The Cost of AI

Beating a dead horse

Truth be told, this horse couldn't have been dead for that long either. The whole Agentic thing is no more than a couple of years old, right? Just think about your personal first touch with the Agent - was it a Midjourney fling back in 2024? I wasn't using Claude Code well into 2026. So, the point I want to make here is fairly new - at least to me. Even though it feels so obvious and so redundant and so outdated at this point:

AI is a tool, don't let it take away your creativity.

And why is that? Because I am (you are) lazy. We are optimization machines; we know how to cut a corner if we see one. But a lot of times, the point is that things need to be hard, to be worth it. Learning, for example - it turns out that when the process feels hard, that's when it really matters, that's when learning occurs. When you sweat. When you suffer. And what's creativity if not a constant struggle, a battle with the unknown, with the uncertainty. The fear of being judged for being novel. And what is AI? A pattern regurgitation algorithm that gives you the mean result, the average. Not only does it have to exist already out there, somewhere, but it kind of is the middle of the road.

As a result, I am now skipping any post that reads like:

"It's not [X], it's [Y]. Meta commentary: [exaggeration]. It's the [expression] of [something]. It's [exaggeration], and that [is important]."

Especially if something in that post has just "landed" somewhere "quietly". Because this tells me 2 things about the person behind the post:

• They are lazy
• They are scared

And there's little I want to learn from such a person.

What I do want to see is that someone spent actual time! Has suffered, has sweated, has cared for the product that came under their fingers. Even if it's flawed, even if it's imperfect.

Another favourite phrase of my dear ChatGPT - "And why is that important"? Well, I will tell you why that's important in the context of the actual Cost of AI. Again, nothing new, just spelling out.

What are we paying for?

This took a detour, but here we go - I am not claiming AI is of no use, or harmful or should go away. It's like arguing that electricity is bad for us, that roads are harming the environment, that internet was an invention of Satan.

AI is infrastructure

That's my hot take. It's not that hot either ... not even lukewarm. But blaming infrastructure is like blaming the weather - it's pointless, yet you can't help it, now can you mate.

So, it's really the cost of that infrastructure I am discussing here. And it's up to you what you use yours for and if you have the money to support it. You can look at funny cat videos all day every day, but the moment they start charging per kitty token, then .. well you'll swap over to dogs, if that's still free.

From seats to usage ... to ?

It used to be that we pay once and we own the damn thing. We then started paying for the update - that's fair, someone works to make it happen after all. Then what? Right - subscription time - and I don't need my click-to-cancel-you-Adobe, you filthy predator. But ... we found the value in that as well, didn't we? So, sure thing. Give me them 10 seats of your best software, kind Sir, and thank you very much! I can now do a yearly budget that looks scary, but at least somewhat predictable.

Enter usage. 5 tokens. 100'000 tokens. Your cooldown period ends in 5 hours.

I don't want to make a huge literature point around this - it's as simple as "you cannot predict your usage". You get cut off mid-stream, you can't finish what you started, or at least not in time. And you become vulnerable to a payment model you have no control over. I hope I don't need to explain why that matters.

The thing is, much like before, the price is justified .. in a way. Someone's got to pay the piper, at the end of the day. Autodesk hinted at this: about 17% of its revenue already comes from usage-based models, and its CFO has said heavier AI automation will be priced on usage too (paraphrasing, check the source for the exact wording and meaning). But that's not the end of it. Autodesk is changing the Data Model API pricing too - from October 6^th, anything pulled out of the AEC Data Model beyond your subscription's allowance gets metered (Data Model APIs included with subscriptions, plus flexible ways to scale). That cost will have to fall on us, the users. Because this is how it goes:

We pay Autodesk -> Autodesk pays Amazon for AWS / Amazon Bedrock -> Bedrock serves the frontier models behind the Autodesk Assistant

The main problem, the way I see it, is when those frontier models start tightening the rope. The above cascade will remain, and our already uncertain bill will balloon out of proportions.

