing the work you've been pretending isn't there.
Five areas, and an honest mirror
If you haven't yet worked through the five questions from last week's piece, do that first — this exercise assumes you've already gotten past the "should we even?" stage and are ready to look at your own org chart with slightly cooler eyes.
What follows is a self-assessment, not a maturity model. I'm wary of maturity models. They tend to make companies feel either too good or too bad about themselves, and neither is useful. Instead, I want to give you five functional areas where AI actually lands in real SMB and mid-market companies, and a small set of signals you can verify by walking the floor on a Tuesday morning.
One number, then I'll drop the numbers. Industry research from 2026 puts SMB generative-AI usage above 58% — more than double 2023 — but only around 12% of SMBs say they have a real strategy. The gap between using AI and adopting AI is, in other words, enormous. Most of that gap closes the moment a company picks the right area to start. Most of it stays open when they don't.
Area 1 — Knowledge and decision work
This is the area people usually want to talk about first, because it's the most visible: proposals, sales follow-ups, internal Q&A, the inbox that drowns your senior people. The work is text-heavy, judgment-heavy, and almost always being done by someone whose time costs the company a great deal.
Signals that this is your area: your sales team rewrites the same kind of proposal five times a month with minor variations. New hires take months to find answers that live in three different SharePoints and one ex-employee's head. A senior person spends Sunday evening writing a follow-up note that, in any honest accounting, three other people in the company could have written if they had access to the right context.
The fit here is what I'd call agent-shaped work — small, focused AI agents that handle a defined slice of knowledge work end-to-end, plus, occasionally, a lightweight internal web tool to make that knowledge accessible to the rest of the org. It's not glamorous. It is, in my experience, where the fastest visible ROI tends to show up.
Area 2 — Visual content production
I had a client last year — a mid-sized cosmetics brand — show me their product-shot pipeline. Eleven steps. Three vendors. A turnaround that, when I pressed, was actually closer to four weeks than the two they told me on the first call. Their competitor was shipping new SKU imagery in three days.
If your company sells anything physical, or runs ad creative at any volume, this area deserves a hard look. Signals: photographer bookings are a recurring line item. Your marketing team has a backlog of "we'll get to that SKU eventually" products. You're paying to reshoot the same product in a different background because a regional channel needed it. Short-form video for social channels is something you've talked about for two years and still haven't started.
AI image and video generation, used well, doesn't replace your photographer for the hero shots. It replaces the long tail — the variations, the localizations, the channel-specific cuts — which is where most of the cost actually hides.
Area 3 — Document and data ingestion
This is the most boring area on the list, and the one I would tell you to look at first if I had to pick one. Boring is good. Boring means the ROI calculation is simple.
Signals: someone in your company touches paper. There is at least one fax machine still plugged in (don't laugh — in Japan and Korea both, this is closer to the rule than the exception). Invoices arrive as PDFs that someone retypes into your accounting system. Field staff fill in handwritten forms that get scanned and then, somehow, retyped anyway. A scanned contract from 2019 cannot be searched, only remembered.
The fit is OCR combined with language models — not the OCR of a decade ago that gave you a 78% accuracy rate and an apology, but the kind that reads a messy form and hands you structured, queryable data. The reason I push clients toward this area first is that the value is auditable. You can count the keystrokes you stopped typing. Boards like that.
Area 4 — Cross-system workflow automation
A manufacturing client of mine has a Friday afternoon report ritual. One person, every week, exports data from four systems, drops them into Excel, runs a set of pivot tables, formats the result in a particular way the founder once asked for in 2017, and emails it out by 5pm. It takes her about half a day. She is also, separately, a very good operations analyst.
If a single Friday-afternoon report takes one person half a day, that's a signal. So is this: any time data leaves one system to be reshaped and dropped into another, by a human, on a recurring schedule. Reconciliations. Inventory feeds. Sales-to-finance handoffs. Dashboards that someone refreshes manually because the integration was "too complicated to set up properly."
The honest answer here is usually a mix of RPA, light AI judgment in the middle, and an automated reporting layer on top. It's the area where I see the most over-engineering — companies trying to build a Tesla when a Toyota would do — so the self-assessment matters more than usual. Start with the report nobody enjoys writing.
Area 5 — Customer-facing operations
This is the area where I have, more than once, told a client to wait. Not because the technology isn't ready — it largely is — but because customer-facing AI is the one place where a bad implementation is publicly visible, and reputational cost compounds in a way internal-tool failures don't.
Signals that you're genuinely ready: your support volume has a clear repeat pattern (the same 20 questions are 70% of tickets). You operate across time zones or languages and your current coverage is honestly thin after hours. Your reservation, booking, or intake process has a step where customers drop off because they couldn't get a human at 9pm on Tuesday. You have a knowledge base that someone actually maintains.
If those signals are present, a chatbot or voice agent in front of your customers can be transformative. If they aren't — particularly the knowledge-base one — wait. The bot will only ever be as good as what you've written down, and a confident bot trained on contradictions does more damage than a slow human.
How to actually do this assessment
Don't run a workshop. Workshops produce consensus, and consensus on this question is almost always wrong because the loudest functional leader wins. Instead, take the five areas above to three people separately: someone in operations, someone in finance, and one frontline manager you trust. Ask each of them, independently, which area causes them the most preventable pain. Don't tell them what the others said.
Where two of the three converge, you have your starting area. Where all three converge, you've probably been ignoring it for two years already.
I'll admit something. I don't have a clean answer for the company where the three answers all diverge. In those cases I usually go with finance's pick, because finance tends to translate pain into numbers fastest, and numbers are what gets a first project funded. But that's a heuristic, not a rule, and I've been wrong about it.
At 5years+, where I work with Korean and Japanese SMB and mid-market clients across these five areas, the pattern we see most often is that companies pick the most exciting area when they should have picked the most boring one. The boring one, eight times out of ten, ships in a quarter and pays for the next three projects. The exciting one becomes a slide.
Next time, we'll take the most common starting answer once a company picks its area — "we should automate something" — and split it cleanly: AI agents versus workflow automation, and why mid-market teams almost always reach for the wrong one first.