| Fixed rules | AI with human review | Unsupervised AI |
|---|
| Example | Updating an indicator at the same time every day. | Drafting an article a person approves before it’s published. | Replying and publishing with nobody checking. |
|---|
| When it fits | The task never changes and the data comes from a reliable source. | There’s language, judgment or synthesis involved, and mistakes are costly. | Low-risk replies, tight guardrails and a safe fallback. |
|---|
| Risk | Low; breaks if the source changes. | Managed: the reviewer catches what the model makes up. | High: a mistake goes straight to your customer. |
|---|
| How we use it | Scheduled indicator updates and backups. | Editorial production, with topics approved before writing. | Only in chats with scoped answers and a WhatsApp handoff to a person. |
|---|
Fixed rules
- Example
- Updating an indicator at the same time every day.
- When it fits
- The task never changes and the data comes from a reliable source.
- Risk
- Low; breaks if the source changes.
- How we use it
- Scheduled indicator updates and backups.
AI with human review
- Example
- Drafting an article a person approves before it’s published.
- When it fits
- There’s language, judgment or synthesis involved, and mistakes are costly.
- Risk
- Managed: the reviewer catches what the model makes up.
- How we use it
- Editorial production, with topics approved before writing.
Unsupervised AI
- Example
- Replying and publishing with nobody checking.
- When it fits
- Low-risk replies, tight guardrails and a safe fallback.
- Risk
- High: a mistake goes straight to your customer.
- How we use it
- Only in chats with scoped answers and a WhatsApp handoff to a person.
General guidance as of October 2026. Where the line falls depends on your industry and what a mistake would cost.