Vendors sell autonomy as a feature. In a system of record it is a cost decision, and the cost is paid later by whoever has to explain a number.
The five level ladder
Level 1: Read. The model sees data and produces nothing inside the CRM. Analysis, summaries for a human, answering questions. Zero risk to the record.
Level 2: Recommend. It produces a suggestion a person acts on. The suggestion lives outside the fields anyone reports on.
Level 3: Draft. It writes, but to a staging field, a note or a separate object that no report reads. Useful for building confidence with real data before anything counts.
Level 4: Write with approval. It proposes the actual change and a person confirms with one click. The approval must contain enough context to decide in seconds, or it becomes a rubber stamp, which is worse than level 5 because it looks like control.
Level 5: Autonomous write. It changes the record.
The three questions that decide the level
What does a wrong value cost? Noise, a wasted call, an angry customer, a regulatory problem. Each answer implies a different level.
Would anyone notice within a day? Silent failures are the expensive ones. If a bad write would sit undetected for a month, it does not belong at level 5 regardless of how accurate the model is.
Can it be reversed? If not, it needs a human in front of it.
A workable default
Level 5, autonomous: call summaries, activity logs, enrichment into empty fields, tags that no automation branches on. The shared property is that these are appends: a wrong value adds noise rather than replacing truth.
Level 4, approval: stage changes, owner changes, contact detail overwrites, anything that triggers another automation, anything a customer will see.
Level 1 to 3: anything where the process is still being defined, anything on data you do not yet trust, and anything irreversible.
Appends and overwrites are different permissions
This distinction does more work than any other on this page.
A model that fills an empty phone number is helping. A model that replaces a phone number a human typed is destroying evidence, and nobody will find out until someone calls the wrong person.
Give appends and overwrites separate permissions, and default overwrites to approval.
The audit trail that makes it defensible
Every AI write should record: what changed, the previous value, which process wrote it, when, and what it was based on.
The previous value is what makes rollback possible. The basis is what makes the decision explainable when someone asks in six months why a record says what it says.
How to move an action up a level
Run it one level below for a period, sample the outputs against what a human would have done, and count the disagreements. When the disagreement rate is low enough that the cost of the remaining errors is smaller than the cost of the approvals, promote it.
That is a measurable test rather than a feeling, and it is the same discipline that applies to any non deterministic system: you cannot reason about it theoretically, you measure it by running real data through it.
The policy in one page
Write down, for each object: who may write, at which level, with what logging, and who reviews the disagreements monthly. One page, agreed by whoever owns the CRM and whoever owns the risk.
Teams that have this page ship AI into their system of record without drama. Teams that do not end up either paralysed or surprised.