Part 1 of this series dissected Copilot Studio's credit mechanics, Part 2 the license logic of Microsoft 365 Copilot. Both platforms hide consumption behind abstractions — credits or flat-rate seat licenses. Azure AI Foundry is the other end of the spectrum: no abstraction at all. Every token, every retrieval, every tool call has its own meter.

That is a blessing and a curse: maximum control and maximum transparency — but only for those who actually read the price list. This article works out what a custom-built agent on Azure AI Foundry costs at 100, 500 and 1,000+ users, using list prices verified in August 2026.

Note: All prices are USD list prices (West Europe region, as of August 2026) from the official Azure pricing pages. CHF amounts are rounded conversions (1 USD ≈ 0.80 CHF); your Azure subscription bills in USD.

The component inventory: what a Foundry agent actually needs

A production Foundry agent is never just "the model". The typical stack:

Notable: the Agent Service itself costs nothing. Microsoft earns on the consumption the agent triggers — the free-printer-expensive-ink logic. And the most expensive single meter per unit is not the model but web search: $14 per 1,000 transactions is roughly ten times a typical GPT-5 answer.

AIOnomics: the complete metering chain of one agent turn

The series formula applies here too — except that on Foundry every variable is its own Azure meter:

Monthly cost = users × queries/user × units/query × price/unit

The animation shows the metering chain of a single agent answer — a meter ticks at every station from prompt to response:

Let's price one single agent answer. Assumptions: system prompt + conversation history + retrieved context add up to ~4,000 input tokens, the answer to ~500 output tokens. The harness multiplier — tool calls, retrieval planning, retries — typically sits at 1.5–2× in well-built agents; we use 1.6×:

Two cents per answer sounds harmless. But here lies the Foundry trap: the price per answer is not a constant — it is the product of design decisions. An agent with a bloated system prompt (8,000 instead of 2,000 tokens), web grounding on every question (+$0.014) and three tool iterations instead of one quickly lands at $0.06–0.10 per answer — a factor of 3–5, with no visible difference to users.

Baseline 1: 100 users

Assumptions: internal knowledge agent, 30 queries per user per month (3,000 queries), small index (Basic suffices), hosting on a B1 plan, logs within the free quota.

CHF 120 per month for 100 users — barely more than a franc per head. Unbeatable against licenses ($30/user would be $3,000). But: this figure contains zero staff cost. Build (20–60 person-days) and operations (1–2 person-days/month) dominate the total at this size — more below.

Baseline 2: 500 users

15,000 queries/month, larger index (S1), staging environment, P0v3 hosting, ~10 GB logs.

The structure shifts: at 100 users 59% of the cost was fixed, now it is 57% — the variable side grows linearly, the fixed side in steps. The price per user drops to ~$1.40.

Baseline 3: 1,000+ users

30,000 queries/month, S1 with 2 search units (availability), staging on S1, P1v3 hosting, ~20 GB logs.

The consumption curve across all three baselines is shown in the animation — a fixed base plus linearly growing usage:

The platform comparison in francs per user per month:

Foundry wins the pure infrastructure comparison hands down. Why not every project belongs on Foundry anyway is what the next section clarifies.

The honest calculation: staff cost beats infrastructure

With Copilot Studio and M365 Copilot, the premium buys you the operations. With Foundry it doesn't:

• Build: chunking strategy, permission model, evaluation, guardrails — realistically 20–60 person-days, i.e. CHF 25,000–75,000 one-off.
• Operations: pipeline maintenance, model updates, cost and quality monitoring — 1–2 person-days per month, CHF 1,200–2,400 monthly. That exceeds the entire Azure bill in all three baselines.

Rule of thumb: Foundry pays off when (a) your requirements fall outside what Copilot Studio can do natively — custom models, custom orchestration, non-M365 data sources, strict latency or compliance constraints — or (b) usage volume is large enough that the infrastructure advantage outweighs the operating effort.

Seven cost levers for Foundry agents

1. Model tiering. Not every query needs GPT-5. A router that sends routine questions to a mini model (10–25× cheaper per token) typically cuts inference cost by 40–70%.

2. Enforce a prompt budget. Input tokens dominate the bill. Tighten the system prompt, summarize history instead of dragging it along, cap the context window — every 1,000 input tokens saved per query saves $37.50 a month at 30,000 queries.

3. Cache answers. 20–40% of internal queries are repeats. A cache saves at the most expensive point: output tokens.

4. Measure the harness multiplier. Tool retries and redundant iterations are invisible cost drivers. If you don't measure the multiplier, you pay it blindly.

5. Size the index to actual need. The S1 → S2 jump quadruples the fixed cost. Partitions can be added later — measure first, scale later.

6. Ration web grounding. $14 per 1,000 searches is the most expensive meter in the stack. Trigger web search only when the internal index has no answer.

7. PTU only at sustained high load. Provisioned throughput (reserved capacity) pays off only with consistently high, predictable utilization — before that, pay-as-you-go is almost always cheaper.

Conclusion: full control, full responsibility

Azure AI Foundry is the most transparent and, per unit, cheapest way to run an AI agent — CHF 120 to 1,200 of infrastructure per month for 100 to 1,000 users. The price is responsibility: every design decision carries a price tag, and operations are entirely yours. Understand the metering chain and pull the seven levers, and you run agents at a fraction of license cost. Ignore them, and you build yourself an open-ended black box.

In the final part of the series we price Replit Agents — and draw the overall comparison across all four platforms.

How CNEXT can help

• Foundry cost assessment: we model your scenarios with current list prices — before the project starts.
• Architecture decision: Copilot Studio, M365 Copilot or Foundry — justified by your requirements and data sources.
• Cost optimization of existing agents: model tiering, prompt budgets, caching — 30–60% savings are typical.

Before you build, we do the math: Get in touch →

Sources (all retrieved August 2026): Foundry Agent Service pricing · Azure OpenAI pricing · Azure AI Search / Foundry IQ pricing · App Service Linux pricing · Azure Monitor pricing

Further reading:

• Part 1: What Does a Copilot Studio Agent Really Cost?
• Part 2: What Does Microsoft 365 Copilot Really Cost?
• Part 4: What Do Replit Agents Really Cost? — with the Grand Platform Comparison
• The Hidden Cost of RAG in Azure
• Building a Company Brain: Copilot Studio, Azure AI Foundry or Replit?
• Agentic AI services