Why this distinction matters
Most teams start with AI by adding it to a product they already have: a chat panel, a "summarise" button, a writing assistant in a sidebar. This is AI-added design. An AI-native product starts from a different question: now that software can read, draft, classify and act, what should this workflow look like?
The gap shows up in results. MIT Project NANDA's July 2025 report The GenAI Divide found that about 95% of the enterprise generative AI pilots it studied produced no measurable profit-and-loss impact. The authors put this down mainly to brittle workflows, weak contextual learning and a poor fit with day-to-day operations, not to weak models. The model is rarely what holds a product back. The workflow around it usually is.
Step 1: Learn the six classification criteria
Use these six criteria to classify any product. For each one, ask which side the product falls on.
- Unit of value. AI-added gives the user an answer (text to read). AI-native delivers an outcome (a task moved forward or finished).
- Workflow shape. AI-added leaves the old steps as they were. AI-native removes, merges or reorders steps because the AI can now do them.
- Who starts the work. In AI-added products, nothing happens until the user prompts. In AI-native products, the system can start work when an event happens (a document arrives, a deadline nears).
- Where the AI lives. AI-added puts the AI in a separate panel the user has to visit. AI-native puts its output in the place where the work already happens: in the form, the queue or the document.
- Feedback loop. In AI-added products, user corrections are lost. In AI-native products, corrections are captured and used to make the system better.
- The removal test. Delete the AI. If the product still works almost the same way, it was AI-added. If the product no longer makes sense, it was AI-native.
Do: Copy these six criteria into a table. You will use it in Steps 2 and 4.
Step 2: Practise classifying real products
Apply the criteria to two well-documented examples.
- Cursor (code editor). Cursor is built by Anysphere as a fork of VS Code, but the AI is part of the editing itself. Tab completion predicts your next edit and where you will jump next, not only the next word. Background agents work on a separate Git branch while you keep working. Outcome-oriented, embedded, can work without constant prompting: mostly AI-native.
- A typical "Ask AI" sidebar in a document tool. The user opens a panel, types a question, reads the reply, then copies it back into the document by hand. Answer-oriented, separate panel, waits for a prompt, no learning: AI-added.
Most products sit somewhere between the two ends. Score each criterion on its own and don't force a single label too early.
Do: Pick two AI products you use at work. Score each against the six criteria (AI-added / mixed / AI-native) and write one sentence of evidence for each score.
Step 3: Understand why chat on an unchanged workflow rarely changes outcomes
Adding a chat window makes one step faster: finding or drafting information. It leaves every other step the same:
- The user still has to know what to ask.
- The user still copies the output into the system of record.
- The same hand-offs, approvals and waiting times remain.
- Nothing is recorded about whether the answer was right.
If the slow parts of a workflow are the waiting, re-keying and checking, a faster draft barely changes the end-to-end time or the quality of the result.
Klarna is a useful warning that volume is not the same as outcome quality. In February 2024 Klarna said its OpenAI-powered assistant was handling two-thirds of customer service chats in its first month. That was 2.3 million conversations, equal to the work of 700 full-time agents, with resolution time falling from 11 minutes to under 2. In 2025 its CEO said the company may have "over-indexed" on AI cost-cutting, and it began investing again in human support quality. The lesson for designers: decide which outcome you are improving and how you will measure it. Don't stop at how much work the AI absorbs.
Do: For one product you scored, list the workflow steps a chat feature leaves unchanged.
Step 4: Produce your one-page comparison
This is the artefact for this level. Pick one familiar product, for example a home-insurance claim form, and design two versions of it.
- Write the job in one line. Example: "Get a valid claim from report to decision."
- Describe the AI-added version. Example: the current claim form plus a chatbot that answers policy questions.
- Describe the AI-native version. Example: the customer uploads photos and a description. The system pulls out the details, fills in the claim, flags missing evidence and sends routine claims to a reviewer's queue with a suggested decision.
- Score both versions against the six criteria in a side-by-side table.
- Name one outcome metric that each version would change, such as time to decision or the share of claims needing rework.
- List one new risk the AI-native version introduces, such as a wrong suggested decision. Later levels show how to design for it.
Keep it to one page. If you can't explain the difference in one page, the AI-native version is probably not clear enough yet.
Checklist
- The job is stated as an outcome, not a feature
- Both versions are scored on all six criteria, with evidence
- At least one workflow step is removed or changed in the AI-native version
- The AI-native version passes the removal test
- An outcome metric is named for each version
- At least one new risk is recorded
Key takeaways
- AI-added design speeds up a step. AI-native design changes the workflow.
- Six criteria help you classify products: unit of value, workflow shape, who starts the work, where the AI lives, the feedback loop, and the removal test.
- A chat window on an unchanged workflow leaves the hand-offs, re-keying and checking in place, so outcomes rarely move.
- Measure the outcome you improve, not just the volume of work the AI absorbs.
- Most products are mixed. Score each criterion separately.