Practical AI for Business Operations: Where to Actually Start

Last updated: August 1, 2026

By Colin Boothe, CIO at Foretopia

Start with the work you would hand a smart intern on day one. High volume, clearly defined, easy to check, cheap to get wrong. Invoice coding. Meeting notes. Support ticket triage. Not strategy, not forecasting, and not anything where a confident wrong answer costs you a customer. AI is fast. It is not careful. You supply the careful part.

A VP of Operations walked me through his AI roadmap last spring. Eleven slides. Four quarters, color coded, with a workstream called Intelligent Enterprise Enablement that I still couldn’t define for you under oath.

Two floors down, a woman named Dana was retyping purchase orders out of PDF email attachments into NetSuite. By hand. About sixty a day, and she had gotten fast at it, which was somehow the saddest part.

Nobody on slide four knew Dana existed.

That’s the whole gap, and I see some version of it in almost every mid-market company I walk into. There’s an AI strategy, and there’s an AI opportunity, and they are usually in different rooms on different floors talking to different vendors.

Who I am and what I sell, so you can discount accordingly. I run an ERP and business systems consulting group. If you finish this and decide your operation needs help, my company is one of the places you might call, and I would take the meeting. The honest offset is that most of what follows costs you nothing but attention, and I flag the parts where you genuinely don’t need someone like me. You’re also reading this on a staffing firm’s site. KORE1 places the ERP and data people who end up doing this work, their IT staffing services practice closes a typical search in about 17 days, and 92 percent of those placements are still there a year later. That number matters later, when I argue that your AI problem is partly a headcount problem.

What Practical AI Actually Refers To

Practical AI for business operations means applying language models and machine learning to specific, repeatable tasks inside workflows you already run, using tools you can buy and configure rather than build. It targets clerical throughput and decision support, not autonomy. The scope is a task, never a department.

Notice what’s missing from that definition. No transformation. No roadmap. No committee.

The reason the definition has to be that narrow is that the broad version has already been tried, at scale, with real money, and it didn’t work. MIT’s NANDA initiative went through hundreds of enterprise generative AI deployments for a 2025 report called The GenAI Divide, and the number that got picked up everywhere, correctly, was that 95 percent of those pilots produced no measurable impact on the P&L. Not a small return. No return.

And the reason wasn’t that the models were bad. The models were fine. The failures clustered around integration and scope, which is a polite way of saying companies pointed a very fast tool at a very vague problem and were surprised when nothing legible came out the other end.

AI Is the High Speed Idiot

This is the model I use with every executive team, and it has survived about three years of contact with reality.

AI is a high speed idiot. Enormously capable. Genuinely fast. Also completely lacking in judgment, context, or the instinct that tells a human being when something smells wrong. It will do the task you described, at a pace no person can match, including when the task you described was the wrong one. It will never stop and ask. That’s the entire risk profile in one sentence.

So you manage it the way you’d manage a brilliant intern on their first Monday. You don’t hand the intern your pricing strategy. You hand them something with a right answer, a clear input, and a supervisor who checks the output before it leaves the building.

Which gives you a test you can run in about ten seconds.

Could you write instructions for this task that a smart intern could follow on day one, without asking you a single follow-up question?

If yes, AI can probably do it, and you should be embarrassed if a human is still doing it in six months. If no, you don’t have a task at all. What you have is a judgment call that somebody dressed up as a process, and if you automate it anyway you will get fast, confident, thoroughly wrong output at a volume large enough that nobody notices the pattern until roughly the end of the quarter.

Three properties make a task safe to hand over. It happens a lot. The answer is checkable by someone who isn’t an expert. And a wrong answer is cheap to catch and fix. Miss any one of those and you’re not automating. You’re gambling with extra steps.

The Numbers Nobody Puts on Slide Four

Here’s what the actual adoption data says, as opposed to what your LinkedIn feed says. I’ve pulled these from government and independent sources rather than from anybody selling software.

Two rows in that table are doing most of the work.

The Federal Reserve’s April 2026 note on AI adoption found that roughly 41 percent of workers use generative AI at work, with about 12 percent using it daily, while only 18 percent of firms had adopted AI as an organization by the end of 2025. Put those two figures side by side and the picture gets uncomfortable. Your people are already doing this. Your company isn’t. Somebody in your finance team pasted a vendor contract into a chatbot last Tuesday and nobody knows about it, and that’s not a hypothetical. That’s just arithmetic on the numbers above.

