Founders keep asking me the same question: our company knows it should be using AI seriously, but where should we actually be? So here is the map I draw on the whiteboard.
Three levels, each with a real tradeoff, and a clear answer to the question every leader is actually asking: where should we be?
One note before we start. This ladder is about where and how you run models, not how good you are at using them. A team at Level One can outperform a team at Level Three. In fact, most do.
Level One: Everything Inside the App
At Level One, your team works inside a single AI platform's own applications. In our case that is Claude: the chat app, the desktop app, Claude Code for technical work. No custom infrastructure, no API keys, nothing to maintain.
People dismiss this level as "just using the app." That is a mistake, because the real work at Level One is learning the two dials that actually matter:
- Model types. Every provider ships a family of models at different capability and price points. Claude currently spans from Haiku, fast and cheap, through Sonnet and Opus, up to Fable 5 at the top. Knowing which tier fits which job is a skill, and it transfers everywhere.
- Effort levels. The newer models let you set how hard they think: low, medium, high. Planning a strategy deserves high effort on a top model. Summarizing a meeting does not. Teams that never touch this dial either overpay on every task or underpower the ones that count.
What you give up at Level One is choice. You are inside one vendor's ecosystem, on their tools, at their pace.
The honest truth: most companies should master this level before touching the next one, and many never need to leave it. The constraint at Level One is almost never the platform. It is how well your people use it.
Level Two: An IDE and Any Model You Want
At Level Two, your team works through development environments and API access. An IDE, an integrated development environment, is the workbench developers write software in, and modern ones like Cursor or VS Code let you plug in any model from any provider, Anthropic, OpenAI, Google, all of them, switchable per task.
This is where optimization lives. You can route each job to the best model for it: one model for planning, a cheaper one for execution, a specialist for code. You can automate workflows, chain models together, and tune cost per task in ways Level One does not allow. For a technical team, the gains are real.
The tradeoff is just as real: data privacy and control get harder. At Level One, one vendor sees your data under one agreement. At Level Two, your data flows to multiple providers under multiple agreements, through API keys that need managing, with logging and retention policies that differ by vendor. Every additional provider is another place your customer data, financials, or product plans might travel. Most companies discover this gap in a security review, after the workflows are already built.
Level Two makes sense when you have technical staff, real volume, and someone explicitly responsible for governance: who can use which model, with what data, under what agreement. If nobody owns that question, you are not ready for this level. You are just at Level One with extra risk.
Level Three: Open-Weight Models You Host Yourself
At Level Three, you run the models on infrastructure you control. Open-weight models are models whose trained parameters, the "weights," are published so anyone can download and run them. Meta's Llama family and Mistral's models are the best-known examples.
Why would anyone take this on? Control. Your data never leaves your infrastructure, which matters enormously in healthcare, finance, defense, and any business handling data that regulators or customers say cannot travel. You are immune to a vendor deprecating the model your workflows depend on. And at very high volume, running your own can beat paying per token.
The cost is that you just became an AI infrastructure company. GPU hardware or cloud GPU commitments, engineers who can deploy and monitor models, security hardening, and a hard truth on quality: open-weight models trail the frontier. The best models in the world are not open-weight, so you are accepting a capability gap in exchange for control.
Level Three is the right call for a narrow set of companies with regulatory mandates, extreme scale, or genuine sovereignty requirements. For everyone else it is an expensive way to feel in control while your competitors ship faster on better models.
Where Should You Be?
Three questions settle it:
- Can your data legally and contractually leave your infrastructure? If no, Level Three is not a preference, it is a requirement. If yes, keep reading.
- Do you have technical staff and an owner for AI governance? If no, stay at Level One and master it. If yes, Level Two is open to you.
- Is your team actually good at Level One? Model selection, effort levels, clear problem framing. If not, moving up the ladder multiplies confusion, not capability.
Notice the pattern. The ladder is not a maturity contest. Climbing it trades simplicity for control, and you should climb only when a specific constraint forces you to.
One more thing the ladder cannot fix. The skills that matter most, knowing what you want, framing problems clearly, matching models to jobs, live inside your people at every level. Models keep changing, and the way you work with them has to change too. The old habits, including the prompts your team wrote for last year's models, quietly expire as new generations ship. I wrote a separate piece on that, Your Prompts Are Expiring, because it catches almost everyone off guard.
The Leadership Angle
Each level demands something different from your people, and this is where most companies get the sequence wrong.
Level One needs trained operators: people who know the models and use them well. That is a training problem more than a hiring problem, and it is the cheapest, fastest lever you have. Level Two needs engineers plus a governance owner. Level Three needs infrastructure and ML engineers, among the hardest hires in the market right now.
Most companies we meet are trying to staff for the level above the one they have mastered. The smarter move is almost always to get great at the level you are on, and climb only when a real constraint, not ambition, forces it.
That is the judgment we build and verify. AI Officer certification tells you whether someone can lead an AI program at any level: mapping the workflows, choosing what to delegate, and owning the outcome. If you are weighing where your company should sit and who could take you there, that is the place to start.