/ Insights
By Daniel
18 May 2025
Dan Martell, the SaaS founder and investor behind Buy Back Your Time, has a five-level framework for how seriously a company has actually adopted AI. It runs from poking at ChatGPT all the way to training custom models on proprietary data.
I like it because it is a diagnostic, not a pep talk. Read the five levels and you know within a minute where your company honestly sits. But Martell talks about it as an operator, and most people who ask me about it are the ones who have to build what he is describing. So here is each level, plus my take on what the jump to the next one actually costs.
This is where almost everyone starts and where most people stay. "Most people, they're just in the chat realm... they're just playing around," Martell says. Someone asks ChatGPT to draft an email or summarize a document when a task feels overwhelming. No process, no integration, and usually weak prompting, so the results are mediocre and the takeaway becomes "AI is overhyped."
Nothing wrong with level 1 as a starting point. Staying there is the mistake, because the value at this level is near zero and the mediocre output convinces teams the ceiling is low.
Level 2 is intentional, recurring use: AI applied weekly to document, streamline, and improve internal processes. Martell's mnemonic is that "Systems stands for Save Yourself Time, Energy, and Money." Think AI-generated checklists for recurring problems, SOPs, troubleshooting guides.
His warning to leaders is the part worth repeating: you almost certainly overestimate how much your team uses AI. People need it modeled for them and they need to be reminded. My take: the jump from 1 to 2 is not a technology problem at all. It is a management problem, and no vendor can do it for you.
At level 3, AI stops being optional. Martell is blunt about it: in his telling, if you are on a team and not using AI in your work, you are the problem, and he mandated AI use across his companies. His examples of what followed are the interesting part. His finance team, initially resistant, went and found SIFT, an anomaly-detection tool, once usage was required. His media team worked with an engineer to build a custom internal tool that automates B-roll insertion for YouTube videos, cutting roughly 80% of the manual effort.
Two things worth underlining from the engineering side. First, mandates drive discovery: teams do not evaluate tools seriously until not using them stops being an option. Second, Martell's sharpest point is that internal tools can become products. The B-roll tool his media team built for themselves is something they could sell. We see this pattern constantly: the internal tool a team builds to fix its own bottleneck is often a better validated MVP than anything that started on a whiteboard. If that is you, that is a real path to market.
"This is, I hope, where you guys live. Because this is where a CTO should live," Martell says. Level 4 means going past off-the-shelf tools: understanding the model ecosystem beyond the two or three big names, and wiring AI into your product and operations through APIs.
His concrete recommendation is to hire what he calls an automation person: a developer focused on AI APIs who runs a Theory of Constraints analysis (the Goldratt, The Goal kind) on each department, finds the worst bottleneck, and attacks that with AI. He claims the role pays for itself within 60 days. Believe the exact number or not, the principle is right: AI aimed at your single biggest constraint beats AI sprinkled everywhere.
Here is what the framework does not tell you about level 4, because Martell is an operator and this part is an engineering problem. An AI feature that works in a demo and one that survives production are different animals, and the gap is evals, fallbacks, and cost control. It is the exact gap where AI-built prototypes stall. Two specifics from our own work: once model calls sit in your hot path, your bill scales with your usage, and once models can read from or act on your systems, you have inherited a security surface most teams never audit.
The top level is building things that did not exist before: custom models, real data science, and proprietary data treated as a strategic asset. Martell notes that some companies are "turning off APIs because they realize that the advantage they have is the data." The story he tells is about his friend Daryl at FlexPay, who picked one narrow, high-value problem (credit card declines), hired a data scientist off Upwork, and built an immensely valuable company in about three years on the strength of proprietary transaction data.
This is the level people assume is out of reach, and it is less exotic than it sounds. We made this jump in our own product: we trained small task-specific models on data we had already paid frontier models to produce, and cut per-session model cost from $6.45 to $1.60 while holding quality behind strict recall gates. The full numbers are in Cheap Model Finds, Expensive Model Thinks. The pattern generalizes: if you have been running LLMs in production for a while, you already own a training corpus. You are just not using it yet. That is the exact engagement our AI model training work is built around.
Honest scoping. The jump from level 1 to 2, and most of 2 to 3, is management: modeling AI use, mandating it, checking that it happened. You do not need an engineering firm for that, and I would be suspicious of one that says you do.
Levels 3 through 5 are where we work. Turning an internal tool into a product (MVP from idea to market). Hardening an AI-built prototype so it survives real users (AI prototype to production). Auditing what your AI integration exposes (AI app security review). Training custom models on your own data when the API bill or the moat argument says it is time (AI model training).
The useful first step is the honest one: figure out which level you are actually at, not the one on your board deck.
Not sure which level you are really on, or what the next jump costs? Book a readiness audit and we will map it together.
Frontier LLMs are spectacular, and spectacularly expensive at scale. We cut per-session AI processing cost from $6.45 to $1.60 by training small task-specific models on our own data, and the deeper win was teaching cheap models to find exactly what the expensive models need.
Long context vs RAG is a false choice. Here's the three-tier memory architecture I expect to replace both within three years, and what to build (and skip) now.
A Lovable app can reach production, but clickable is not ready. What we check first: Supabase RLS policies, auth, secrets, the release path, and who owns the code.
