How to Tell If You’re Being Sold AI Snake Oil (and What It Actually Costs)

Every AI vendor sells one level of AI investment. Here’s what each of the four costs a manufacturer to own, and the answer nobody in the room gets paid to give.

Marko GrahovacGrowth ManagerPublished 2 October 2026 · 13 min read

01

You already run this decision every week

A customer sends you a drawing for a part you've never made.

Most of what happens next happens before anyone says a number. Somebody checks the quantity. Somebody sees if it runs on what you already have or whether a fixture has to be built first. Somebody reads the tolerance on one feature and realises that changes which machine it goes on. Then somebody decides, without anyone calling it a decision: we run this ourselves, we buy it finished, or we send it out.

At forty pieces a year, you don't tool up. You find a shop that already has the setup and you take the margin you can get. Nobody in the building will treat that as a defeat.

Then somebody arrives with AI, and none of that happens. There's a demo on data that isn't yours, a six-week timeline, and a monthly figure. The only question on the table is whether you want the thing they walked in with.

I've sat on the selling side of that meeting, because my firm is one of the vendors selling AI to manufacturers. This piece is about how the meeting gets built. The person in it usually believes what they're selling.

Sourcing a part comes down to a handful of possible answers, and you choose between them on things you can count. What's the volume. Is this something we want to be good at, or is it a one-off. Can it run on equipment we already own, or does it need new tooling, and how long until that's paid for. Who maintains the setup once it exists, and what happens when the customer revises the drawing.

You do that without calling it a framework while software people argue about the same decision under a different name, build versus buy, and mostly with less discipline than you bring to a part.

Point those same questions at an AI purchase and you get the same shape of answer. There's a version that runs on what you already own, a version your own people put together, and a version you have built. And there's the forty-pieces-a-year answer: don't tool up for this one yet.

The gap between the answers is large. MIT's NANDA initiative, reporting in July 2025 on interviews with 52 organisations, describes a company that spent $50,000 on a specialised contract analysis tool and then watched the lawyer it was bought for go back to a general-purpose chatbot for the same work. This is what they concluded: "a $20-per-month general-purpose tool often outperforms bespoke enterprise systems costing orders of magnitude more, at least in terms of immediate usability." Their findings are preliminary and self-reported, and I'll set out the rest of the caveats where I lean on them harder later. As a picture of the spread it stands: same job, two answers, three orders of magnitude apart.

I'm assuming you've decided to spend something on this. Whether you should be spending at all is a longer question, and I went through this properly in the "Should You Invest In AI: Honest answer for Manufacturers" piece, so I won't re-argue it here.

02

Every AI purchase sits on one of four levels, and one answer comes before all of them

Our CEO Dušan Stamenkovic laid out the four levels of AI investment in full, for any enterprise company. Here's how they look from a plant.

LevelWhat it looks like in a plantWho sells it to youWhere the cost of owning it sits
1. The subscription as boughtCopilot or ChatGPT seats, used as an empty chat windowThe assistant vendor, per seatIn the subscription
2. The configured subscriptionThe same seats, connected to your SOPs, work instructions and quality procedures, with permissions set and a short list of uses per roleRarely sold at all. The seats are already paid forWith whoever owns the setup, every year it runs
3. The DIY TrackYour own people build a specific tool on one or two sources, with a no-code builder or an AI coding assistant. A quote-prep checklist off the ERP export, sayAutomation platforms and the consultancies that build on themOn the person who built it, usually without anyone deciding that
4. The Custom Build TrackEngineers build a production system that can reconcile several systems that disagree, write into your records, or answer customersDevelopment firms, us includedIn plain view, as engineering time that never stops

The answer people usually miss comes before Level 1, and it's "not yet," and volume usually decides this one. If the process you'd point this at runs twice a week, there's nothing there to earn back a setup cost, the same way there's nothing in tooling up for forty pieces. That's arithmetic about one process, and it can change next quarter when a customer doubles an order.

