AI agent decision classification
Which decisions can AI agents make · 150 decisions classified
AI agent decision classification sorts an organization's business decisions into three zones: green, where the agent acts alone; orange, where the agent prepares and a human validates; and red, where the decision stays human and the agent passes it on. ACF offers a classification of 150 common decisions across 8 functions as a starting point.
Why start from a classification
Most teams facing their first agent get stuck on the same blank page: everyone agrees that "critical decisions stay human" and nobody has written which ones they are. A ready-made list changes the conversation, since people react far more quickly to a concrete proposal they can dispute than to an abstract principle. Many frameworks choose not to prescribe a universal classification, and rightly so, since no list fits all companies. A worked classification is therefore a starting point to adjust, and the thresholds shown here are examples to set according to your own activity.
One decision per zone, for each of the 8 functions
| Function | Green: the agent acts alone | Orange: a human validates | Red: stays human |
|---|---|---|---|
| Customer service | Telling the customer where their case stands | Accepting a request for a refund | Answering a customer who threatens legal action |
| Sales and marketing | Qualifying an inbound prospect | Granting a discount outside the grid | Speaking for the company to the press |
| Purchasing | Ordering a listed supply at the negotiated price | Listing a new supplier | Signing a framework contract for more than a year |
| Finance | Reconciling bank statements with the accounts | Preparing a tax return | Changing a supplier's bank details |
| Human resources | Answering a question about a written rule | Shortlisting job applications | Rejecting an applicant |
| IT and development | Resetting an employee's password | Merging code and deploying it to production | Deleting data or backups in production |
| Legal | Answering from an approved legal note | Negotiating contract clauses | Signing a commitment for the company |
| Operations | Assigning tasks according to established rules | Deciding what happens to a non-compliant batch | Restarting equipment after a safety shutdown |
A decision can change zone
The zone comes with the conditions that move it. Paying a supplier invoice matched to its order and delivery note can be green, turns orange above a set amount such as 5,000 euros, and turns red if the supplier's bank details have changed since the last payment. Reconciling bank statements turns orange when a gap exceeds a set amount or stays unexplained after a few days. These conditions are what make the classification usable, since they are precisely what gets written into a mandate.
From the decision to the guardrail
Each decision can then be tied to its guardrails. ACF pairs the 150 decisions with 100 guardrails sorted across 8 types of agent: customer service, sales, purchasing and payments, finance, human resources, marketing and content, code and data, and internal assistant. Each one holds an instruction to place in the agent, a setting to place outside the agent, and a public incident that shows why it exists. The outside setting is the one that holds, and ACF Control enforces it before the action takes place.
The full classification and the guardrails are distributed free of charge on request at contact@acfstandard.com.
Frequently asked questions
What decisions can an AI agent make on its own?
Those that are frequent, follow written rules and are easily reversed: answering from a validated knowledge base, reconciling data, entering a matched invoice, resetting a password. Each comes with the condition that moves it to human validation.
How do I adapt the classification to my company?
Start from the list, remove what does not apply, add your own decisions, and rate each one by criticality. Change amounts and delays to fit your activity, and have the business owner of each decision validate the zone.
Is a decision in the green zone without risk?
No. Green means the agent may act alone, and the decision remains traced and supervised. Green-zone decisions are the ones most easily forgotten when conditions change, which is why their change-of-zone conditions matter.
Where does your organization stand?
ACF Score measures your agentic governance in about 10 minutes, the AI Act checker tells you which obligations apply, and ACF Control applies your rules before an agent acts.
Related terms
Written by Vincent Dorange, creator of the Agentic Commerce Framework®. Last updated .