The Census Bureau’s Business Trends and Outlook Survey puts national use at 19.8 percent, but Information sits at 39.7 percent and Finance and Insurance at 33.9 percent while Retail Trade limps along near 14 percent. If you’re a distributor or a manufacturer, your competitors are mostly not doing this either. That’s a window, and windows close.

Then McKinsey. Their State of AI survey found 88 percent of organizations using AI somewhere, 39 percent able to attribute any EBIT impact to it, and most of that 39 percent putting the figure below 5 percent. Roughly 6 percent qualify as high performers. The high performers were nearly three times more likely to have actually redesigned the workflow instead of bolting a chatbot onto the old one.

So the score is: everybody adopted, almost nobody redesigned, and the value went to the ones who redesigned. Fine. Noted.

Four Jobs You Can Hand It This Month

These are the ones I see work over and over in $50M to $1B operations. None require a data scientist. Two of them your controller can set up without telling IT, which I’m not technically endorsing.

Document intake. Purchase orders, supplier invoices, bills of lading, remittance advice. Anything that arrives as a PDF and leaves as a data-entry job. This is Dana’s job and it’s the single highest-ROI thing most distributors can do, because the volume is enormous, the format is semi-structured, and a human still approves the result before it posts. NetSuite, Dynamics, and Sage all have vendors doing this now, and the good ones learn your specific supplier layouts over a few weeks.

Meeting and call synthesis is the one nobody argues about. Notes, action items, follow-ups, pulled out of an hour of recording in about ninety seconds. It’s unglamorous, it costs almost nothing, most of your existing software vendors are shipping some version of it already, and the project managers on my own team got back something in the neighborhood of four hours a week once they stopped writing recaps by hand. Start here if you want a quick internal win to buy political capital for the harder stuff.

Support and internal ticket triage. Not answering tickets. Routing them, tagging them, and drafting a reply a human sends. There’s a real difference between those two things and most of the failed deployments I’ve cleaned up got that difference wrong. Draft, then a human sends. Never straight through.

The fourth one is code, if you have developers. My own technical team hasn’t grown in headcount in two years while our delivery demand went up substantially, because nine years of accumulated codebase turns out to be excellent context for a model that’s otherwise guessing at your conventions. Your developers already know this. Go ask which tools are on their machines right now. They’ll tell you, and they may be slightly nervous about the question, which tells you something about your policy situation.

What Will Eat Your Whole Quarter

Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, and their stated reasons are escalating costs, unclear business value, and inadequate risk controls. Their analyst was blunter than analysts usually are, describing most of these projects as early-stage experiments driven by hype.

Agents aren’t fake. They are early, and they are being sold four years ahead of where they actually are. Gartner also expects a third of enterprise software to have agentic features baked in by 2028, which is the more useful prediction, because it means the thing you want is going to show up inside the software you already own. Wait for it. Don’t fund it.

Skip these for now:

• Demand forecasting, if your item master is a mess. Garbage in, garbage out, except now it’s wrong at machine speed and has a confidence score attached.
• Anything customer-facing that sends without human review. The math on one bad automated email to your largest account never works out.
• Custom model builds. MIT’s data found that buying from specialized vendors succeeded roughly 67 percent of the time against about one third for internal builds. You’re not going to beat that spread.
• A company-wide chatbot trained on your intranet. Everyone tries this. It answers HR questions adequately and nothing else, and it costs about $200K to find that out.

That last one deserves more space than I’m giving it. It’s the most common expensive mistake I see and it always starts as somebody’s good idea in a leadership offsite.

A First 90 Days That Does Not Require Hiring Anyone

Run this in order. It’s deliberately boring.

Weeks 1 and 2. Find out what’s already happening. Ask every department head which AI tools their team uses. Don’t make it a compliance exercise or you’ll get lies. Frame it as wanting to know what works. Given the Fed’s 41 percent figure, you’ll find more than you expect, and you’ll find at least one thing that should worry you.

Weeks 3 and 4. Count the retyping. Walk your AP, order entry, and customer service desks and count how many times a day a human moves data from one screen to another screen. Write the number down. That number is your business case and you won’t need a consultant to build the slide.

Weeks 5 through 8. Pick one. One process, one owner, one number that has to move. Buy a tool, don’t build one. Set a baseline before you turn it on, because if you skip this you’ll spend the following year arguing about whether it worked.

