“An AI agent is $20 a month”: the promise that wrecks budgets
When a small business owner discovers ChatGPT Business, Microsoft 365 Copilot or Lindy, the sticker prices of roughly $20 to $30 per user per month look exciting. Six months later, the bill piles up multiple subscriptions, LLM usage beyond the plan, custom integrations, tracing tools, EU AI Act compliance work and outside prompt engineering help. The sticker price of an AI agent platform never represents more than 20 to 35 percent of the real 3-year TCO of a small business AI agent program.
This article breaks down the full cost structure of an AI agent program in 2026, by company size (from 1 to 30 agents in production), with real ranges in euros validated on French programs I have supported since 2024. You will find the 7 cost items, the actual list prices of the main platforms (Copilot Studio, CrewAI, LangChain/LangGraph, n8n AI Agent, Lindy), LLM usage rates (OpenAI GPT-4o, Anthropic Claude Sonnet 4.5, Mistral Large, Google Gemini 2.5), the hidden costs to plan for, the budget impact of EU AI Act compliance, and a 3-year ROI calculation showing when an AI agent program is one of a small business’s best digital investments.
This article is part of our series on AI agents for business. To go further, see our comparison of AI agent platforms to choose a tool, and our fractional CIO guide to secure the program’s governance.
The 7 cost items of a small business AI agent program
Every AI agent program, whatever the company’s size or chosen platform, breaks down into 7 main cost items. Understanding this structure is the first step to analyzing a quote, choosing the right architecture and avoiding nasty surprises in years 2 and 3, when the agents scale up and the LLM bill explodes.
Scoping, design and prompt engineering (18 to 28 percent of 3-year TCO)
Initial and ongoing cost of business scoping, agent design, writing system prompts, reasoning chains, tools, memory and fallback rules. In-house: 5 to 25 days per agent depending on complexity, by a trained AI lead or data engineer. Outsourced, in France: €4,000 to €18,000 per simple-to-intermediate agent, €18,000 to €60,000 per business-critical multi-tool agent. Prompt engineering is not a one-off cost: it takes 15 to 25 percent of the team’s time in steady state (model changes, use-case drift, continuous improvement).
Platform licenses and user seats (15 to 25 percent of 3-year TCO)
Recurring SaaS cost billed according to the platform’s model: per user (Microsoft 365 Copilot Business at $21 per user per month billed yearly, Lindy from $29.99 per user per month, Copilot Studio at $200 a month for 25,000 Copilot Credits, as of November 2026), per orchestration seat (CrewAI Enterprise, pricing not published), or included in an existing no-code platform (n8n AI Agent built in, free when self-hosted). For a company with 30 to 80 users and 5 to 10 agents in production, expect €600 to €2,500 a month in licenses on French projects, or €22k to €90k over 3 years.
LLM usage (API tokens) (20 to 35 percent of 3-year TCO)
The item that explodes most in production. The LLM bill depends on the model (GPT-4o, Claude Sonnet 4.5, Mistral Large, Gemini 2.5 Pro), the monthly volume of interactions and the length of prompts and responses. For 5 to 10 active agents in a small business: €350 to €2,800 a month in API tokens. An agent running long-context RAG (vector search over an internal document base) can reach €800 to €3,500 a month on its own. Over 3 years: €18k to €100k of LLM spend for a mature small business, without a token optimization strategy.
Integrations, connectors and vector database (10 to 18 percent of 3-year TCO)
An AI agent only has value when connected to the company’s business tools: CRM (Salesforce, HubSpot), ERP (SAP, Sage, Cegid, a French ERP), SharePoint/Notion/Confluence for RAG, Slack/Teams for delivery. In France, expect €1,500 to €6,000 per custom, non-native connector. For RAG, budget €200 to €1,500 a month for a vector database (Pinecone, Weaviate, Qdrant Cloud) or host it in-house (PostgreSQL + pgvector, self-hosted Qdrant). Over 3 years for a mature small business: €12k to €45k in total on this item.
Observability, tracing and evaluation (5 to 12 percent of 3-year TCO)
No agent should go into production without tracing tools (LangSmith, Helicone, Langfuse, Arize Phoenix) to measure quality, latency, hallucinations and cost per agent. Expect €50 to €400 a month depending on trace volume and number of users. Add continuous evaluation tools (test sets, automated evals, quality drift alerts). This often-forgotten item drives 60 to 80 percent of operational quality and EU AI Act compliance. Over 3 years: €3k to €18k.
