Part 1 of the series "What does an AI agent really cost?" — coming up: Microsoft 365 Copilot, Azure AI Foundry Agents and Replit Agents.

For many Swiss companies, Copilot Studio is the fastest path to their own AI agent: connect SharePoint, define topics, publish. The platform hides the complexity of models, orchestration and grounding — but it also hides the cost mechanics. Billing happens neither in tokens nor in users, but in Copilot Credits. And how many credits a single answer consumes depends almost entirely on how the agent is built.

This article works through the official Microsoft prices, verified in August 2026: what goes into a Copilot Studio agent, what it costs at 100, 500 and 1,000+ users — and which levers change the bill by a factor of 5–10.

A note on freshness: all prices are US list prices from Microsoft's official pricing and licensing pages (as of August 2026). CHF amounts are rounded conversions (assumption: 1 USD ≈ 0.80 CHF); your contract's billing is authoritative.

The components: what's inside a Copilot Studio agent

A production Copilot Studio agent is more than the visible chat window:

The platform itself costs nothing — you pay exclusively for usage, measured in credits.

The credit mechanics: an answer is not an answer

Microsoft bills every agent capability at a fixed credit rate. The official rates (August 2026):

The crucial property: credits stack per interaction. A single question to an agent with tenant graph grounding and a generative answer costs 10 + 2 = 12 credits — six times more than the same question to an agent that only uses its SharePoint knowledge source (2 credits). Add one action (say, creating a ticket) and you're at 17 credits.

The animation shows how the credit meter fills depending on agent design:

Three ways to pay

1. 1Pay-as-you-go: $0.01 per credit via an Azure subscription. No fixed fee, no commitment — ideal for pilots and fluctuating load.
2. 2Credit packs: $200 per month for 25,000 credits (annual commitment), pooled tenant-wide. Effectively ~20% cheaper than PAYG ($0.008/credit) — but unused credits expire monthly.
3. 3Pre-purchase (CCCUs): prepaid commit units ($1 = 100 credits) with tiered discounts from 5% (from $3,000) to 20% (from $3M). Unused credits don't expire monthly, and PAYG kicks in automatically as overflow.

And a fourth, often overlooked path: users with a Microsoft 365 Copilot license ($30/user/month) consume no credits at all for standard capabilities — their agent usage is included in the license. More on that below, because it is the biggest lever in the whole calculation.

AIOnomics: why credits are just the packaging

Throughout this series we use a simple mental model that applies to every agent platform — we call it AIOnomics. Every AI agent has three cost layers:

1. 1Tokens — the raw material. Every word in and out of the language model costs fractions of a cent.
2. 2Inference — the processing. Retrieval, ranking, reasoning: more context and more thinking steps mean more model work per answer.
3. 3Harness — the rig around it. Orchestration, security, channels, governance, operations — what turns a model into a product.

Copilot Studio sells all three layers in one unit: the credit. The base formula for any agent platform is:

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

In Copilot Studio the "unit" is the credit, and the biggest multiplier is "units per query" — that is, agent design. Tokens and inference are flat-rated into the credit; you pay a premium for harness convenience, and in return model hosting, grounding infrastructure and operations disappear. Which is exactly why the comparison with self-built RAG stacks matters: there you pay tokens and infrastructure separately — more transparent, but with far more work on your side.

Baseline 1: 100 users — the pilot

Assumptions: internal knowledge agent, 10 queries per user per working day, 20 working days → 20,000 queries per month.

Same user count, same query frequency — a factor of 8.5 between lean and agentic. The lean design fits in two credit packs ($400); the agentic design needs 14 packs, or better yet CCCUs at a discount.

Baseline 2: 500 users — the rollout

100,000 queries per month (same assumptions):

At this size, comparing against the license world becomes mandatory: 500 Microsoft 365 Copilot licenses cost $15,000 per month — and make those users' standard agent usage credit-free. A heavily used, grounding-hungry agent can effectively finance the license, with Copilot in Word, Excel, Outlook and Teams thrown in on top.

