The functions where agents land first
Customer operations: triage, account actions, and resolution drafting, with [conversational and voice agents](/learn/ai-voice-agents) at the front line. Sales and revenue operations: enrichment, follow-up drafting, CRM hygiene — work that dies of neglect when humans own it. Finance and back office: reconciliation across systems that almost agree, exception chasing, report assembly. IT service: provisioning, access requests, first-line incident response. The common shape is unmistakable: a recognisable workflow, many small decisions, an end state someone can verify. Where work is novel, ambiguous at the ends, or irreversible, agents assist rather than own.
How to pick the first deployments
Score candidate workflows on three axes: volume (does it happen enough to matter), verifiability (can anyone state what done-correctly looks like), and reversibility (what does a wrong action cost to undo). High-volume, verifiable, reversible work is the green zone — and it is where the [build discipline](/guides/build-an-agentic-ai-system) produces evidence cheaply, which is what later, riskier deployments will be approved on. The classic mistake is starting with the impressive workflow instead of the verifiable one, discovering the agent is wrong sometimes, and having no evaluation record to say how often.
The bite: costs that arrive with the value
Three costs are reliably underestimated. Oversight is not free — every agent needs the [identity, gates, and audit work](/guides/govern-agentic-ai), and ten agents built ten ways cost more to govern than the work they replaced. Integration is the real build — the model is hours, the connection to your systems of record is the project. And workflow drift is permanent: the process the agent automated keeps changing, so an agent without an owner and a review cadence decays into confidently doing last year's process. Budget for the agent as you would for an operational system, because that is what you are buying.