What to keep human, forever

The goal is not a CRM that runs itself with no humans, it is a CRM where humans spend their scarce judgement on the handful of calls that actually need it and nothing else. The productivity lift from AI agents comes mostly from eating the repetitive admin, the data entry, the dedup, the chasing of stale fields, the work that consumes that 17% of the week. That is the labour. The judgement stays with people, and being deliberate about that line is what keeps the system trustworthy rather than merely fast. The cautionary note in the research is worth holding: deploying agents widely does not automatically lift productivity, drawing the autonomy line carelessly is how teams get the tools without the gains.

Keep these out of the agent's autonomous reach, permanently, regardless of how good it gets:

• Killing or reviving a real deal , the agent flags the phantom; a human pulls the trigger, because they may know something the activity stream doesn't (a verbal yes, a board delay).
• Moving a deal's forecast category , commit, best-case, pipeline. This is a promise to the business and a human owns it.
• Reassigning an owner , a territory and relationship decision, never a data one.
• The final committed number , the agent assembles the most honest possible forecast; a human signs it.

A hygiene-and-forecasting system lives or dies on trust, and trust is lost the first time the agent does something irreversible and wrong. Draw the autonomy line conservatively and widen it with evidence, every rule earns its autonomy by being right in suggest mode first. That is the whole discipline: relentless automation of the truth-keeping, careful human ownership of the truth-telling. Build it that way and the forecast stops being a quarterly act of optimism and becomes a number the business can actually run on. If you are doing this without a sales team behind you, running a pipeline solo with AI shows how the same agents carry a founder who is also the whole revenue org.