AI Workforce Sizing: How Many Agents Do You Actually Need

Three years into the agentic AI era, enterprises have moved past "should we use AI agents" into "how many." The question sounds simple. It is not. Too few agents and the AI program is a footnote on the operating budget that delivers limited value. Too many agents and the operations cost, the maintenance burden, and the user confusion outweigh the benefits.

The right number depends on your workflows, your maturity, and your tolerance for governance overhead. This is the framework for deciding.

What an agent is, for this purpose

For sizing purposes, an agent is a discrete capability with:

• A defined purpose and scope
• A named business owner
• A documented governance posture (risk class, oversight, audit trail)
• An evaluation harness measuring its outcomes

A "multi-LLM chat interface" is not an agent for sizing purposes; it is a platform feature. A "customer support triage agent" with the four properties above is an agent.

The distinction matters because the maintenance cost of an agent scales with the four properties — each agent has documentation, owner, governance, and evaluation that must be maintained.

The sizing question in three dimensions

Breadth: how many distinct functional domains do agents serve? Sales, marketing, support, finance, ops, HR, engineering, compliance? More breadth means more change-management surface and more diverse governance challenges.

Depth: per domain, how many agents? One catch-all agent per domain or many narrow agents? Depth affects per-domain capability but also per-domain operational cost.

Sophistication: are the agents simple (single-LLM call with clear input/output), composite (multi-step with retrieval and tools), or pipelines (multi-agent orchestration with human approval)? Sophistication affects per-agent cost and per-agent value.

The three dimensions interact. An enterprise can be wide and shallow (one agent per domain across all domains), narrow and deep (many agents in one strategic domain), or broad and deep (the mature steady state, which takes years to reach).

The sizing pattern by enterprise maturity

Year 1: 1-3 agents. One strategic agent that proves out the value, plus 1-2 supporting agents that establish the operational pattern. Focus is on getting the platform, governance, and operations right rather than maximising scope. The breadth is intentionally narrow.

Year 2: 5-15 agents. Expand into 2-4 functional domains with 2-5 agents per domain. The operational pattern is now repeatable. Compliance evidence is being produced consistently. The platform team has learned the failure modes.

Year 3: 15-40 agents. Coverage across most functional domains, with depth in the strategic ones. Multi-agent pipelines are becoming common for cross-functional workflows. The platform team has stabilised at the necessary headcount.

Year 4+: 30-100+ agents. Mature steady state. New agents ship in weeks, not quarters. Agents retire when they no longer deliver value. The AI capability is operational infrastructure, not a project.

The progression is not strict. Some enterprises move faster on the first two years; few skip the years entirely without paying for it later.

Where the breakpoints fall

The operational breakpoints that change the sizing equation:

5 agents: the breakpoint where ad hoc agent management stops working. Need documented governance, evaluation harness, and a named platform team.

15 agents: the breakpoint where per-agent prompts and tools need versioning, deployment pipelines, and rollback procedures. Hobby-grade operations stop working.

30 agents: the breakpoint where the agent estate needs its own change-management process. Coordination overhead becomes visible. The AI program manager role becomes a thing.

100 agents: the breakpoint where the platform itself needs more sophisticated capabilities — agent discovery for users, agent deprecation processes, portfolio-level cost optimisation, automated evaluation at scale.

Below 5 agents you can run informally. Between 5 and 30 you need basic infrastructure. Between 30 and 100 you need a real AI organisation. Above 100 you need platform engineering invested in the agent operations as their own discipline.

The "too many agents" failure mode

The opposite of too few is too many. Symptoms:

• Multiple agents doing similar work with slight variations, each accruing maintenance cost
• Users not knowing which agent to use for a given task
• Compliance evidence stretched thin because every agent needs separate documentation
• Cost-per-outcome rising because operational overhead grows with agent count
• New agents being created instead of existing ones being improved

This pattern is more common than too few in mature programs. The remediation: agent consolidation reviews quarterly, retirement of low-value agents, merging of overlapping agents, governance gates on new agent creation that ask "could this be done by extending an existing agent?"

The decision questions

For your specific organisation, walk these:

1. What is the highest-value workflow that has not been automated? That is the first agent. The first agent's success criteria drive the second agent.

2. What is the available capacity to operate agents responsibly? This caps the number of agents you can have running well. Underestimating this is the most common sizing error.

3. What is your compliance posture? More agents means more compliance documentation. Mature compliance functions can absorb 30-50 agents with documented governance; thinner functions struggle past 10.

4. Where is the strategic value concentrated? A handful of strategic agents that change the business meaningfully outweigh dozens of small productivity agents in financial terms. Optimise for strategic value first.

5. What does the agent retirement process look like? If you have no retirement process you will accumulate agents indefinitely. Plan retirement from day one.

The portfolio view

Healthy AI programs at scale look like portfolios:

• A handful of strategic agents that drive material business outcomes (5-15% of the estate, 60-80% of the value)
• A larger set of operational agents that handle routine work across domains (40-60% of the estate, 15-30% of the value)
• A long tail of niche agents serving specific needs (30-50% of the estate, 5-10% of the value)

The mix is normal. The mistake is treating all agents as equally important and applying the same operational discipline to the long tail as to the strategic agents. They do not need it; the cost is wasted.

What AgentWorks supports

The platform supports the full sizing range. For small estates (1-10 agents), the platform is overpowered but the cost is reasonable; the operational simplicity is the value. For mid-size (10-50 agents), the platform's per-agent operations are where the platform pays for itself. For large estates (50+), the platform features for agent discovery, portfolio observability, and bulk governance become essential.

The pricing page shows the cost progression at different estate sizes. The honest curve: cost-per-agent decreases as estate size grows, because the platform overhead is amortised across more agents. Above 50 agents the per-agent platform cost is typically dominated by per-run inference costs rather than platform subscription.

The honest sizing answer

Most enterprises end up at 20-50 agents in a mature program, with the exact number determined by industry, scale, and the maturity of the operational organisation. Above 50 is fine for very large organisations with sophisticated AI functions; below 20 is fine for smaller organisations or earlier maturity.

The number itself matters less than the discipline. An organisation with 20 well-operated agents delivers more value than the same organisation with 80 poorly-operated agents. Sizing for governance capacity, not for theoretical scope, is the win that compounds.