Baseline 3: 1,000+ users — enterprise scale

200,000 queries per month and beyond:

For perspective: 1,000 M365 Copilot licenses = $30,000 per month. Which means concretely: a grounding-heavy Copilot Studio agent for 1,000 active users costs roughly as much as the full Copilot license for everyone — a license that additionally zeroes out the agent's credit bill. Beyond this threshold, pure pay-per-use is almost never the cheapest option anymore.

The hidden side costs

• Dataverse capacity: conversation transcripts and agent tables live in Dataverse. Included capacity is limited; expansion starts at $40 per GB per month (database) or $2 per GB (files). For chatty, logging-heavy agents this is a real line item.
• Agent flows with their own triggers: only flows using the "When an agent calls the flow" trigger are credit-free for M365 Copilot licensed users. Scheduled or event-based flows always pay.
• Deep reasoning and premium AI tools: 100 credits per 10 responses — one premium prompt in the wrong place costs fifty times a generative answer.
• Expiring packs: unused pack credits expire monthly. If your load is spiky, CCCUs or PAYG overflow serve you better.
• Bring your own model: your own Azure Foundry models are billed separately — the credit rates only cover Microsoft's built-in models.

The honest calculation: even low-code isn't built for free

The credit bill is only the platform side. A production agent additionally needs build and upkeep effort — whether by internal makers or a partner:

• Build: curating knowledge sources, topic architecture, grounding strategy, testing with real questions — realistically 3–10 person-days, i.e. CHF 4,000–12,000 one-off. Far below a full-stack project (cf. Part 3: 20–60 person-days), but not zero.
• Upkeep: credit reviews, source hygiene, topic tuning and agent governance — around 0.5 person-days per month, i.e. CHF 600. For a pilot with CHF 320 of credit costs, that is the bigger item on the monthly bill.

Six levers to cut the credit bill

1. Only enable tenant graph grounding where it adds value. This item multiplies cost per answer five- to six-fold. For agents with well-defined SharePoint sources, generative answers (2 credits) are almost always enough.

2. Model frequent questions as classic answers. Your top-20 FAQ as predefined topics cost 1 instead of 2–12 credits — and answer more consistently, too.

3. Give heavy users M365 Copilot licenses. From roughly 250 grounded queries per month, the $30 license beats the same user's PAYG credits ($30 / $0.12 per grounded answer); with actions the break-even comes even earlier. The mixed calculation — licenses for heavy users, PAYG for the rest — almost always beats a one-size-fits-all approach.

4. Start on PAYG, switch to CCCUs with volume. Measure first, commit second: the usage estimator plus two or three pilot months provide the data; from $3,000 of pre-purchase you get 5–20% discounts without monthly expiry.

5. Bundle agent flows. 13 credits per 100 flow actions is cheap — but only with the agent trigger. Consolidate flows and trigger them correctly instead of firing per event.

6. Review credit consumption per agent monthly. The admin center shows consumption per agent and capability. A design review after four weeks of real usage almost always finds a grounding or premium-tool item that can be replaced.

Conclusion: the agent design is the price

Copilot Studio has no fixed costs and no infrastructure of its own — in exchange, a consumption price that is controlled almost entirely by agent design, plus a modest but real build and upkeep effort. Between a lean and an agentic design lies a factor of 8–9 at identical usage. If you know the credit rates, you decide the bill while building — not while paying.

In the next part of the series we work through the license side: what Microsoft 365 Copilot really costs — and when the $30 license beats credits.

How CNEXT can help

• Credit forecast before the project: we model your scenarios with the official estimator and realistic assumptions — you know the bill before the agent goes live.
• Design review of existing agents: grounding strategy, topic architecture, flow triggers — typically 30–60% credit savings without quality loss.
• License mix calculation: PAYG, packs, CCCUs or M365 Copilot licenses — we compute the cheapest combination for your user structure.

Before the credit meter becomes a surprise: Get in touch →

Sources (all retrieved August 2026): Microsoft Learn: Copilot Studio billing rates · Microsoft Copilot Studio Licensing Guide · Copilot Studio plans & pricing · Azure: Copilot Studio pay-as-you-go

Further reading:

• Part 2: What Does Microsoft 365 Copilot Really Cost?
• Part 3: What Does an Azure AI Foundry Agent 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?
• AI Company Brain: Architecture, Components and Data Sources
• Agentic AI services