The Cost of AI: How to Think About Pricing and ROI

# The Cost of AI: How to Think About Pricing and ROI AI pricing is confusing. Token costs, per-seat fees, enterprise contracts, free tiers — how do you figure out what AI actually costs and whether it's worth it? ## The three pricing models in 2025 ### Model 1: Per-token / per-API-call You're charged based on usage. This is the pricing for raw AI model APIs (OpenAI, Anthropic, Google) and for tools that pass through those costs. The unit is "tokens" — roughly 0.75 words for English text. Pricing in 2025 has come down dramatically: - GPT-4o class models: $2-5 per million input tokens, $8-15 per million output tokens - Claude Sonnet 4: $3 per million input, $15 per million output - Smaller/faster models (GPT-4o-mini, Claude Haiku, Gemini Flash): $0.10-0.30 per million input What this means in practice: a typical AI interaction (summarize a document, draft an email) costs fractions of a cent. Even a heavy user spending 4 hours/day on AI might cost $20-50/month in API fees. ### Model 2: Per-seat subscription You're charged a flat monthly fee per user. This is the model for most end-user tools (ChatGPT Plus, Claude Pro, GitHub Copilot, Notion AI). Typical pricing: $20-30 per user per month for individual tools. Enterprise tiers: $30-60 per user per month. The advantage: predictable costs. The disadvantage: you're paying regardless of usage, so heavy and light users pay the same. ### Model 3: Enterprise / usage-based For high-volume enterprise deployments, vendors offer custom pricing that often combines a base subscription with usage overages. This is where deals can get complex, but the total cost is usually 20-40% less than per-seat at high volumes. ## The hidden costs everyone forgets The sticker price is the easy part. Here are the costs that often get missed: ### Implementation cost For any non-trivial AI deployment, you need engineering time to integrate the tool, set up the data flows, configure permissions, and handle edge cases. For a mid-sized deployment, this is typically 2-8 weeks of engineering time. For off-the-shelf tools, this is much lower (hours, not weeks). For custom AI applications, it's much higher. ### Change management cost Training people to use new tools, updating workflows, gathering feedback, iterating. This is often the largest hidden cost. A 200-person organization rolling out an AI tool might spend 50-100 person-days on change management in the first year. ### Data preparation cost AI quality depends on data quality. If your data is in 15 different systems and inconsistent, you need data preparation work before the AI will be useful. ### Maintenance cost AI tools need ongoing maintenance. Model updates can change behavior. Edge cases need to be handled. Bugs need to be fixed. This is usually 10-20% of the initial implementation cost per year. ### Vendor lock-in cost If you build your AI strategy around a single vendor's tools and pricing changes, you have a problem. The lock-in cost is real but hard to quantify upfront. ## How to calculate ROI The honest ROI calculation is straightforward but requires discipline: ### Step 1: Define the current state Measure how long the current process takes, what it costs, and what the error rate is. Be specific. "Customer support" is too vague; "Average time to resolve a tier-1 support ticket" is specific. ### Step 2: Define the target state With the AI tool, what's the new time, cost, and error rate? Use the vendor's case studies as a starting point, but adjust based on your own pilot results. ### Step 3: Calculate the per-use savings For each use of the AI tool, calculate the savings. Example: - Current: 30 minutes per task, $25 in labor cost - With AI: 8 minutes per task, $7 in labor cost - Per-use savings: 22 minutes, $18 ### Step 4: Multiply by volume How many times per month will this be used? If your team handles 500 support tickets a month and AI helps on 70% of them, that's 350 AI-assisted tickets per month. Monthly savings: 350 × $18 = $6,300 Annual savings: $75,600 ### Step 5: Subtract all costs Subtract the AI tool cost, implementation cost, and change management cost. What's left is the net annual benefit. Net annual benefit: $75,600 - $36,000 (tool) - $20,000 (implementation, amortized) = $19,600/year ### Step 6: Add the value of quality improvements The above calculation only captures time savings. If the AI also improves quality (e.g., fewer errors, better customer satisfaction), add an estimate of that value. This is harder to quantify but real. ## When AI is NOT worth it Despite the hype, there are plenty of cases where AI doesn't make economic sense: **Low-volume workflows**: if you only do something 10 times a year, the time savings don't justify the tool cost. **Workflows where errors are catastrophic**: if the cost of an AI error is $10,000 and the AI makes errors 1% of the time, the expected cost of errors may exceed the savings. **Workflows that don't have measurable outputs**: if you can't measure the current state, you can't measure improvement. Some work is too qualitative to benefit from AI ROI calculations. **Workflows where the human touch is the value**: customer relationships, creative direction, complex negotiations. AI can support these but not replace them, and the value of the human element is real. ## The honest answer to "is AI worth it" For most mid-sized organizations in 2025, the answer is yes, with three caveats: 1. **Focus on workflows, not tools**: pick the workflow, then find the tool, not the other way around 2. **Measure from day one**: if you can't measure, you can't prove ROI 3. **Plan for the hidden costs**: implementation, change management, and maintenance often exceed the tool cost The companies getting real ROI from AI aren't the ones with the most sophisticated deployments. They're the ones with disciplined measurement, clear use cases, and a willingness to walk away from tools that don't deliver.