The vendor quote is rarely the cost. A platform that lists at $5K/month often has a real TCO of $12K to $20K/month once integration build, model usage, maintenance, governance, and change management are added. This piece is the TCO model that surfaces those numbers before the contract is signed. Companion to ROI calculator, cost attribution, and agent pricing.
Six TCO components
The full TCO of an agent breaks into six categories. The mix varies widely by use case; the categories do not.
- Platform fees. What the vendor charges you.
- Model usage. Token costs if not bundled into platform fees.
- Integration build. One-time engineering to connect the agent to your systems.
- Ongoing maintenance. Prompt tuning, eval, integration drift, model upgrades.
- Governance and security. Compliance reviews, audits, security tooling, DPA review.
- Change management. Training, process redesign, handling team resistance.
For most internal use cases, the typical mix is platform 25-40%, model 15-30%, integration (amortized over 3 years) 10-20%, maintenance 10-15%, governance 5-10%, change management 10-15%. The variance is large; the rule is to count every category, not just the invoice-visible ones.
Platform fees
The line item the vendor sells. Five pricing families and how to compute them at your real volume.
- Per-seat. Number of users × monthly rate × 12. Easy to project; misaligned for backend agents.
- Per-run. Runs per month × per-run rate × 12. Variable; can spike if volume grows faster than expected.
- Per-token (pass-through plus markup). Tokens per month × per-token rate × 12. Most variable; highly sensitive to use case (RAG agents consume far more tokens than simple Q&A).
- Credit/unit. Credits per month × cost per credit × 12. Pre-purchase volume gives discounts; unused credits often expire.
- Flat enterprise fee. Annual contract. Predictable, only economical above a usage threshold.
Read the contract for overage rates. A vendor whose base fee is $5K/month but whose overage is $0.50/run at 30 percent above quota can easily double the bill.
Model usage costs
If model costs are not bundled, calculate separately. For a customer-facing agent on a top-tier model (Claude Sonnet 4 class, GPT-4 class):
- Average tokens per run: input 4,000 + output 500 = 4,500.
- Approximate cost per million input tokens: $3 (Claude Sonnet 4.5 input rate); per million output: $15 (Anthropic pricing, 2025).
- Per-run cost: (4,000 × $3 + 500 × $15) / 1,000,000 = $0.0195.
- At 100,000 runs/month: $1,950/month, or $23,400/year.
Adjustments. Prompt caching can reduce input costs by 50 to 90 percent for repeated context, depending on hit rate (Anthropic prompt caching, 2025). RAG-heavy workloads with long retrieval contexts push the per-run number higher. Tool-call agents with multiple model invocations per run multiply the cost.
Integration build
One-time engineering to connect the agent to your systems. Per-integration estimates:
- Native vendor integration. 0 to 1 week to configure. Usually OAuth setup, permissioning, testing.
- Partner connector via the platform. 1 to 3 weeks. Connector exists; you handle field mapping, error handling, and edge cases.
- Custom integration. 2 to 8 weeks per integration. SDK or framework provided; you write the connector, auth, error handling, retries, schema mapping.
- Greenfield integration to a closed system. 6 to 16 weeks. Includes API design or scraping logic for systems without one.
Convert weeks to dollars using your loaded engineering rate (typically $200 to $400/hour fully loaded). A 4-week custom integration at $300/hour fully loaded = $48,000 one-time. Amortize over 3 years for the annualized number.
Ongoing maintenance
The line item most TCO calculations miss. Annual maintenance for a deployed agent.
- Prompt tuning and eval. 1 to 5 hours/week per agent capability. Model behavior drifts; new edge cases emerge; the eval suite needs new tests.
- Integration drift. Vendor APIs change. Auth tokens rotate. Schemas evolve. Budget 1 to 3 hours/month per integration.
- Model upgrades. Each new model snapshot needs eval, possibly prompt updates. 1 to 3 days per upgrade × 2 to 4 upgrades per year.
- Bug triage from on-call. Issues discovered in production that require platform-level fixes. Variable; 0.5 to 2 days/month typical.
For a single agent with 3 integrations, maintenance typically runs 0.2 to 0.5 FTE per year of engineering time. At $300/hour loaded, that is $125K to $310K/year of engineering cost per agent at scale.
