Your 2027 budget has an AI line. Here's what should be under it.

The budget review is on page six. Facilities, fleet, IT. Then a new row somebody added after the last board meeting: "AI tools." No vendor, no owner, just a number with a question mark next to it.

Two rows up is the software subscription you bought last spring after a good demo. Everyone got a login. Pull the usage report and you'll probably find a handful of people still opening it, and none of them can tell you which step of the job it changed. It renews in January.

Why that row gets forgotten

AI spending this year mostly comes from pressure, not from a problem. Somebody asks "what are we doing about AI?" and the easiest answer is a subscription. It costs little enough to approve without a fight, which also means nobody fights to make it work.

The real cost isn't the license fee. It's that the constraint you actually have, the order desk that can't keep up or the dispatcher who's maxed out, is still there next year. When volume grows next year you're back to the same problem, posting the same job, and now your people have quietly decided that "we tried AI and it didn't do anything."

How to tell a capacity problem from a subscription

Before anything goes under that line, I'd put each idea through five questions.

1. Can you name the person and the step? Not "sales" or "operations." The customer service rep who builds the order confirmation by hand, three times a day.
2. Does it grow with volume? If you double orders, does this work double? That's what makes it a capacity problem. Work that stays flat as you grow is rarely worth building for.
3. Is it rules or judgment? Matching invoices to receipts, re-keying an order from email into the ERP, chasing missing paperwork: that's rules. Deciding which customer to push back on is judgment. Build for the first. Support the second.
4. Can you show what "done right" looks like? If you can pull ten past examples and say "this is correct," a tool can be tested against them. If nobody agrees what correct is, no software fixes that.
5. Does someone own it? One person who will notice when it's wrong and has the authority to fix the rules.

Five yeses is a candidate. Two or three is usually a subscription, a process fix, or nothing at all.

What I'd put money on, and what I wouldn't

Put money on finding out first. Map where your team's hours actually go on the handful of processes that grow with volume, in writing, before anyone picks a vendor. That map tells you which one or two problems are worth building for and which aren't.

Then put money on one working prototype, at a fixed price, run against your own real documents. Not a demo on the vendor's sample data. Your purchase orders, your messy emails, your weird customer. If it works on last month's actual work, fund the full build. If it doesn't, you've spent a small amount to learn something true.

Make sure you own what gets built. Code you own can be changed, moved, and maintained by someone else if your vendor disappears. A platform you rent can change its price or its features without asking you.

What I wouldn't fund:

• Seat licenses for everyone "to experiment," with no named problem attached.
• Anything that needs a data team you don't have to keep it running.
• A multi-year contract before you've seen a working version on your data.
• A custom build for something your current ERP or CRM already does. Check this year's release notes first. Gartner notes that enterprises are using the simpler AI features built into software from their existing vendors, and for plenty of problems that's the right answer and the cheapest one.

Buy it the way you'd buy a press brake or a new forklift. Spec it against the job. Trial it on your material. Agree on what acceptance looks like. Know who maintains it.

What the spending numbers say

The money is moving fast. Gartner forecasts worldwide AI spending of about $2.7 trillion in 2026, up 49.5% from last year, while placing generative AI "firmly in the Trough of Disillusionment." That's Gartner's phrase for the stretch where the hype has worn off and people start asking what they got.

Plenty of projects won't make it. Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.

And if you feel behind, you probably aren't. The US Census Bureau found that as of May 2026, 32% of firms with 100 to 249 employees and 37% of firms with 250 or more reported using AI in their operations. Most companies your size haven't started. Nobody's going to win this by spending first. You can win it by being specific.

Try this before your budget meeting

List every software renewal coming up in the next six months that was sold with "AI" in the pitch. For each one, write two things: the specific step of work it changed, and the name of the person who'd notice if it disappeared. Any line where you can't fill in both is a candidate to cut or renegotiate, and that money can go toward something you can name.

One next step

If you'd rather walk into the budget meeting with a map than a guess, that's what the Growing Without Hiring workshop is for. It's a half-day diagnostic with your team that ends in a written capacity map: where the hours go, what's worth building, and what isn't. Then the AI line in your budget has something specific under it.