Why twelve — what the paper says, and what we added
OpenAI's 2026 paper "How Organizations Use AI" sums up enterprise AI adoption in one sentence: "Adoption is only the beginning of deployment." After buying AI, a company has to find where to use it → invest in complementary capabilities → redesign workflows → integrate it stably into daily work. (The paper is summarized in Part 0 of the series.)
That's what the paper says. To apply those four stages to a real company, questions arise that have to be answered. We distilled the ones we run into on the ground every time into twelve. These twelve are not the paper's claim but MAEUM's extension — the paper said "integration is needed," and we write "to integrate, you have to decide these."
The 12 questions — which design area each belongs to
| # | Question | Design area | What happens without an answer |
|---|---|---|---|
| 1 | Who will use it | Permissions | Anyone uses it, or no one does |
| 2 | For what | Task mapping | Stuck at "what should we do with AI" |
| 3 | Which data may go in | Data governance | Customer personal data flows into personal accounts |
| 4 | Which model | Model routing | Expensive model for simple classification, cheap model for contract review |
| 5 | How far does AI act on its own | Autonomy scope | The owner finds out later about a text the AI sent |
| 6 | Who reviews | Human review | Wrong answers go straight to customers |
| 7 | Can AI operate real systems | Execution rights | Orders and refunds go out with no human check |
| 8 | Who approves the result | Decision authority | Decisions with no owner pile up |
| 9 | Who is accountable when it fails | Accountability structure | After an incident it ends with "the AI did it" |
| 10 | How do we know it was good | Evaluation | Effects described only as feelings |
| 11 | What does it cost | Cost management | Surprise at the month-end bill |
| 12 | How do we improve next time | Organizational learning | The same mistake repeated department by department |
Each of the twelve is one part of the series. Permissions (1·7·8) → Part 2; autonomy and review (5·6) → Part 3; task mapping (2) → Part 4 (Korean); model (4) → Part 5 (Korean); data (3) → Part 6 and Part 9 (Korean); evaluation (10) → Part 8 (Korean); cost (11) → Part 10 (Korean); improvement (12) → Part 12 (Korean).
Self-check — count your yes / no / don't know
Next to each of the twelve questions, write one of yes · no · don't know. "No" is fine — it means you know nothing has been decided. The problem is "don't know": nothing was ever decided, and you don't know who should decide.
| "Don't know" count | Stage | Where you are |
|---|---|---|
| 0–3 | Integration in progress | AI is inside the workflow; only the gaps need filling. |
| 4–7 | Individual-use stage | Employees use it individually, but there's no company-level design. The paper's "adopted but not deployed" state. |
| 8–12 | Pre-adoption | Even with accounts bought, you're effectively pre-adoption. It's faster to decide everything together from the start. |
Most small and mid-sized companies are in the middle row. Nothing to be embarrassed about — the authors write that the paper's 1,764 large companies are also "still learning where it belongs."
An accounting-firm scenario — how the 12 questions actually come up
Say a six-person tax and accounting firm decides "let AI handle the reminders for clients who haven't submitted their bookkeeping documents." The twelve questions land like this.
- 1 Who — Only the two staff in charge. The principal gets result reports only.
- 2 For what — Drafting reminder texts and emails to clients who haven't submitted by the 5th and 15th of each month.
- 3 Which data — Client name, missing items, deadline only. No revenue figures or national ID numbers.
- 4 Which model — A low-cost model for text drafts; a higher-tier model for client-specific exceptions.
- 5 How far on its own — Up to drafting. No sending.
- 6 Who reviews — The staff member checks the draft and presses send.
- 7 System operation — Read-only on the client database. No write or send rights.
- 8 Approval — Changes to message templates require the principal's approval.
- 9 Accountability — For a wrongly sent message: the staff member → the principal. Not the AI.
- 10 Evaluation — Compare submission rate and number of reminder calls month over month.
- 11 Cost — Set a monthly cap on AI calls, alert when exceeded.
- 12 Improvement — Each month, look at "how much did people edit the AI drafts." When the edit rate drops, widen the autonomy scope one step.
Writing those twelve lines takes 90 minutes. That's what the first scoping session actually looks like.
The order for filling in all 12 within five scoping sessions
MAEUM's ₩250K scoping pass (5 sessions, up to 90 min each) is structured to fill in these twelve questions.
| Session | Questions filled | Output |
|---|---|---|
| 1 | 1 Who · 2 For what | List of repetitive tasks + one first seat chosen |
| 2 | 3 Data · 4 Model | List of data allowed in, model selection criteria |
| 3 | 5 Autonomy · 6 Review · 7 Execution · 8 Approval | Permission table (the 20-permission matrix from Part 2) |
| 4 | 9 Accountability · 10 Evaluation · 11 Cost | Named owners, 3 metrics, monthly cost cap |
| 5 | 12 Improvement + demo | A working demo (free) — the contract comes after |
When the five sessions end, the twelve answers remain as a one-page document. Even if you don't sign, that document is yours.
Sources — what the paper says / MAEUM's extension
- What the paper says: After adoption, enterprise AI requires a process of use-case discovery, investment in complementary capabilities, workflow redesign, and daily integration. Companies are learning "where to put it," not "whether to use it." (Chatterji et al., How Organizations Use AI: Evidence from ChatGPT, OpenAI 2026-08-11, original PDF)
- MAEUM's extension: The 12 questions and the self-check criteria needed to execute that process. Not in the paper; drawn from our field experience.
KRW is authoritative; USD figures are approximate. SEO and AI-search visibility work is included in every build at no charge. The demo is free — try it, and sign only if you like it.