Maintenance, governance and in-house FTE (12 to 22 percent of 3-year TCO)
In steady state, a program of 5 to 10 agents needs 0.2 to 0.5 dedicated in-house FTE (updating prompts as models change, quality monitoring, RAG tuning, user support, quarterly governance committee). Over 3 years for a small business in France: €35k to €110k in fully loaded internal cost. Without an identified in-house lead and genuinely freed-up time, this is the number one reason small business AI agent programs fail. Governance is less than 5 percent of TCO but drives 80 percent of real ROI.
EU AI Act, GDPR and security compliance (5 to 10 percent of 3-year TCO)
Starting in August 2026, the EU AI Act requires high-risk AI systems (HR, credit scoring, automated decision-making) to have a register, technical documentation, an impact assessment, robustness testing, human oversight and a user complaint channel. Expect €5k to €12k up front for compliance and €2k to €5k a year to maintain the documentation. Add GDPR requirements (record of processing, a DPA with each LLM vendor, a privacy impact assessment) and security (handling prompt injection, data leaks, jailbreaks). Over 3 years: €10k to €30k.
AI agent pricing grid by company size (3-year TCO)
Here is the reference grid I use when scoping projects with small and mid-sized companies in France. The ranges, in euros, are validated on programs I have supported since 2024 across the 5 main platforms (Copilot Studio, CrewAI, LangChain/LangGraph, n8n AI Agent, Lindy). Assumptions: mixed proprietary/open-source LLMs, a mix of native and custom connectors, in-house FTE at €65k fully loaded per year.
| Program profile | Year 1 setup | Annual recurring cost | 3-year TCO |
|---|---|---|---|
| Very small business - 1 to 2 simple agents | €3k–8k | €3k–7k | €12k–30k |
| Small business - 3 to 5 agents | €12k–28k | €9k–22k | €35k–90k |
| Mature small business - 5 to 10 agents | €22k–55k | €18k–42k | €55k–140k |
| Industrial SMB - 10 to 20 agents | €55k–120k | €40k–95k | €140k–310k |
| Multi-department SMB / mid-sized company - 20 to 30 agents | €120k–220k | €80k–170k | €280k–460k |
These ranges include the 7 items detailed above, including in-house FTE and EU AI Act compliance. The low end corresponds to a rollout on Microsoft 365 Copilot or self-hosted n8n AI with Mistral Large and a junior in-house FTE. The high end corresponds to a CrewAI or LangGraph rollout with GPT-4o and Claude Sonnet, an outside provider for critical agents, and reinforced EU AI Act compliance. To understand the differences between platforms, see our comparison of AI agent platforms.
List prices of the main AI agent platforms (2026)
Pricing models vary radically from one platform to another: per user, per seat, per message, per premium message, or included in a no-code subscription. Understanding these units is essential to avoid getting caught out in year 2. Below are US entry list prices as of November 2026, alongside the monthly spend actually observed in 2026 on French small business projects (in euros).
| Platform | Pricing model | Entry price (US list, Nov. 2026) | Active SMB spend per month (French projects) |
|---|---|---|---|
| Microsoft 365 Copilot | Per user | $21/user (Copilot Business, billed yearly) | €800–2,800 (30–100 users) |
| Copilot Studio | Credit capacity pack | $200/month (25,000 Copilot Credits) | $400–1,600 (50k–200k Copilot Credits) |
| CrewAI Enterprise | Per seat / crew | Pricing not published | €800–2,200 (10–20 seats) |
| LangChain LangSmith | Per user + traces | $39/user (Plus) | €200–900 (5–15 users) |
| n8n AI Agent (Cloud) | Per execution | $20–50 (2,500–10,000 executions) | €200–700 (50k–200k executions) |
| n8n AI self-hosted | Open license + infrastructure | €10–50 (VPS) | €200–600 (infra + DevOps) |
| Lindy | Per user + credits | From $29.99/user (Team, 3,000 credits) | $250–1,000 (5–15 users) |
Important: add these on top
None of the prices above include LLM usage (API tokens), except Microsoft 365 Copilot and Lindy, which include a limited allowance. For 5 to 10 active agents in a small business, expect an extra €350 to €2,800 a month in OpenAI, Anthropic or Mistral tokens on French projects. This line is often the largest budget item in years 2 and 3. See the detailed LLM grid below.
