What makes an agent workforce (and not a demo)
An AI agent is "workforce" when it clears three operational bars: it has a named human owner (not "the leadership team"), it has a measurable SLA (response time, expected quality) and it has cost-per-task metrics logged. Without all three, it's an experiment — with all three, it's a digital employee. The distinction isn't academic: it changes how you budget, how you measure and how you maintain it.
The 4 roles that work best in the enterprise
- Research assistant. Structured searches, executive summaries, competitive monitoring.
- Administrative process operator. Processing invoices, contracts, forms, validations.
- Junior commercial assistant (NOT senior). Initial qualification, booking meetings, operational follow-ups.
- L1 support agent. Resolving repetitive tickets with escalation to a human for L2/L3.
How to assign a human-in-the-loop
Human-in-the-loop is not "someone reviewing every piece of the agent's work" — it's someone stepping in at specific checkpoints. The right setup:
- Operational agent owner. 1 person, 5-20% of their time depending on agent volume.
- Automated checkpoints. High-impact decisions that require human confirmation before executing.
- Weekly sampled review. The owner reviews 30-50 random cases and flags hits/misses.
- Monthly iteration. 30-min meeting with the technical partner to tune prompts and configuration based on findings.
Governance, escalation and QA
Governance starts before the agent does: if nobody decided what identity it works under and how far its scope reaches, everything else is built on sand. That prior work is in what permissions to give an AI agent.
Without governance, the agent is technical debt dressed up as productivity. The 4 mandatory pieces:
- Clear usage policy — what the agent can do, what it can't, which decisions need a human.
- Auditable logs of every decision and every action with context.
- Rollback mechanism — when the output goes off, the system reverts to the prior state.
- Periodic review with automated evals + human sample review.
Real cost per task
The correct calculation includes all five components — not just "LLM tokens":
- LLM cost (tokens) per execution.
- Infrastructure cost (compute, storage, vector store) prorated.
- Tooling cost (CRM API, etc.).
- Human supervision cost (% of the operational owner's time).
- Amortized setup cost (split over expected useful months).
When it doesn't pay off
- Volume <50 agent runs/week — setup doesn't pay back.
- Process about to be killed or transformed.
- No operational owner available or assigned.
- Cases where the cost of error far exceeds expected savings.
- Teams without bandwidth for the minimum monthly iteration.