"Not yet" is narrower than it sounds. It means don't build or buy anything new for that process yet. AI can still earn its keep elsewhere in the same building, and the seats you may already pay for are the place to start. A quarter spent measuring the process and getting real use out of what you have puts you in a better position to choose a level when the volume shows up.

Anyone giving you this answer honestly will also tell you what would have to change for it to become a yes. If you can't say what the process costs you today, measuring it comes first, ahead of choosing a level. That measurement is what an opportunity and readiness assessment produces, and it costs a fraction of whatever comes after it.

Two of the levels get skipped, for the same reason.

Level 2 is where most of the value of seats you already pay for sits, and it's the level most companies jump straight past. In a plant it's the difference between a chat window that knows nothing about you and one that can find the current revision of a work instruction and tell you who owns it. The seat vendor gets paid whether that work happens or not, so nobody shows up to sell it to you.

Level 3 is the one almost nobody puts in front of a manufacturer, because there's no licence revenue in recommending it. For a lot of office processes it's genuinely enough. It also has the least understood cost structure of the four, which I'll get to.

03

Everyone selling you AI is a paid advocate for one level

A seat vendor will tell you more seats solve it. An automation consultancy will tell you to build it on their platform. A development firm will tell you it needs building properly. Each of them is right some of the time, and each is describing the one level they get paid for.

Nobody shows up to tell you it's "not yet."

We do Level 4. Commissioned builds are the dearest of the four, so discount what I say about them and read the next section carefully.

There's no outside referee on any of this either. Regulators have been busy on AI claims, and none of it reaches you: the cases are consumer products and statements made to investors. No industrial trade body has a code on it. So you're the only check on whether the pitch in front of you is honest, which is why knowing the four levels matters more than learning to read a salesperson.

04

What kills you is the cost of owning it, and your tooling already taught you that

Sourcing decisions rarely go wrong on the purchase price. The question really is what it costs to own the capability afterwards. The fixture that needs re-machining, the one person who knows how the setup is dialled in, the requalification when the customer revises the drawing. You already price that in without being asked.

Nobody prices it into AI, and the number that gets quoted for it is wrong.

You'll hear that maintaining a system you've had built runs 15% to 20% of the build cost a year, usually credited to Gartner. I couldn't find a Gartner document behind it. Every citation I followed led to a software agency's blog citing another agency's blog. Stop using it.

The figure that does still hold up today is far less comfortable. An older study by Robert Glass, in IEEE Software, put maintenance at 40% to 80% of a software system's total cost over its life, around 60% on average. Put plainly, keeping it running costs about one and a half times what it cost to build. It covers software in general rather than AI, so read it as a rough scale.

Level 1. Buried in the subscription and the renewal. You're renting somebody else's maintenance, usually the cheapest way to buy it, but you can't change the product and the price isn't yours to set.

Level 2. Somebody has to keep the documents current, the permissions right, and decide which version of a procedure is the real one. That's an owner and a budget line, and none of it ends.

Level 3. It sits on one person's shoulders. One practitioner auditing a small school's automations found thirty workflows across two platforms, a few broken for weeks, one still sending new leads to a colleague who had left. It's one account rather than a study, but you'll recognise it: the undocumented fixture only one setter understands.

Level 4. In plain view as engineering time, and it never stops: watching whether it still works when your product mix moves, retraining when it doesn't. A system that loses that budget line quietly stops being used. Keeping it earning costs something every year.

05

Building is the right answer less often than it's pitched

Which raises the obvious objection. If a subscription is cheapest to own and a build is dearest, why would anyone build?

Because a product can only solve a problem that lots of companies have. The workflows where money actually sits in a shop your size tend to be the ones that are yours alone: how you price work, what your estimators know that isn't written down anywhere, the sequence of judgment calls between an enquiry arriving and a quote going out. When the answer has to pull from your ERP, your old job files and your drawings, and those three disagree, no seat and no weekend workflow fits it. When one of those workflows carries enough money, building is the right answer.