LLM API prices for AI agents (as of November 2026)
The choice of LLM drives 30 to 60 percent of an agent’s operating cost. Price gaps between models can reach a factor of 10 or more, for nearly equivalent quality on many use cases. Here are the official US list prices as of November 2026 and the recommended uses for a small business.
| LLM model | Input price (1M tokens, USD) | Output price (1M tokens, USD) | Recommended use |
|---|---|---|---|
| OpenAI GPT-4o | $2.50 | $10 | Complex multi-tool agents |
| OpenAI GPT-4o mini | $0.15 | $0.60 | Classification, routing, simple tasks |
| Anthropic Claude Sonnet 4.5 | $3 | $15 | Document analysis, reasoning |
| Anthropic Claude Haiku 4.5 | $1 | $5 | Real-time agents, tier-1 support |
| Mistral Large (current version) | $0.50 | $1.50 | EU data sovereignty, strict GDPR |
| Google Gemini 2.5 Pro | $1.25 | $10 (prompts up to 200k tokens) | Long context (1M tokens), multimodal |
| Llama 3.3 70B (self-hosted) | 0 (license) | €400–2,500/month GPU (in France) | Very high volumes, sensitive data |
Optimization strategy: a well-designed agent uses a premium model (GPT-4o or Claude Sonnet 4.5) only for complex decisions, and hands 70 to 85 percent of calls to a low-cost model (GPT-4o mini, Claude Haiku 4.5, Mistral Small) for classification, routing and secondary tasks. This multi-model routing typically cuts the LLM bill by a factor of 3 to 5, with no noticeable degradation for the end user. For a company with EU data constraints, Mistral Large offers an excellent quality/price ratio with EU hosting and data sovereignty.
ROI calculation: a 3-year worked example for an 80-employee company
The ROI of an AI agent program rests mainly on three levers: automating repetitive cognitive tasks, raising the productivity of targeted employees, and speeding up business cycles. Here is the detailed calculation for a French B2B company with 80 employees (revenue €22M, a 15-person sales team, a 6-person tier-1 support team, an 8-person finance team), with a program of 7 critical agents live by the end of year 1.
Direct operational gains over 3 years
| Gain lever | Annual gain | 3-year total |
|---|---|---|
| Lead qualification agent (1.5 sales FTE freed up) | €95,000 | €285,000 |
| Tier-1 support agent (40 percent of tickets handled with no human) | €78,000 | €234,000 |
| Internal document search agent (productivity +35 percent) | €52,000 | €156,000 |
| Invoice and expense report analysis agent | €38,000 | €114,000 |
| Sales proposal and reporting generation agent | €42,000 | €126,000 |
| Total operational gains over 3 years | €305,000/year | €915,000 |
Strategic benefits over 3 years
| Benefit | Annual value | 3-year total |
|---|---|---|
| Faster sales time-to-quote (-65 percent) | €28,000 | €84,000 |
| Better customer satisfaction (NPS +12 points) | €45,000 | €135,000 |
| Hires avoided (1 support FTE, 0.5 sales admin FTE) | €82,000 | €246,000 |
| Total strategic benefits over 3 years | €155,000/year | €465,000 |
3-year ROI summary
Total AI agent program cost: €120,000 (3-year TCO, mature small business profile with 5 to 10 agents, see the pricing grid).
Cumulative 3-year benefits: €915k (operational gains) + €465k (strategic benefits) = €1,380,000.
Net 3-year benefit: €1,260,000.
Break-even point: 10 months after the first rollouts. 3-year ROI: 1,050 percent (an 11.5x multiple). Key assumption: real adoption above 70 percent and active governance. Without governance, ROI can drop to 220 percent (still positive). These figures explain why deploying AI agents has become one of the highest-multiple digital investments for small businesses, provided the priority use cases are scoped properly up front.
The 6 hidden costs AI agent quotes never show
Beyond the 7 official items, here are 6 costs that are systematically hidden or downplayed in quotes and published pricing, and that can weigh up to 40 percent of real 3-year TCO.
Hidden cost 1: LLM tokens exploding in production
The classic scenario: an agent planned for 500 interactions a month in the POC ends up at 5,000 in real production, with context 3 times longer than expected (RAG, conversational memory, history). The LLM bill jumps from €80 to €1,200 a month. Over 3 years, a poorly sized agent can generate €20k to €40k in unplanned extra cost. Solution: systematic per-agent cost tracing from go-live, threshold alerts, multi-model routing (GPT-4o mini for 70 percent of calls, GPT-4o for the critical 30 percent).
Hidden cost 2: shadow AI and duplicate subscriptions
Without central governance, every department ends up subscribing to ChatGPT Business, Claude Pro, Copilot or Perplexity (marketing, HR, sales, sales admin, management). A 100-employee company can end up with 5 to 8 AI subscriptions running in parallel, or €800 to €3,000 a month more than under central governance. Typical small business audit: 30 to 45 percent immediate savings by consolidating. Add GDPR and data leak risks when everyone pastes internal documents into a personal account.