Governance and security
The cost of compliance and risk management.
- Security review of new vendor. 40 to 120 hours one-time.
- DPA review and negotiation. 20 to 60 hours one-time, mostly legal.
- Annual SOC 2 / ISO review of platform. 10 to 20 hours/year.
- Internal compliance audit support. 20 to 40 hours/year.
- Privacy impact assessment. 20 to 60 hours one-time, more if scope expands.
- Ongoing access reviews, log monitoring. 5 to 10 hours/month, depending on regulatory posture.
Higher for regulated industries (healthcare, finance). The EU AI Act adds documentation and conformity assessment requirements for high-risk applications (EU AI Act Article 17).
Change management
The cost of getting humans to use the agent. Easily 10 to 20 percent of Year 1 TCO. Components.
- Training. Sessions, documentation, ongoing support. 4 to 16 hours per affected user.
- Process redesign. Mapping how work changes when the agent handles part of it. Days of cross-functional time.
- Communications. Internal launches, FAQ updates, escalation channels.
- Resistance handling. Some teams resist agent deployment. The cost is meetings, executive air cover, and patience.
Underbudget change management and the agent ships but does not get used; ROI fails not because the agent is bad but because nobody adopted it.
Three-year TCO example
A mid-size customer support agent serving 8,000 tickets/month.
Year 1.
- Platform fees: $60,000
- Model usage: $30,000 (passthrough)
- Integration build (3 integrations, $48K each): $144,000 one-time, amortized $48,000/yr
- Maintenance: $25,000
- Governance: $25,000 (incl. one-time PIA + security review)
- Change management: $30,000 one-time + $5,000 ongoing
- Year 1 total: $223,000
Year 2-3 run-rate.
- Platform fees: $60,000
- Model usage: $30,000
- Integration amortization: $48,000
- Maintenance: $25,000
- Governance: $10,000
- Change management ongoing: $5,000
- Year 2-3 total: ~$178,000/year
Three-year TCO: ~$579,000. Annualized: ~$193,000.
Hidden costs vendors do not list
Six costs that almost never appear on a vendor quote.
- Eval infrastructure. Building the eval suite, running it weekly, monitoring quality drift. 0.5 to 1 FTE per platform.
- Tenant isolation testing. For multi-tenant platforms, ongoing tests that one tenant cannot see another's data. Quarterly to monthly.
- Vendor migration option cost. What it would cost to migrate off this vendor. Compute it; it scopes your real lock-in.
- Premium support tier. Often required for production SLAs; often 20 to 40 percent above the base.
- Sandbox or staging environment fees. Some vendors charge for non-prod environments. Read the small print.
- Data egress. When the platform stores your audit logs and prompts, exporting them out can carry data transfer fees.
FAQ
- What is the total cost of ownership for an AI agent?
- The all-in cost over its useful life. Six components: platform fees, model usage, integration build, ongoing maintenance, governance/security, and change management. Hidden costs add 40 to 80 percent on top of the invoice.
- What is the breakdown of agent platform TCO?
- Platform 25-40%, model 15-30%, integration 10-20%, maintenance 10-15%, governance 5-10%, change management 10-15%. Mix varies by use case.
- Which agent platform costs are most often underestimated?
- Integration build and ongoing maintenance. Custom integrations run 2 to 8 weeks. Maintenance runs 10 to 20 percent of integration build per year.
- How long should I project TCO?
- Three years. Year 1 one-time-heavy; Years 2-3 run-rate. Beyond 3 years only for capital allocation.
- Does model cost dominate agent TCO?
- Sometimes. For high-volume customer-facing agents, model usage can hit 30 percent. For internal agents, model is often 10 to 15 percent; integration and maintenance dominate.
- How do I compare TCO across vendors with different pricing models?
- Normalize to monthly cost at your real projected volume across all 6 components. Add vendor-specific hidden costs.
Sources
- Anthropic, "Claude API pricing", 2025, anthropic.com
- Anthropic, "Prompt caching", 2025, docs.anthropic.com
- European Commission, "EU AI Act Article 17: Quality management system", 2024, artificialintelligenceact.eu
- Forrester, "Total Economic Impact methodology", 2024, forrester.com
- Gartner, "How to Calculate the TCO of AI", 2024, gartner.com
