--- title: The AI Usage Spectrum source: https://steadman.ai/newsletters/david/ai-usage-spectrum.html published: 2026-04-12 summary: A 3x3 framework mapping three types of AI tool against three levels of usage, helping leaders understand where their people are and where they need to be. --- # The AI Usage Spectrum > The question for leadership isn't "are we using AI?" It's "which cell are most of our people in, and is that good enough?" Author: David Boyle, Steadman Published: April 2026 ## The framework Three types of AI tool. Three levels of usage within each. Most organisations have people scattered across all nine cells without knowing it. The goal isn't to get everyone to advanced. It's to give each person a deliberate nudge one or two cells to the right. ## Free / Constrained (Gen 1a) Examples: Free ChatGPT, free Claude, basic Copilot, bundled tools **Poor usage:** Treating it like a search engine. Generic answers pasted into slides. Unedited robot emails. > Treating it like a search engine. "What is a SWOT analysis?" Getting a generic answer, pasting it into a slide, calling it done. Or worse: asking it to write an email, sending the output unedited, and wondering why it sounds like a robot. **Good usage:** Using it as a thinking partner within its limits. Sharper prompts, heavier editing, realistic expectations. > Using it as a thinking partner within its limits. Drafting a first pass of meeting notes from memory. Asking it to challenge an argument before presenting it. Generating three alternative subject lines for an important email. Accepting the tool is weaker and compensating with sharper prompts and heavier editing. **Advanced usage:** Working around the ceiling. A personal prompt library. Knowing exactly when the tool will fail. > Recognising the ceiling and working around it. Using voice dictation to get thoughts down quickly, then running them through the tool for structure. Building a personal library of prompts that reliably produce useful output despite the model's constraints. Knowing exactly when the tool will fail — anything requiring multiple documents, long context, or precise numbers — and not asking it to do those things. ### The real problem with Gen 1a Free tools do teach foundational skills. That's helpful. But the limits are so severe that the opinion people form is "AI is mediocre." They learn to expect poor results. They stop pushing. The tool poisons the well: not by being useless, but by being just useful enough to seem representative. ## Pro / Competent (Gen 1b) Examples: Paid ChatGPT, Claude Pro, Gemini Advanced **Poor usage:** Owning a Ferrari and driving it in first gear. No custom instructions. Starting every chat from scratch. > Using a paid tool exactly like a free one. Never setting up custom instructions. Never uploading a document. Starting every chat from scratch with no context. Asking for a "summary" of a transcript instead of asking targeted questions. Accepting the first output without pushing back. Essentially owning a Ferrari and driving it in first gear. **Good usage:** Custom instructions configured. Projects with files loaded. Pushing back on first outputs. Check, Edit, Own. > Custom instructions configured with role and preferences. Projects set up for recurring work with relevant files loaded. Uploading the actual document and asking specific questions: "What are the three weakest arguments in this proposal?" Using the Check, Edit, Own cycle on every output. Starting fresh chats for fresh tasks. Knowing when to use thinking mode versus quick answers. **Advanced usage:** A personal operating system. Skills or Custom GPTs for recurring tasks. Sharing them with your team. Infrastructure, not a tool. > A personal operating system. Meeting transcripts processed daily into structured reflections. Style guides and templates loaded into projects so every output matches your standards. Multi-step workflows: first extract key takeaways, then pull supporting quotes, then synthesise across documents. Using voice input for all first drafts. Building Skills or Custom GPTs for the four or five tasks you do every week, and sharing them with your team. The tool is no longer something you "use." It's infrastructure you work inside. ### The real problem with Gen 1b **The skill gap.** It still takes considerable skill and knowledge to do long, complex tasks. You can generate a PowerPoint slide well, but not a hundred. You can do a piece of research well, but not a whole market research project in one go. The tools are capable; the techniques for sustained, complex work are not obvious. **The expectation gap.** People get excited by what these tools can do for individual tasks. They immediately jump to wanting longer and more complex work, and get frustrated at the limits. The gap between excitement and sustained value is where most people get stuck, and where most quietly give up. **The absorption gap.