Hidden cost 3: prompt maintenance and quality drift
AI agents are not “set and forget.” Models change (GPT-4o to GPT-4.5, then GPT-5), business use cases shift, quality drifts. Expect 15 to 25 percent of the AI team’s time in steady state for prompt maintenance, continuous evaluation and tuning reasoning chains. Over 3 years for a mature small business in France: €35k to €90k in fully loaded cost, rarely budgeted at the start. Solution: build in an automated evaluation pipeline and an identified lead FTE from day one.
Hidden cost 4: EU AI Act compliance (August 2026)
Starting in August 2026, the EU AI Act requires high-risk AI systems to have a central register, technical documentation, an impact assessment, robustness testing, human oversight and a complaint channel. Expect €5k to €12k up front for compliance, plus €2k to €5k a year to maintain the documentation. AI Act penalties: up to €35M or 7 percent of worldwide revenue. Most small businesses I see during scoping had planned for neither the timeline nor the budget of this compliance. Planning ahead avoids a costly last-minute scramble.
Hidden cost 5: ongoing user training
An AI agent deployed without training business users stays underused: adoption typically sits at 30 to 40 percent instead of the targeted 75 to 85 percent. Initial training (1 to 2 hours per user) plus ongoing follow-up (monthly sessions, use-case documentation, internal ambassadors) costs €3k to €8k a year for a small business in France. It is one of the biggest ROI swings: doubling adoption nearly doubles the program’s ROI. Never deploy without an identified training and change management plan.
Hidden cost 6: prompt injection security and data leaks
AI agents connected to business tools (CRM, ERP, mailbox) are new attack targets: prompt injection (manipulation through a booby-trapped document), jailbreaks (bypassing guardrails), data exfiltration through the agent’s responses. Putting serious protection in place (input sanitization, monitoring sensitive outputs, a sandbox for tool execution, least privilege) costs €3k to €10k of initial scoping and €1k to €3k a year in monitoring. Underestimating this item exposes you to incidents with an average remediation cost above €50k per incident.
8 levers to cut AI agent TCO by 35 to 55 percent
A well-thought-out scoping, purchasing and governance strategy can lower the 3-year TCO of an AI agent program by 35 to 55 percent, or €25k to €200k for a company with 50 to 150 employees. Here are the most effective levers I use with my small business clients.
Multi-model LLM routing
Reserve premium models (GPT-4o, Claude Sonnet 4.5) for the 20 to 30 percent of critical calls and route the remaining 70 to 80 percent to GPT-4o mini or Claude Haiku 4.5 (about 3 to 17 times cheaper at US list prices). Typical savings on the LLM bill: 50 to 70 percent with no noticeable degradation. The number one lever for any high-volume program.
Caching system prompts and static context
OpenAI, Anthropic and Mistral now offer prompt caching, which cuts the cost of repeated input tokens (system prompt, fixed RAG base, few-shot examples) by a factor of 5 to 10. Turning caching on properly can save 30 to 50 percent of the LLM bill on high-volume agents. A few hours of setup for lasting savings.
Focus on priority use cases
Rather than 15 scattered pilot agents, deploy 3 to 5 priority agents with high volume and high ROI. Concentrate the prompt engineering, observability and training budgets. Savings over the first 18 months: 30 to 50 percent compared with a diluted strategy. Once the first agents are running smoothly and profitably, expand gradually.
Centralized AI subscriptions and catalog
A single central subscription (Microsoft 365 Copilot Enterprise or ChatGPT Enterprise) rather than scattered accounts. A documented catalog of all active agents, each with a business owner. Savings versus shadow AI: 30 to 45 percent, immediately. A quarterly committee to prioritize new requests (IT, business, finance).
Mistral and EU models for data sovereignty
For agents handling sensitive data or customers with strict EU requirements, Mistral Large (hosted in France) offers an excellent quality/price ratio: at US list prices ($0.50 / $1.50 per million tokens versus $2.50 / $10 for GPT-4o), roughly 80 percent cheaper than GPT-4o, with native EU hosting and easier GDPR compliance. It also avoids transferring data to the US and the complex impact assessments tied to Schrems II (the EU court ruling on EU–US data transfers). A differentiating lever in regulated B2B.
