Lesson 4 of 10 · 8 min read
The Graduated Exposure Ladder
In one paragraph
The Graduated Exposure Ladder is IOanyT's framework of five levels for deciding how much language-model behaviour is allowed at run time: Level 0 build-time only, Level 1 internal assist, Level 2 bounded extraction, Level 3 constrained conversation, and Level 4 promotion, where proven patterns are compiled back into code.
In this lesson
- Name the five levels of the ladder and what each allows
- Place an existing AI feature on the ladder
- Explain why steps move up only after proving themselves, and why they move back down
The video version of this lesson is in production. The full lesson is below.
Compiling everything would be wrong. Some steps involve messy inputs, like a scanned document or a customer’s free-text question, where a language model genuinely helps at run time. The question isn’t “AI or no AI?” but how much model, at which step, under what controls? The Graduated Exposure Ladder is a way to answer that consistently.
The five levels
Level 0: Build-time only. The model writes code, rules and tests before release; nothing calls a model at run time. Example: eligibility, pricing or routing rules compiled into code. This is the default starting point for any decision that matters.
Level 1: Internal assist. A model produces output for staff, who review it before anything happens. Example: drafting a customer reply that an agent edits and sends, or summarising a case file for an underwriter.
Level 2: Bounded extraction. A model fills a strict, predefined form, code checks the result, and a person takes over when it doesn’t fit. Example: reading invoices or claim forms into fixed fields, where code checks that totals add up and dates are valid.
Level 3: Constrained conversation. A model handles language with customers, understanding questions and phrasing answers, while deterministic flows control every action, price, policy statement or commitment. Example: an assistant that understands “can I change my booking?” but whose answer about fees comes from code, not from the model’s memory.
Level 4: Promotion. Patterns that have proven stable at Levels 1–3 are compiled back into deterministic code, so the amount of runtime AI shrinks over time. Example: the extraction rules for a supplier’s invoice layout that never changes become ordinary parsing code.
The rule that makes it work
Every step starts at Level 0 and moves up only after it has proven itself on your data. “Proven” means measured against criteria agreed before the test (accuracy, consistency, and what it does when unsure), not a successful demo.
Two consequences follow:
- Moving up is a decision, not a drift. Someone signs off, with evidence.
- Moving down is normal. If a model update or new data makes a Level 2 step less reliable, it drops back until it’s re-proven.
The line that never moves
At every level, commitments stay with code or people: prices, approvals, refunds, legal or policy statements, anything involving money. Level 3 conversation can be fluent and helpful, but when it reaches a commitment it hands off to logic your team approved. That single line is what separates a controlled system from the failures in lesson 1.
Try it yourself: place your AI on the ladder
List the AI features your organisation runs today. For each one, write down:
- Which level it actually operates at (be honest: a chatbot quoting prices from its training is above Level 3, which means it’s off the ladder).
- Whether anyone decided it should be at that level, with evidence.
- What one step down would look like.
Features that are off the ladder, or on it by accident, are where to look first.
What comes next
“Proven on your data” needs a method. Lesson 5 covers the evaluation practice behind it: golden datasets, regression evaluation and error analysis.
Key takeaways
- Deterministic AI isn't all-or-nothing. Each step of a workflow sits on its own level.
- Every step starts at Level 0 and moves up only after it has proven itself on your data.
- At every level, code controls commitments: prices, decisions, money, policy.
- Level 4 runs the ladder in reverse: patterns that prove stable get compiled back into code.
Check yourself
Pick an answer, then open the card to compare.
-
1. A model drafts replies to customer emails, and an agent edits and sends each one. Which level is that?
- A.Level 0, build-time only
- B.Level 1, internal assist
- C.Level 3, constrained conversation
Show the answer
B. Level 1, internal assist The model's output is reviewed by staff before any action. That is internal assist.
-
2. A model reads invoices and fills a fixed form; code checks the totals; unclear invoices go to a person. Which level?
- A.Level 2, bounded extraction
- B.Level 3, constrained conversation
- C.Level 4, promotion
Show the answer
A. Level 2, bounded extraction Strict schema, validated by code, with a human fallback. That is bounded extraction.
-
3. What happens at Level 4?
- A.The model is given full autonomy
- B.Proven, stable patterns are compiled back into deterministic code
- C.All AI is removed from the product
Show the answer
B. Proven, stable patterns are compiled back into deterministic code Promotion reduces runtime AI over time by turning what the model reliably does into tested code.
Common questions
Can one product have steps on different levels?
Yes, and most do. A support assistant might understand questions at Level 3, extract order details at Level 2, and decide refunds at Level 0.
How do you decide when a step can move up?
By measuring it on your own data against criteria agreed before the test, covering accuracy, consistency and how it behaves when unsure. The criteria depend on what an error costs in that workflow. Lesson 5 covers the measurement practice.
Is there a level where the AI acts entirely on its own?
Not as a destination. Autonomous action fits only reversible, low-value tasks with hard limits on what the system may do. Commitments that matter stay with code or people.