Even when building is the right call, many attempts don't make it. NANDA found that only about 5% of enterprise-grade AI systems, custom or bought, reached production at all, and that tools a company built in-house got there roughly half as often as ones brought in with an outside partner. The authors also say the gap may come down to what kind of company chooses to build in-house.

Read the full case here →

06

Price it at the volume you'll run if it works

Say the offer is $4,000 a month. I made that up, because nobody publishes pricing for this category, which tells you something on its own. Best if you put your own number in.

At $4,000 a month that's $48,000 a year, $144,000 over three. At most companies, $48,000 a year of operating spend gets signed inside a department's budget. The same $144,000 asked for as a capital purchase goes up to finance, and at a lot of family-owned shops to the owner. Monthly pricing keeps the decision with the person least likely to ask the hard questions, and that's no accident.

Then add what use costs. The DIY Track is the confusing one here, and also the only level where you can check the arithmetic yourself in an afternoon.

The catch is that every automation tool counts usage differently, so their price lists can't be compared side by side. Take one process with six steps, run once. n8n counts that as one run. Make counts each step, so six. Zapier counts five, and more if one of the steps uses AI. Same work, three different bills. Pricing in this industry also changes often, and that is a whole different challenge for long term planning.

Two things there will catch a company your size specifically. n8n's Business tier, at €667 a month billed annually, is described on their own pricing page as being for companies with fewer than 100 employees, which is smaller than almost anyone reading this, and the features your IT lead will insist on start at that tier. And since November 2025, Make charges 25% more for credits bought beyond your plan than for the ones inside it, so the bill accelerates exactly when the thing starts working.

That holds beyond Make. Usage pricing means the bill grows when people adopt it, which is the same event as success. We design usage pricing for software clients ourselves, and there's nothing dishonest in it. It does mean the only honest way to price an offer is at the volume you'd be running if it worked, which is rarely the volume in the pilot.

Doing that properly against your own numbers, before any capital moves needs feasibility and ROI validation. It takes weeks rather than months. Running it before the spend matters more than which level you end up choosing.

07

So, how do you tell?

You won't tell by reading the person selling to you. They may be entirely sincere and still be wrong for you, because they sell one level and you have four plus "not yet."

Work out which one your situation needs before the next meeting.

If you can't say what the process costs today, that's your answer for now, and it's the cheap one.

If you're already paying for seats, configure them before you buy anything else. Connect them to the documents your people actually look things up in, and give each role a short list of things to use it for.

If the workflow is yours but it runs on one or two sources and a person checks the answer, someone in your own building can probably build it. You won't hear that from the people who profit from the alternative.

If one workflow carries real money, depends on systems that disagree, and nothing on the market fits it, build it, and put the cost of owning it in the request from the start instead of discovering it in year two.

At Level 4 the answer often reaches a customer or changes a record before a person reads it. So ask for one thing before money moves: run it on my data, on my worst month, and show me where it gets things wrong. A seller who has built this before will hand you the failures without being asked, because they've seen them. A seller who can't tell you how often it's wrong, how you'd know, and who checks, is telling you they haven't run it anywhere that matters.

You've been pricing make-or-buy decisions your whole career, and you're better at it than the people pitching you are at theirs. What's been missing is anyone willing to lay out every level when they can only be paid for one.

Do that once and the next meeting runs differently. You're sourcing on your own terms, which is work you've done a thousand times.

One last move, and it points the other way. Before you go looking for holes in their proposal, go looking for them in your own brief, because every line of yours that stays vague is a line somebody else gets to price. "Improve quoting with AI" can be priced at almost any number, and it will be.

Growth Briefings · for business-case ownersMore in Insights →

Talk to the partner who leads the work.

Schedule a call with a partnerFree until 11 December