Month-end reconciliation, before AI, ran like this: finance exports a file from the ERP, the warehouse exports one from the shipping system, two people block out an afternoon to match them, anything that doesn't match goes in an email to sales, sales replies the next day, and someone goes back to fix it. Three days, done.
After adopting AI, it looks like this: AI reads both files and lists the discrepancies in ten minutes, finance spends a morning checking whether AI misread anything, anything that doesn't match still goes in an email to sales, and sales still replies the next day. Two and a half days, done.
Half a day saved. That half day cost an annual subscription, a one-time implementation fee, and two training sessions.
The owner usually draws one conclusion from this: AI isn't as magical as they'd imagined. That conclusion is aimed at the wrong place. The problem isn't whether AI reads accurately enough. It's that from start to finish, nobody went back and asked one question about this workflow: do these steps still need to exist?
01Why the workflow looks like this in the first place
Any workflow that has been running in a company for more than five years wasn't designed out of thin air. Four things forced it into shape.
Traditional workflows split up work, add approvals, keep records, and hold alignment meetings to compensate for four limits: people are expensive, people are slow, people forget, and passing information from person to person is costly.
Look back at reconciliation. Why do two people each export a file? Because one person can't finish it alone, and because if one person does it, nobody can catch their mistakes. Why email sales? Because the discount in a salesperson's head isn't written in any system. Why keep a record? Because three months from now, nobody will remember why it was approved.
None of these steps deals with the business itself. They deal with human limits.
Lay out any company's workflow diagram and a sizable share of the boxes are buying insurance against "people make mistakes, forget, and quit."
02Move it over and you get a more expensive old system
Here's a counterexample that will look familiar to any company that has rolled out AI:
Ticket comes in → AI classifies → human review → AI drafts reply → human review → send result
On a slide it looks fully AI-powered. In substance, nothing has changed. None of the original six steps is gone. Two of the boxes swapped a person for a model, then each got a human review step added next to it.
Run the numbers and the account is actually negative. The old workflow's costs stay exactly where they were, and a new layer of cost sits on top: subscription fees, implementation fees, training time, and the one most often underestimated — people now spend time reviewing AI. And because they don't trust it, they review it more carefully than they ever reviewed a colleague.
Putting AI in every box doesn't mean the workflow has been rebuilt for AI. It's the same road with a more expensive crew.
There's a very useful reverse test: take AI out of the workflow — does it still run? If it does, AI is just a plug-in, not the heart. Take AI out of the reconciliation workflow and finance and the warehouse can still block out an afternoon and finish it. So that isn't AI transformation. That's paying for a faster calculator.
03Which kinds of steps should be deleted
The work that's genuinely ready to be restructured falls roughly into three groups. Translated into what happens every day at a small or mid-sized business, they look like this:
Steps that mostly move information around. Export from system A, paste into system B; type the orders from a LINE group chat into Excel; fill in the same customer details on three different forms.
Steps that mostly sync context. The weekly meeting held so everyone knows where things stand; the email with seven people on CC; the two days waiting for a manager's sign-off.
Steps that mostly make up for memory. Handover documents, monthly reports, and re-summarizing everything you did once a quarter.
| What this step compensates for | The old solution | After AI |
|---|---|---|
| People can't read it all at once | Split it among three people | Read in one pass — the splitting itself disappears |
| People forget | Handover docs, monthly reports, review meetings | Context stays in the system — memory-patching steps disappear |
| Information distorts as it's passed along | Alignment meetings, emails for the record, lots of CCs | Everyone shares the same context in real time |
| People make mistakes | Add a human review step | Keep it, but only in the few high-risk boxes |
The first three rows get deleted. The fourth stays.
Let's be precise about the line here: what gets deleted is steps, not people.
When reconciliation, file conversion, chasing paperwork, and filling in the same forms disappear from the diagram, the people who did them don't disappear with them. Low-judgment work gets absorbed by agents; high-judgment work gets amplified by people. The afternoon finance saves should go to "why has this client's margin dropped three months in a row." That question used to always sit behind reconciliation, and there was never time for it.
