Mullin  /  Articles  /  AI Project Stages

Field Notes · Building with AI

The five stages of an AI project, from start to finish.

Most AI work doesn't fail at the model. It fails at the sequence — teams reach for the revolutionary before they've earned the elementary. Here's the order a system actually has to climb.

M Mullin Agency June 5, 2026 9 min read The "-ary" Maturity Sequence

Every AI project we've ever shipped — and every one we've been called in to rescue — was somewhere on the same ladder, whether the team knew it or not. The ones that worked climbed it in order. The ones that stalled tried to skip a rung: a board demanding something revolutionary from a system that didn't yet have a reliable elementary. The model wasn't the problem. The sequence was.

We call the ladder the "-ary" Maturity Sequence. It's five stages — Elementary, Supplementary, Exemplary, Visionary, Revolutionary — and the rule is simple: each stage is load-bearing for the next. You can't add value to a foundation that doesn't exist. You can't set a gold standard on something that isn't yet reliable. And you certainly can't reshape an industry with a system your own team doesn't trust on a Tuesday afternoon.

What follows is how that sequence plays out specifically for an AI build — what each stage demands, what "done" looks like, and the signal that tells you it's time to climb.

Ambition isn't the problem. Skipping rungs is.

The climb · one rung at a time Elementary → Revolutionary
Elementary 01 · Foundation Supplementary 02 · Add-On Exemplary 03 · Gold Standard Visionary 04 · Strategy Revolutionary 05 · Ultimate Shift

Each rung is load-bearing for the next. The summit isn't a starting point — it's what four finished rungs add up to.

The Sequence

Five rungs, in order.

01
Elementary · The Foundation

Make it work once.

The absolute, must-have baseline. Without this, the system cannot function — so nothing else on this page matters yet.

Elementary is the unglamorous core: one clearly-defined problem, the data that actually feeds it, and a single path from input to a useful output. Not a roadmap — a working spine. For an AI build that means a real use case (not "AI for the business"), data you've inspected and can trust, a model or prompt that produces a usable answer, and the plumbing to get that answer in front of one real user.

The discipline here is subtraction. Most failed AI projects are over-scoped Elementaries — five use cases, none of them finished. Pick the one where a correct answer is obviously valuable and a wrong answer is survivable. Ship that. The goal isn't impressive; it's true.

What "done" looks like

One use case produces a correct, useful result end-to-end for a real user — repeatably.

Signal to climb

People stop asking "does it work?" and start asking "can it also…?"

02
Supplementary · The Add-On

Make it trustworthy.

The helpful extras. Not strictly required to function — but they're what turn a working demo into something people actually rely on.

Once the spine works, you layer on the things that make it genuinely useful: guardrails, evaluations, a feedback loop, error handling, the UX that lets a non-expert get value without reading a manual. None of these are the AI — and all of them decide whether anyone keeps using it.

This is where evals earn their keep. A handful of test cases you score on every change is the difference between "it felt better" and "it's measurably better." Add the small integrations — it writes to the doc, posts to the channel, updates the record — that remove the copy-paste tax. Supplementary is where adoption is won or lost.

What "done" looks like

Real users return without being told to, and you can prove quality is improving, not drifting.

Signal to climb

Demand outpaces what the current setup can safely handle.

03
Exemplary · The Gold Standard

Make it excellent.

High-quality execution that sets a prime example — the move from "it works" to "it works so well people point to it as how this should be done."

Exemplary is where craft separates a tolerated tool from a beloved one. The system already functions and is trusted; now you raise the bar on every detail. Latency that feels instant, answers that are not just correct but well-shaped, a reliability record people stop worrying about, and an experience polished enough that new users get it without a tutorial. Nothing here is strictly required to function — and all of it is what makes the system the standard others measure against.

This is also where the quality bar becomes deliberate rather than accidental: tighter evals, edge cases handled gracefully, errors that explain themselves, observability that catches drift before anyone notices. Teams that rushed past Supplementary feel it most here — you can't set a gold standard on a shaky base. Exemplary is the rung where the system earns its reputation.

What "done" looks like

Quality is high and consistent enough that the system becomes the reference point — internally, or in your market.

Signal to climb

It's so good at today's job that the question becomes where it could go next.

04
Visionary · The Strategy

Make it a moat.

Forward-thinking design that anticipates future needs and puts the system ahead of competitors — instead of merely keeping pace.

By now the system works, is trusted, and scales. Visionary asks a sharper question: what compounds? Usually the answer is proprietary data and feedback — the loop where every interaction makes the next one better, in a way no competitor can simply buy off the shelf. You stop thinking in features and start thinking in platform: shared capabilities other teams build on top of.

This is the strategic rung, and it's where a generic AI tool diverges from a durable advantage. The work is as much positioning as engineering: where is the industry heading, what will users expect in eighteen months, and how does this system put you there first? Visionary is the difference between adopting AI and owning a position with it.

What "done" looks like

The system creates compounding advantage — proprietary data, network effects, or capability rivals can't quickly copy.

Signal to climb

The advantage is real enough to reshape how the work itself is done.

05
Revolutionary · The Ultimate Shift

Make it change the game.

A groundbreaking transformation that changes the category — not a better version of the old way, but a different way entirely.

Revolutionary is rare, and it's earned, never declared. It's the point where the system doesn't just help you do the work faster — it changes what the work is. The org chart shifts around it. Customers expect things that weren't possible before. Competitors are now responding to you. Very few projects reach this rung, and none of them reach it by aiming at it from day one.

The uncomfortable truth: every revolutionary system was, not long before, a deeply unglamorous Elementary that someone refused to skip. The transformation is the visible tip of four rungs of compounding, boring, load-bearing work. You don't build the revolution. You build the ladder, and the revolution is what's standing at the top.

What "done" looks like

The system redefines expectations in its category — and the rest of the field is reacting to you.

The catch

It's only reachable by climbing the four rungs below it, in order.

The sequence, in one breath
  1. 01Elementary — one use case, working end-to-end and true.
  2. 02Supplementary — guardrails, evals, and UX that earn trust and adoption.
  3. 03Exemplary — craft, consistency, and polish that make it the standard to beat.
  4. 04Visionary — proprietary loops and platform thinking that compound into a moat.
  5. 05Revolutionary — the system changes what the work is, and the category follows.

Where most teams actually are

If you're being honest, most "AI initiatives" are stuck between Elementary and Supplementary — a promising demo that never became dependable enough to scale. That's not a failure of ambition or talent. It's almost always a sequence problem: pressure to look Revolutionary pulled attention away from the unglamorous rungs that make Revolutionary possible.

The honest distribution

Most AI work clusters at the bottom two rungs.

Roughly two in three never make it past the add-on stage — not for lack of ambition, but because the rung they're standing on was never finished. Illustrative of the pattern we see, not a formal survey.

The fix isn't more ambition. It's knowing exactly which rung you're on, finishing it, and refusing to skip. Name your current stage. Define what "done" looks like for it. Climb deliberately. The systems that change industries are the ones that respected the ladder when no one was watching.

The one rule

Each stage is load-bearing for the next. Climb in order, finish each rung, and the top takes care of itself.

Now booking · Scottsdale, AZ

Not sure which rung you're on?

That's usually the most valuable conversation to have first. We help teams place their AI work on the ladder, finish the stage they're actually on, and design the climb to the next one — without skipping the parts that make it last.

Book a call
Climb.