Not Every Job Needs a Genius: The 4 Levels of AI Intelligence

Most companies pick one AI model and point it at everything, which is the equivalent of staffing an entire business with either interns or Nobel laureates, and it explains why so many AI budgets produce so little.

The question almost nobody asks before deploying AI is the one that determines whether the project pays off: How much intelligence does this task actually need?

Having spent nearly two decades building lead-generation and marketing technology, and the last several years investing in and building companies in the application layer of AI, I’ve seen that across all of it, the same pattern holds: Organizations run on thousands of distinct tasks, and those tasks sit at very different levels of required intelligence. Matching the model to the task, and the scaffolding to the stakes, is the whole game.

Start with value, not technology

Before any model decision, sort the work with one test: Would a customer pay for this?

Most companies’ weeks are split into three piles.

Tasks that add no value, such as retyping data between systems, answering “where’s my order?” for the hundredth time and categorizing receipts.

Tasks that add value but repeat: first drafts of proposals, personalised outreach, the monthly report clients actually read.

And tasks where the value is the human judgment: pricing, market selection, the call that saves a churning customer, the key hire.

If a task vanishes tomorrow and no customer would notice, it doesn’t add value. That first pile is where AI pays for itself in the first month. The third pile is where founders should be spending their own hours. The mistake I see constantly is the reverse: leaders automating the judgment call and personally retyping invoices.

The data supports this task-level view.

The U.S. Department of Labor’s O*NET database catalogues roughly 19,000 distinct work tasks across about 900 occupations, and the emerging research on real-world AI usage shows businesses automating specific tasks such as email handling, document processing and scheduling rather than whole jobs.

Meanwhile METR, a non-profit that measures how long a piece of work AI can complete unsupervised, finds that horizon has been doubling every four to seven months, from a few minutes in 2023 to roughly a working day today.

The four levels

New models ship every few months, so memorizing names is pointless and what stays stable are the categories. Every major lab sells essentially the same ladder, from a cheap-and-fast tier at the bottom to a flagship at the top. If you read any price list top to bottom, you’re looking at these four levels.

One definition first: scaffolding. Scaffolding is everything you build around the model: connecting it to your data and tools, validating its outputs, setting permissions on what it may do, and putting a human in the loop at the right moments. Models are rented, and scaffolding is owned.

Level 1: Reflex work. Seconds-to-minutes tasks that happen thousands of times: routing tickets, pulling totals from invoices, tagging leads, answering routine policy questions. High volume, low stakes. Use the smallest, cheapest tier any provider offers, or a small open-source model on ordinary hardware. Scaffolding is minimal: a clear instruction template, a format check, and a rule that anything uncertain goes to a person.

Level 2: Skilled drafts. Minutes-to-an-hour tasks a competent professional handles alone: marketing copy, sales outreach, contract summaries, substantive customer replies. Moderate stakes, because the output represents your company. Use the mid-tier workhorse model, the default in any major chat product.

Scaffolding now matters: the model needs your brand guide, product catalog, pricing and past work or it writes generic prose, and a human should review anything that ships externally. Going open source here means building what the paid platforms hand you as ready-made features. Budget for that, or the “free” model gets expensive.

Level 3: Agentic workflows. Multi-hour tasks that cross systems and take real actions: reconciling invoices against accounting, researching a prospect and updating the CRM and drafting outreach, screening and scheduling candidates.

This is where business AI usage is concentrating, and where most projects fail. Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027, because the stakes jump from “bad draft” to “wrong action.”

Use the flagship tier. Open-source options trail here by roughly a tier, so choosing them means compensating with engineering: orchestration, strict limits on what the agent may touch, approval gates before consequential actions, and a full audit log.

My rule of thumb is that open source at Level 3 roughly doubles your engineering investment in exchange for data control and lower running costs. For regulated industries that trade can be worth it. For most others, not yet.

Level 4: Frontier work. This tier barely existed a year ago: day-to-weeks projects that previously required a senior expert, a whole team or a consultant’s invoice. A new class of long-horizon models that began shipping this year handles them whole, and every major lab is racing to field one.

For a founder, the unlock isn’t the software demo. It’s what this does to the third pile, the high-value work you previously couldn’t afford: the competitor analysis you’d have paid five figures for, a full financial model for an expansion, a due-diligence pack on an acquisition target, research across hundreds of contracts and PDFs.

Work that used to be broken into a dozen supervised chunks can be delegated whole, with the founder supplying the judgment at the start and the decision at the end. There is no open-source equivalent yet. The alternative is old-fashioned project management, a person slicing the work into Level 3 pieces, which is precisely the overhead this category eliminates.

The scaffolding inversion

Here is the part of the framework I consider most important, and the one companies most often get backwards.

At the bottom of the ladder, scaffolding compensates for the model’s weakness: checks, narrow instructions, human review. At the top, scaffolding governs the model’s strength: permissions, checkpoints, audit trails. Yet I routinely see teams wrap their cheap email-sorter in bureaucracy and hand their most powerful agent the keys.

What this means for founders and investors

For operators: Make the three value piles, then inventory 20 recurring tasks across your departments, not just the technical ones, and place each on the ladder by duration and consequence. Automate the no-value pile this month with the cheapest tier, prove the savings, and reinvest them upward.

For investors, and for anyone building in the application layer: The durable value is not in the model. Models are commodities on a ladder every lab sells. The defensible asset is scaffolding, the proprietary connection between a model tier and a specific, high-stakes workflow, and the judgment about where a human belongs in the loop. That is what I look for, and it is what I’d build.

The companies winning with AI aren’t the ones with the smartest model. They’re the ones who stopped asking “which AI should we buy?” and started asking two better questions: “Does this task deserve anyone’s time, human or machine?” and “Which level of intelligence does it need?” Answer those for every task in the building, and the roadmap writes itself.