If a company adopts AI and the only thing it thinks of is cutting two people, what it bought is a short-term saving on payroll, not a new capability.
04Which boxes have to keep a person
Most companies ask: which steps can we bring AI into? That's addition, and if you keep adding you end up with that six-box ticket workflow with AI in every box.
The question to ask is: which steps still need to keep a human? That's subtraction. Start from six boxes and delete until you can't delete any more. What's left is the workflow's real skeleton.
Which boxes can't be deleted? Three things are considered irreplaceable by humans: choosing the problem, aesthetic judgment, and decisions and accountability. Add one practical red line: high-impact actions must be confirmed by a person — human-in-the-loop — and the threshold can be very concrete, such as any contract above a certain amount must be signed by a person.
So once the reconciliation workflow is rebuilt, a more sensible version looks like this: exporting, matching, the discrepancy list, and the follow-up emails all run on their own in the system, and people show up in only two places — the discrepancies above the threshold, and the judgment call on "should we renegotiate this client's terms."
From six boxes to two. That's what it means for AI to restructure a workflow. Going from three days to two and a half was just AI brushing up against it.
05Why most companies can't get the deleting done
Knowing what to delete and being able to delete it are two different things.
Most companies get stuck for the same set of reasons: too few key people who understand the business, understand AI, and have influence; leaders who read articles but never get hands-on, so expectations drift from reality; overestimating general-purpose models while underestimating the knowledge orchestration complex business workflows need; documents and standards that never got captured as raw material AI can read; and nobody with time to iterate.
The most hidden failure of all has a very fitting name: applause for the demo, no sign of Beta 2. A small team builds a first version, everyone claps in the meeting, guidelines get posted in the group chat, and then nobody actually uses it. New lessons never get written back into the SOP, and version two quietly never happens.
The first action is to drop the slogan "let's make the whole company AI-powered." Replace it with a working version within thirty days: pick a workflow that's high-frequency, repetitive, involves several people, and varies a lot in quality; put one person in charge who understands the business and has the authority; fill in the rules, examples, boundaries, and acceptance criteria; get AI genuinely into the workflow instead of stopping at a chat window; and finally, be clear about who uses it, who checks it, and who writes the new lessons back in.
Reconciliation, quotes, customer service replies, monthly reports — most companies have three to five workflows like this. Pick the one that hurts most first.
The road is fixed. Some roads shouldn't exist.
The last piece was about fixing the road before running AI: when data is scattered across five systems, three formats, and two naming conventions, the answers AI gives you aren't usable. That still has to be done, and it has to be done first.
But once the road is open, you hit a second problem: some roads should never have existed. They were paved only because people used to be expensive, slow, forgetful, and prone to distortion when passing information along. Once those four premises loosen, the roads are still there, but the reasons are gone.
The first layer OAKOrc deleted from its own work was exactly this kind. Projects used to be split three ways — strategy, design, and development subcontracted to different parties — so every handoff meant producing a handover document, holding an alignment meeting, and re-explaining the previous party's judgment. Not a minute of those hours went into design or code. All of it went into covering the cost of "changing hands." One team, from strategy to launch didn't cut anyone's job. It cut an entire layer of steps that existed only for handoffs.
So before you evaluate any AI tool, take one workflow and question it box by box: is this box handling the business, or handling human limits? For the ones handling human limits, first ask whether they can be deleted. Only if they can't, ask whether AI can take them.
You can start on this without buying a single tool. Our AI workflow diagnostic doesn't start from architecture. We walk through your most time-consuming workflows with you, mark which boxes should be deleted, which need to keep a person, and which workflow is worth making your first cut within thirty days.
Let AI move fast. Let people hold the anchor. Tools will keep getting faster, but deciding whether a step stays will always be a person's call.
Rooted in craft. Built for the new wild.