Self-hosted open source for extreme volumes
For an industrial SMB or mid-sized company above 5 million tokens a month, hosting Llama 3.3 70B or Mistral Mixtral 8x22B on a dedicated GPU (European hosts Hetzner, OVHcloud, Scaleway) becomes cheaper than proprietary APIs. Total GPU infrastructure plus outsourced DevOps cost in France: €6,000 to €18,000 a year. Versus an equivalent GPT-4o setup: savings of €30k to €100k a year. Prerequisite: an in-house technical team or a trusted DevOps provider.
Pooling with a no-code automation project
A company that already runs self-hosted n8n or Make in production can add its first AI agents through n8n AI Agent or Make AI without a new platform. Governance, connectors and the technical lead are shared. Savings at startup: €8k to €25k. See our no-code automation guide.
Independent upfront scoping (the most profitable lever)
Bring in an independent consultant or an outside fractional CIO (in France, €8k to €18k of fees) to: map candidate use cases and prioritize the 3 to 5 most profitable, size LLM volumes over 3 years, choose the right platform and architecture, run a competitive vendor consultation, structure governance and the catalog, anticipate EU AI Act compliance, and secure ROI with a defensible business case. A modest cost with a typical ROI: net savings of €40k to €150k on the program and a time-to-value 4 to 8 months faster. It is the investment with the best multiple in any small business AI agent program.
Frequently asked questions about the cost of AI agents for business
Is Microsoft 365 Copilot’s $21 per user per month price representative of the real cost?
Partly. Microsoft 365 Copilot Business, at $21 per user per month billed yearly as of November 2026, includes the assistant built into Word, Excel, Outlook and Teams and access to Microsoft’s LLMs. But as soon as you want to build custom business agents (lead qualification, tier-1 support, RAG document search), you need Copilot Studio ($200 a month for a pack of 25,000 Copilot Credits, plus pay-as-you-go beyond). Add the costs of scoping, prompt engineering, CRM/ERP integrations and governance. On French projects, for a 50-user company with 3 to 5 custom agents, the real monthly cost lands between €2,200 and €4,500, several times the per-user sticker price.
Should you choose a proprietary LLM (OpenAI, Anthropic) or self-hosted open source (Llama, Mistral)?
The rule of thumb: start with a proprietary LLM (GPT-4o, Claude Sonnet 4.5, Mistral Large via API) for the first 12 to 24 months, then evaluate self-hosting once volumes have stabilized and use cases have matured. Self-hosting pays off above 5 to 8 million tokens a month OR if you have very strict sovereignty constraints (healthcare, defense, ultra-sensitive data). Below that, managed SaaS remains cheaper once you factor in GPU infrastructure, monitoring, backups and in-house MLOps skills. For details, see our comparison of AI agent platforms.
What public funding can help finance an AI agent program?
In France, several programs exist in 2026. The France 2030 AI plan offers regional grants (Île-de-France, Auvergne-Rhône-Alpes, Hauts-de-France, Grand Est) covering 25 to 45 percent of the cost of a structuring AI project. The Research Tax Credit (Crédit d’Impôt Recherche) can cover 30 percent of R&D spending on developing innovative AI agents. France Num assessments, run through the chambers of commerce (CCI), support initial scoping free of charge. Bpifrance, the French public investment bank, offers French Tech Émergence loans and Deep Tech grants for innovative programs above €80k. Bpifrance’s 2025-2027 SME AI Plan specifically targets industrial SMBs with grants of up to €60k.
How much does EU AI Act compliance cost a small business?
It depends on how your agents are classified. For limited-risk AI systems (customer chatbot, productivity assistant, content generation), only transparency obligations apply (telling users they are interacting with an AI), at a marginal cost below €1k. For high-risk systems (HR, credit scoring, automated decision-making): a register, technical documentation, an impact assessment, robustness testing, human oversight and a complaint channel, or €5k to €12k up front and €2k to €5k a year. Most AI agents deployed in small businesses fall under limited risk, but you need to classify them systematically and document the decision.
How do you keep an AI agent program from drifting into uncontrolled shadow AI?
Five concrete rules make the difference. First, a written company AI policy signed by leadership (what is allowed, what is forbidden, governance, sensitive data). Second, a single central company subscription rather than personal accounts (ChatGPT Enterprise, Copilot, Claude Team). Third, a documented catalog of all active agents, each with an identified business owner and a description sheet. Fourth, a quarterly committee to prioritize new requests (IT, business, finance, legal). Fifth, an annual review of each agent to validate its use, quality and compliance (regularly pruning the 15 to 25 percent of agents that have become obsolete). This governance costs less than 5 days a year but prevents 30 to 50 percent of budget drift and 80 percent of GDPR incidents.