Weeks 9 through 12. Redesign the workflow around it rather than dropping it into the existing one. This is the step that separates McKinsey’s 6 percent from everybody else and it’s also the step that everybody skips, because it involves telling people their job is changing, and that conversation is harder than any piece of the technology.

Then write down what you learned and pick the next one. That’s it. That’s the program.

The Part Where the Software Is Not Your Bottleneck

I need to be straight with you about something that costs my own business money to say.

About half the AI conversations I get pulled into aren’t technology conversations. The tooling is fine. The tooling has been fine for eighteen months. The company has one overloaded ERP admin, no analyst, and a controller doing three jobs, and there’s simply nobody with the hours or the authority to own a process change from start to finish. You can’t consult your way out of that. I’ve tried, on behalf of clients who really wanted me to, and what happens is you get a beautiful implementation that decays in four months because its owner went back to their actual job.

The tell is easy to spot. If your answer to “who owns this after go-live” is a name that’s already on three other projects, you don’t have an AI initiative. You have a wish.

Fixing that is either a hire or a contractor, and honestly it’s often a contractor first, because you usually don’t know what the permanent role should look like until somebody has done it for a quarter. KORE1 handles that side, not mine, and their contract staffing model is built for exactly this shape of problem. If you want the longer argument about where AI headcount actually goes, they have written it up in their piece on taking an AI pilot to production, and their AI and ML engineer staffing practice covers the deeper builds. For the ERP-side version of this same argument, I made it at length in getting more value out of NetSuite after go-live, and the diagnosis is nearly identical.

What Gets Asked Once the Buzzwords Run Out

Nobody on our team is a data scientist. Are we already out of the game?

Almost none of what I described requires one. Document intake, meeting synthesis, and ticket triage are configuration work, not modeling work, and the people who should own them are your controller, your ops manager, and whoever already administers your ERP. You need a data scientist when you’re building something custom, which is the thing I just spent a section telling you not to do first.

Every vendor is pitching us agents. Should we buy one?

Wrong question, slightly. The better one is whether the software you already pay for is shipping agentic features this year, because Gartner expects a third of enterprise applications to include them by 2028 and most major ERP platforms are already partway there. Ask your existing vendors what’s on their roadmap before you sign with a new one. It’s usually a cheaper answer and a shorter integration.

What should a mid-market company budget for year one?

Under $50K for the first two or three use cases, if you scope them the way I described. Document intake tools run a few thousand a month at mid-market volume. Meeting synthesis is often already bundled into software you own. The budget line that actually hurts is people time, and that one doesn’t show up in any vendor quote.

Our data is a mess. Does that disqualify us?

Depends what you point it at, and this is where most advice gets it wrong. Forecasting and analytics need clean data, so those wait. Document intake doesn’t, because the model is reading a PDF from your supplier and the quality of your item master is irrelevant to that job. Pick the use cases that route around your data problem while you fix it separately.

Half our staff is already using ChatGPT and nobody approved it. Now what?

Get them a licensed account this week and write a one-page policy, in that order. A ban buys you nothing except worse visibility, and you already suspect that. What you’re actually managing is where company data ends up, and the fix for that is a paid tier with the right data terms plus a short written list of what never gets pasted into a prompt.

Six months in, how do I tell whether this paid for itself?

You pick the number before you start, or you’ll never know. Hours on a task, error rate, days to close, tickets resolved without escalation. One number, measured for two weeks before go-live and eight weeks after. If you can’t name the number, you’ve picked the wrong use case, and no amount of enthusiasm downstream will fix that.

Pick the Boring One

Every business I walk into is underutilizing the technology it already pays for. That isn’t a fashionable opinion but it has been true in every single engagement, and AI hasn’t changed it. It has just given everyone a more exciting way to not do the basics.

You don’t need an AI strategy. You need one process, one owner, one number, and ninety days. Then another one. Let’s graduate our tech stack from the 90s to at least the 2000s before we start talking about autonomous agents running the business.

Go find your Dana. She is in your building right now, and she is fast at something no human should have to be fast at.

Want to argue with any of this, or get a second opinion on a use case before you sign something? Hit me up on LinkedIn. If the gap you found turns out to be an empty chair rather than a software problem, talk to the KORE1 team. They fill these roles across 30-plus U.S. metros and their recruiters average 15-plus years in the seat, which is more institutional memory than most internal teams have on the subject.