The Complete Guide to What AI Actually Costs a $5M to $100M Company

Ask a founder running a $20M company what AI costs him and he quotes his seat count. Forty people on paid chat plans, a batch of Copilot licenses, done.

That number sits on the surface. Underneath it sit three bigger line items that never make the software budget: the metered usage, the people who run the systems, and the pilots that die quietly with money still in them.

I run AI and growth at NuVision Auto Glass, a $48M US auto glass company, and my growth company builds AI systems for US brands between $5M and $100M in revenue. The same four layers show up in every budget at this scale, and the one everybody tracks is the smallest.

One caution before the numbers. Prices in this category move fast, and vendors change tiers every few months. Treat every figure below as a checkpoint from July 2026, and confirm the current page before you budget anything.

Layer 1: the seats, the line you already track

The per-person subscriptions. Consumer plans on the major chat tools run about $20 a month per person. Team and business plans land in the $25 to $30 per user range, and give you the data protection, admin control, and shared workspaces that personal plans don't. At your size, the business tier stops being optional. You have contracts, customer data, and a brand to protect, and personal accounts put all three on someone's private login.

Microsoft's Copilot for 365 sits around $30 per user per month at list. Its real cost runs higher than the sticker, since it requires an underlying Microsoft 365 subscription. When you compare it against a standalone tool, compare the total.

Multiply it out and the line stays small next to revenue. 40 seats at the $30 tier cost $1,200 a month, $14,400 a year. Against $20M in revenue, that rounds to nothing. Which is exactly why most founders stop reading the budget right there.

Two rules keep this layer clean. First, pay for daily users only. Your admin panel shows last activity per seat. People who touch the tool twice a week work fine on a free tier, and you revisit the split every quarter. Second, hunt duplicates across departments. The pattern repeats at every company I've seen: marketing signs up for one chat tool, ops signs up for a second, and finance discovers two vendors doing one job six months later. One concrete check, run today, closes it. Pull vendor names from your card statements and AP ledger, group them by function, and merge every pair.

Seats are rent. Predictable, visible, easy to cap. The next layer behaves nothing like rent.

Layer 2: the meter

An API is the pipe your software uses to send work to a model and get an answer back. You pay per unit of work. If seats are rent, API usage is the electricity bill, and it climbs with volume whether or not anyone watches it.

This layer barely exists at a 5-person shop. At your scale it becomes real, because the systems worth building touch every record. A chat widget answering questions on your site. A workflow that summarizes every inbound sales call. An automation that scores every lead in the CRM. None of those bill per person. They bill per event, and your event count grows with your revenue.

The automation platforms that carry this traffic, Zapier, Make, n8n, or the automation inside your CRM, start between $20 and $100 a month on entry plans. That figure describes the door, not the room. Task volume drives the real bill, and a company processing thousands of orders or calls a month sits nowhere near the entry tier. Before you pay for any connector, check what you already own. Your CRM, your form tool, and your accounting software probably include automation you're already paying for, and plenty of companies buy a connector to link two products that already link natively.

The meter also fails in a way rent never does. A workflow stuck in a retry loop burns through a month of budget over a weekend, and the invoice tells you about it two weeks later. So treat every API key like a company card. Put a monthly cap on it the day you create it, set a spend alert at half the cap, and give the whole meter one named owner who reads the usage dashboard weekly. One concrete example of the habit: at NuVision, no automation goes live without a cap and an owner attached, because the one time a loop runs unwatched pays for years of the discipline.

Layer 3: the people who run it

The layer missing from every software budget, and at your scale, the biggest one.

Every system from the first two layers needs a human attached. Someone maintains the prompts, the written instructions the model runs on. Someone fixes the workflow when a vendor changes their product, which happens every few months. Someone retrains the team, checks output quality, and manages the vendor relationships. That adds up to a fraction of an ops manager, a fraction of an engineer, and a fraction of you.

