Revenue decision guide · AI authority
AI CRM write controls: keep people accountable
Let AI capture, structure and propose while protected CRM fields remain behind explicit authority, evidence and approval.
The operating question
Automation becomes risk when it silently changes the commercial record.
An AI agent runs overnight and updates the close date on forty opportunities. By the time the team opens the forecast, the quarter has moved four per cent. Nobody approved the individual changes and nobody can reconstruct which sources the agent used.
Some updates may be right. A buyer email did mention a later date. Others may have treated a meeting proposal as an agreed commitment. The immediate debate becomes whether the model is accurate enough, but the commercial record has already changed before that question can be answered.
The first control isn't model accuracy. It's permission. What may the system read, what may it propose, which fields may it write, whose approval is required and what evidence must travel with the change?
Turning it off and trusting it are both incomplete answers
Turning automation off protects the record and gives up useful work. Trusting every update saves time and gives a model authority it may never have been granted.
Most organisations move between those positions because the control is applied to the whole tool. The better boundary is set by action and field. Capturing, retrieving, structuring, summarising and proposing are assistance. Changing a forecast category, close date, stage or amount can alter what leadership commits to and should follow explicit authority.
The argument is about permission, not accuracy
A model can be right most of the time and still need approval for a consequential write. Accuracy describes performance across cases. Authority determines whether this system may change this field in this situation.
Keep the two questions separate. First, was the action allowed? Second, was the proposed value supported? A usually accurate model isn't a control, just as a usually accurate person isn't automatically authorised to approve every commercial decision.
Require the evidence with the proposal
A proposed write should arrive with the target field, current value, proposed value, source, relevant extract, applied date and uncertainty. The reviewer needs enough context to accept, correct or reject it without hunting through another system.
Record the model contribution, the human decision, the configured authority and the final write outcome. If the proposal is corrected, keep both values. If it's rejected, keep the reason so the system and process can be reviewed later.
Make authority specific
Customers should decide which roles can approve which fields, at what scope and under which conditions. A Manager may control a forecast category for their team while RevOps owns stage definitions and Admin controls delegation and account safety.
An external AI tool should not gain write authority simply because it has a CRM connection. PROOF's control layer can require evidence and approval before protected fields change, while allowing low-consequence administrative updates to continue automatically where the customer permits them.
What the external evidence says
Forrester's 2026 B2B Marketing, Sales and Product Predictions says untested generative AI functionality and lagging user skills will lead to incidents costing more than $10 billion in enterprise value. This is an analyst prediction published on 28 October 2025, not a measured outcome. Read Forrester's prediction.
Boston Consulting Group's August 2024 Executive Perspectives describes sales AI transformation as 90% sales change management and 10% technology, data and AI. BCG presents this as an executive view from client work and publishes no research base in the document. It's useful context, not a statistic PROOF can generalise as measured fact. Read BCG's perspective.
Where write controls go wrong
Don't create an approval queue for every field. People will approve in bulk and the control will become theatre. Apply review to changes that alter commercial meaning, and tune it to the customer's authority model.
Don't let a technical administrator approve commercial truth merely because they configured the connection. Configuration authority and decision authority are different. The record should always show who held each one.
How PROOF approaches it
Keep the evidence, decision and action connected.
PROOF keeps the source, timing and accountable decision together so the team can act while the outcome can still change and learn from what happened afterwards.
See the working product