AI agents in revenue operations are useful — and they're useful in a much narrower band than the demos suggest. After 18 months of running them in production for our own GTM stack and a half-dozen client engagements, three categories of work pay back consistently. One does not.
This piece is what we wish someone had handed us at the start: where to deploy them, how to measure them, and the one place we keep watching teams burn money.
Where AI is now meaningfully cheaper than people
Three categories. Not exhaustive — but every one we trust in production today fits into one of these.
Category one: structured-data shape-shifting
Taking dirty CRM data and forcing it into a known shape — companies, contacts, deal stages, pick-list values. The work is high-volume, low-judgment, and the success criterion is a regex you can write down. LLMs are dramatically cheaper than data analysts here, and the failure mode (a row mis-classified) is recoverable.
We use this for: account enrichment, ICP scoring, pipeline cleanup, lead-source consolidation. None of these need creativity; they need consistency at scale.
Category two: drafted-then-edited content
First drafts of standard-shape content: meeting recap emails, deal-stage updates, internal QBR slides. The model produces 80% of a competent draft; a human edits the last 20% and ships it. Time-to-publish drops 60–80%, quality stays comparable.
What kills this is teams forgetting the editing step. The output is competent, not good — ship it unedited and you train the audience to ignore the channel.
Category three: structured triage
Reading a freeform inbound — support ticket, sales-objection email, churn signal — and routing it. Plus a draft response. The model doesn't decide; it proposes. The human ratifies in under 30 seconds. Compared to triage by hand, this is 4–6× faster.
If the failure mode of your AI deployment is 'agent makes a final decision a human would have made differently,' you've put it in the wrong job. Move it earlier — to suggestion, not commitment.
Where it doesn't pay back
Net-new analytical reasoning over your data. The category we keep watching teams pour money into.
The pitch is seductive: ask the model a question, it queries your warehouse, it returns the answer. In practice the model gets the schema partly wrong, the joins partly wrong, and confidently produces a number that disagrees with last quarter's by 15% — for reasons no one can reverse-engineer.
A purpose-built BI tool with a defined semantic layer beats this every time, for less money. The exception is when you have a strict, governed schema, a small set of approved questions, and a human-in-the-loop on every answer — at which point you have a chatbot wrapped around your existing BI tool, which is fine but not magic.
How to measure whether it works
Three metrics. Not engagement. Not 'hours saved' (which nobody believes). Three concrete operational metrics:
- Time-to-output for the work the agent replaced — measure before and after.
- Edit distance: how much human work happens after the agent. If the human re-writes 60% of the output, the agent is theatre.
- Defect rate: how often does a human rate the output as wrong-and-needs-redo, not just imperfect-but-shippable?
If you can't measure all three, you can't tell whether the deployment is working. We've watched teams declare success by adoption rate while edit distance stayed at 70%.
What we'd build first
If you're starting from zero and have one quarter of headcount to spend: build the structured-data shape-shifter. Pick one specific dirty data problem (lead source consolidation, account enrichment, pipeline cleanup) and ship it end-to-end before touching anything more ambitious.
It's the deployment with the highest payback per engineering hour, and the one that lets the team learn how to operate AI in production without a flagship project at risk.