Back to blog Perspectives

From Assistants to an AI Workforce: The Three Stages of Agent Adoption

Phinite teamNov 5, 20263 min read

Gartner's 2026 CIO survey found that 17% of organizations have deployed AI agents, against more than 60% expecting to within two years. McKinsey's State of AI put the share of organizations scaling agents in any single function at around 10%. The gap between intent and scale is where most of the enterprise sits today, and it helps to be precise about which stage you are in, because the obstacles are different at each one.

Stage one: humans and systems, with assistants

This is where nearly everyone starts. People operate software; an assistant drafts, summarizes and answers questions; a person reviews and acts. The unit of work is still a human decision. Value shows up as time saved per task and is hard to measure because the task was never timed before.

The ceiling at this stage is the review. If every output needs a person, the assistant can make each person faster but cannot change the shape of the operation.

Stage two: humans, agents and systems

The shift happens when agents take on the recurring work itself: a ticket triaged and routed, an invoice matched and filed, an account opened, a claim prepared for decision. People move to exceptions, judgment and direction. The workforce changes shape.

This is the stage where most programs stall, and not because the agents fail. They stall because operating an agent needs things a pilot never had: versions and environments, a runtime with scoped credentials, traces and alerts, evaluations before and after release, policies on what the agent may do, and a registry so the second team can reuse what the first one built. Gartner's forecast that more than 40% of agentic projects will be cancelled by 2027 names the causes: cost, unclear value and weak risk controls. All three are operating problems.

The other thing stage two needs is a number. One per process, agreed with the owner, built from data the agent produces on every run, with guardrails beside it. Without that, "unclear value" is guaranteed.

Stage three: a measured AI workforce

At the third stage the organization runs agents across functions the way it runs teams: with owners, budgets, approvals and a scorecard. Agents hand work to each other. The registry says what exists, who owns it and what it may do. Value is reported with honest labels, modeled, validated or realized, and Finance confirms it.

Industry by industry the scorecard looks different but the structure is the same. Banking and lending measure decision time and loss. Healthcare operations measure admin time and throughput. Retail and customer operations measure resolution and returns. Each function keeps its own number; the platform keeps the evidence.

What moves an organization between stages

The move from one to two is a choice of process, not a choice of model: one workflow with volume, rules, a cost line and a number the owner already tracks. The move from two to three is an operating decision: whether build, deployment, observability, evaluation and governance run as one connected lifecycle or as five tools stitched together. Databricks reported that organizations with active governance had twelve times more AI projects in production, and those with evaluation tooling six times more. That is the difference between a program that compounds and one that restarts with every agent.

Stage three is not a destination most companies will announce. It is the point at which nobody asks whether the agents are working, because the business numbers already say.

Written byPhinite teamWrites on building, running and governing AI agents in production.

Prove what your agents changed in the business.