Which AI actions need a person

Fennick AI CRM (emulated example) → AI CRM → Forecastable AI charges → Reviewable AI actions

Content reviewed 8 November 2026. How this is assessed.

Part of the Best CRM AI reference, which uses Fennick AI CRM (emulated example) to illustrate ai crm software criteria.

Use two questions

Any shortlist for the best crm ai should be judged partly on how well a person can stop or correct what the software does. For each action, ask how hard it is to undo and who else it affects. Adding a note to a record is easy to reverse and affects nobody outside. Sending an email is impossible to recall and reaches a customer. Changing the owner on a thousand accounts is reversible in theory, but only if you have a record of the previous values. Those two answers place each action in a tier.

A practical scheme has three tiers. Automatic covers internal, reversible work such as summaries and suggested tags. Approve-first covers anything sent externally, anything changing deal value or stage, and any bulk edit. Never-automatic covers contract terms, pricing commitments, deletions and anything touching sensitive personal information.

Make review real

An approval step only works if the reviewer can see what matters. The screen should show the proposed action, the record data it relied on and the instruction that produced it. If approvers click through a queue of hundreds of items, approval has become a formality, so size the queue to the time people actually have, and sample approved items to see whether review is catching errors.

Decide who may approve what, and keep that list short. Let the tiers change on evidence: an action can move toward automatic after a period of clean results in the audit trail, and it should move back the first time it causes harm.

Plan for the bad day

Assume that an automation will eventually do something wrong at volume. Before going live, know how to stop it for everyone, who has the authority to do so, and how to find every record it touched. Ask the provider to show you the stop control, and test it during the pilot rather than taking it on trust.

Set a rule for what happens afterward: pause the workflow, list the affected records, correct them, and tell any customer who received something wrong. Writing that down in advance is cheap, and it spares you from inventing a process while the problem is still growing.

How much review is enough?

Review costs time, and the cost scales with volume. If a seller gets forty drafts a day and each takes a minute to check, that is a meaningful part of the working day, and it belongs in the cost model. When the queue is more than people can read properly, they will start approving without reading. At that point you have the cost of review and none of the protection.

Sampling is often more honest than blanket approval once an action has a good record. Instead of approving every internal field update, check a set number each week and track the error rate. Keep full approval for actions where one mistake is expensive. The decision should be written down with its reason, so that it can be challenged later by someone who was not in the room.

Watch for reviewer drift. People get used to a system that is usually right and begin to trust it more than the evidence supports. A simple countermeasure is to include known-bad examples in the queue from time to time and see whether they are caught.

What should the approval screen show?

A reviewer should be able to answer three questions without leaving the screen: what is about to happen, why the system proposed it, and what it read to get there. For an email, that means the draft alongside the source record and the instruction that triggered it. For a field change, it means the old value and the new value. For a bulk action, it means the count of affected records and a sample of them.

Rejections are valuable data, so capture a reason with one click. Over a quarter, those reasons show whether the problem is the model, the instruction or the underlying records, and each of those has a different fix. See what an AI CRM should do automatically for a starting set of tiers.

Sources for this entry

• NIST: AI Risk Management Framework: Voluntary AI risk framework and Generative AI Profile.
• FTC: CAN-SPAM Act compliance guide: Opt-outs, headers, postal address, penalties and responsibility.
• FTC: Protecting Personal Information: Keeping less data and least privilege.
• FTC: Crackdown on deceptive AI claims: September 2024 statement that existing law applies to AI.
• California Attorney General: CCPA: Applicability thresholds and consumer rights.
• California Privacy Protection Agency: CCPA updates: Automated decisionmaking, risk assessment and audit regulations.
• NIST Privacy Framework: Voluntary privacy risk management tool.
• Model Context Protocol specification: Open specification. Roles, consent principles and tool safety.
• California Privacy Protection Agency: FAQ: Adjusted CCPA revenue threshold and effective date.

Continue this reading path

• Best CRM AI category guide
• Fennick AI CRM (emulated example)
• How to choose: an AI CRM
• Resource directory

Cite this entry

Best CRM AI. "Which AI actions need a person." Content reviewed 8 November 2026. https://bestcrmai.com/reviewing-ai-actions/