I mean, of course, it cannot be that catastrophic, and I am sure the market will ultimately figure itself out. Honestly, there isn't an alternative - if companies start going bust because of their AI bills ... that's not really a plausible scenario anymore, now, is it?

Is AI the new UI?

My main point is this - don't be silly. Don't use AI for things that you can use traditional toolage for. If there is a button for it - use the button!! Don't type it in a chat box if you can clickity-click-click your work in the same amount of time.

In that regard, distrust AI wrappers that build on existing functionality. Anything you can already do and doesn't take much time - just skip the AI offer altogether, not worth it! On the other hand, if AI is indispensable to a novel workflow you simply can't do without - no brainer, use that AI!

A short glossary of AI terms

• Frontier model - the biggest, most capable general-purpose AI models of the day (Claude, GPT, Gemini and co). Expensive to build, expensive to run, and what most AI features are calling under the hood.
• Token - the unit AI models read and write in, roughly three-quarters of a word. You pay per token, in and out. Not to be confused with Autodesk Flex tokens, which are a billing currency.
• Context (window) - how much a model can "hold in its head" at once, measured in tokens. Bigger context, bigger bill.
• Inference - running a model to get an answer. This is the compute you pay for every time you hit enter.
• Agent - an AI that doesn't just answer but acts: calls tools, edits files, loops until the job is done. Burns far more tokens than a chat.
• MCP (Model Context Protocol) - an open standard that lets AI models plug into software like Revit. The protocol is free; the servers built on it may or may not be.
• API - the door software uses to talk to other software. Increasingly, the door has a turnstile.
• Usage-based (consumption) pricing - paying for what you use instead of per seat. Fair in theory, hard to budget in practice.
• Credits / Flex tokens - vendor currencies you prepay to cover usage. Mind the expiry date: Autodesk Flex tokens expire 12 months after purchase.
• AI wrapper - a product that puts a chat box over existing functionality or over someone else's model.
• Vibe coding - building software by describing what you want to an AI and iterating, rather than writing every line yourself.
• Hallucination - when the model confidently makes something up.

My rule of thumb

Autodesk Assistant (AA)? Yes, sir! Ships with Revit, folded under your existing bill (for now) - use, use, use!

That new tool I saw a presentation of that bundles 3 plugins, but has an Agentic interface - hard pass.

Use it if it's 10 times faster, if I can trust the output, and if I can double-check it.

Good use cases:

• (vibe) coding - absolutely outstanding, but you have to remain the architect
• Research - deep or wide
• Document formatting

Bad use cases:

• Writing emails, social media, blog posts
• Design
• Ideation and reasoning

Anything that serves me and saves my time - yes! Anything facing outward - nah, not so fast little agent.

Why is that? Because, for heaven's sake, what are we even doing here? The freaking Matrix? I didn't create a life and a career to be lived by some swarm of agents. I want my life to be lived by me, fully. That includes any creative endeavours, and mostly my interactions with others.

If anything, the last couple of years have made it easier to come to terms with my personal imperfection. I don't know about you out there, but this kid right here - far from a finished product! I bring so much crap with me, all the time. And my emails suffer from it. And my social persona suffers from it. And so, I burn bridges. But at least I get to live my version of this life.

A smart way to use AI in the AEC industry

The way forward is this, in my opinion - use AI as an agent of democratization of code creation. Where possible, automate your own workflows using the frontier model of your flavour, to the expense you are comfortable with. Vibe code. Create workflows. Connect APIs in a way that you wished you could in the past but didn't have the capacity to do, or the time, or money. Wrap it in a robust old-school software, cloud-based or otherwise. Make sure you know your long-term spend, once the AI layer is no longer there. And have fun doing it!

If you want to know more about how we at Archilizer use AI, follow us on LinkedIn, or subscribe to our Newsletter. We shall soon reveal what we've been tirelessly working on in the last months. Spoiler alert - it's agentic.