At NuVision, the time my team and I spend running and improving our AI systems costs more than the tools themselves. That ratio holds at every company I've helped at this scale. The tool line looks like the budget. The payroll fraction behind it is the budget.

The build cost lives here too. Standing up a real workflow takes either your people's hours or an outside project with a beginning and an end, and the ongoing subscription stays separate from both. Founders who skip this line item end up owning shelfware. A tool with no workflow around it just keeps billing you every month.

The rule that protects this layer: every tool gets a named owner and one number it moves. A concrete version, taken straight from how the failure looks in practice. A department head champions a new tool, builds the workflow, then leaves. Nobody inherits it. The output quality slides for a quarter, the team quietly routes around it, and the subscription bills for another year. A named owner in a shared doc, reviewed quarterly, kills that story before it starts. No owner means cancel it.

Layer 4: the failed pilots

A pilot is a small trial of a tool or system before you roll it out company-wide. At your scale you run several a year, and most of them die. Budget for that instead of pretending otherwise.

The evidence sits in public now. MIT's NANDA initiative studied 300 public AI deployments, interviewed 52 executives, and surveyed 153 leaders for its 2025 report, The GenAI Divide: State of AI in Business. It found that 95% of enterprise generative AI pilots produced no measurable P&L impact. The report blames poor integration into real workflows, not weak models. Tools that never learn your process stall, however impressive the demo looked.

The money leaks in a familiar shape. A pilot starts with energy, a vendor contract, and no exit criteria. The trial becomes a subscription. The subscription outlives the problem that justified it, and a tool bought for one campaign keeps billing for a year. A five-person shop leaks a forgotten $20 chat subscription this way. A company your size leaks a vendor contract and three months of an engineer's time.

3 rules turn pilot losses from a surprise into a line item.

1. Every pilot gets a kill date, written before the contract starts. On that date it graduates to a funded rollout with an owner, or it dies. No third state. 2. Every pilot names one number it must move, decided upfront. The MIT finding gives you the reason: pilots fail on integration, and a pilot that touches no real workflow number skipped integration from day one. 3. Budget pilots as R&D. Set a fixed annual amount you expect to mostly lose, because the winners pay for the graveyard. Founders who fund pilots ad hoc feel each failure as waste. Founders who fund them as a pool feel them as the price of finding the two that work.

How to budget the whole thing against a hire

Now put the four layers together, because the total only becomes useful next to something.

The right comparison is your next hire. You already know how to judge headcount. You look at the fully loaded cost, you name what the person will own, and you check in ninety days whether the hire moved the number. AI spend deserves the same test, and at $5M to $100M it has quietly become the same magnitude as a hire. Forty seats, real API volume, the payroll fraction that runs it all, and a pilot pool add up to a decision-sized number, and nobody judges a pile of subscriptions the way they judge a decision.

Run the exercise this week. It takes an afternoon.

1. Pull card statements and the AP ledger and list every AI charge, including seats your team expenses on personal cards.
2. Sort every line into the four layers: seats, meter, people, pilots. Estimate the people layer honestly, as fractions of real salaries.
3. Cancel anything with no login in 60 days. Admin panels show last activity.
4. Give every survivor a named owner and one number it moves. Write both down. No owner, no renewal.
5. Cap every API key and set spend alerts the same day.

6. Write the four-layer total as one annual figure and put it on the same page as your hiring plan. Review both on the same quarterly cadence, and ask of each: what did this produce that the other wouldn't.

That last page is the whole point. The founder who quotes his seat count is budgeting the smallest of four layers and skipping the comparison that matters. Price the full stack the way you price a person, and AI stops being a category you're vaguely for. It becomes a role on your org chart that either earns its cost or loses its seat.

If you want the whole stack reviewed against what it actually returns, with a plan for what to keep, cut, and build, we do that at NuroSparx. See what a proper setup costs or send us your current stack and we'll tell you where the waste is. Most of the waste hides in layers 3 and 4, and cutting it funds the first automation that earns its own line on the sheet. The four-layer total takes one afternoon, so run it before the next renewal hits your card.