** Even when people sustain good usage, the organisation often fails to capture the value. People do better work, but without changes to roles, teams, and processes, the gains are absorbed rather than banked. This is an organisational design problem, not a tool problem — and it's why the [diagnostic](/newsletters/david/ai-maturity-diagnostic.html) matters alongside the spectrum. ## Agentic / Frontier (Gen 2) Examples: Claude Code, ChatGPT Codex, Google Gemini CLI **Poor usage:** Failing to set up your workspace. Letting the agent run unsupervised. "Build me an app" without clearly specifying the problem. Automating a broken process. > Failing to set up your workspace — no custom instructions, no rules files, no tool permissions configured. Letting the agent run unsupervised and trusting whatever comes back. Asking it to "build me an app" without clearly specifying what problem the app solves, who uses it, or what good looks like. Automating a broken process — making the same mistakes faster. Using agents for tasks where a simple prompt in a general-purpose app would be quicker and more reliable. **Good usage:** Setting up your custom instructions, rules, and tool permissions. Supervised stages. Explicit boundaries. Checkpoints where a human reviews before continuing. > Setting up your workspace properly: custom instructions that tell the agent who you are and how you work, rules files for recurring patterns, tool permissions so it can act without asking you to approve every step. Then breaking complex tasks into supervised stages. "First, catalogue these 40 transcripts. Show me the catalogue before you do anything else." Setting explicit boundaries: iteration limits, stop conditions, checkpoints where a human reviews before the agent continues. Understanding that agents are best for tasks with clear inputs, clear outputs, and repeatable logic. Using them for batch processing, file organisation, data transformation, and prototype building — not for judgment calls. **Advanced usage:** Entire workflows designed around agents. Humans design the system and check the output. Teams of people with teams of agents. > Designing entire workflows around what agents can do. A consulting team that records every client meeting, auto-transcribes, auto-extracts key takeaways by theme, and surfaces contradictions across interviews before a human analyst touches anything. An operations leader who has agents monitoring incoming data, flagging anomalies, drafting responses, and queuing them for human approval. The human role shifts from doing the work to designing the system and checking the output. ### The real problem with Gen 2 Almost nobody in a corporate environment is here yet, and the main reason is perception, not capability. People think these tools are for developers. They're just as powerful for senior executives and junior analysts, but they're not perceived to be for them. And because they're so powerful, they require care: clear boundaries to stop individuals making costly mistakes, and governance to manage security. But the first barrier is simply that most of the people who would benefit most don't know these tools are for them. ## What to do about it If you're on Gen 1a, push to Pro as soon as you can. The foundational skills transfer; the ceiling doesn't. If free tools are all you have for now, push to advanced usage as fast as you can — it's genuinely powerful, and it will change how your team works. Simultaneously, advocate for Gen 2 being deployed in your organisation. That conversation usually starts with a single senior sponsor who's seen what's possible and can make the case to IT and leadership together. It doesn't need to be a company-wide rollout. A controlled pilot with the right people is enough to demonstrate the value. We have a demo environment you can try. If you want to explore what agentic AI looks like in practice, ask us for a login and we'll set you up. We're happy to work with your IT and leadership teams to help them understand what this enables. The question for leadership isn't "are we using AI?" It's "which cell are most of our people in, and is that good enough to make us work better, quicker, and happier?" This spectrum maps how well individuals use each generation of tool. A separate but related question is how mature your organisation is at capturing the value — whether gains stay with individuals or compound across the business. The [organisational diagnostic](/newsletters/david/ai-maturity-diagnostic.html) measures that, and the [value map](/newsletters/david/ai-value-map.html) shows where the prize sits. ## Related - [Gen 1 vs Gen 2 AI](https://steadman.ai/newsletters/david/gen1-vs-gen2.html): What the generations mean - [Organisational AI Diagnostic](https://steadman.ai/newsletters/david/ai-maturity-diagnostic.html): Which phase has your organisation reached? - [AI Value Map](https://steadman.ai/newsletters/david/ai-value-map.html): Where's the value and how much are you capturing? - [Saturday AI Thoughts newsletter](https://steadman.ai/newsletters/david/): David's weekly emails on AI - [Steadman](https://steadman.ai): Advisory, coaching, strategic leadership ## Contact David Boyle: david@steadman.ai