TJE Ventures

Why AI Costs Are Hard to Understand

Most software is priced by seat or subscription. You pay a fixed amount per user per month and you know what you are getting. AI pricing breaks this model. Usage-based pricing means costs fluctuate with how heavily your team uses the tools. New concepts --tokens, context windows, model tiers --require translation before they mean anything to a budget owner. And the market is moving fast enough that pricing changes frequently.

This guide gives you the vocabulary and mental models you need to evaluate AI spend intelligently, without requiring a technical background.

Understanding Tokens

Most AI pricing is based on tokens. A token is roughly equivalent to a word fragment --about four characters of text on average. The sentence you just read contains approximately 25 tokens. Tokens cover both input (what you send to the AI) and output (what the AI sends back), though these are often priced differently.

Pricing is typically expressed as cost per million tokens, often abbreviated as per-MTok or per-1M tokens. Common ranges as of 2025:

• Budget models --smaller, faster models optimized for high-volume tasks: $0.10–$0.50 per million tokens
• Mid-tier models --strong general capability at moderate cost: $1–$5 per million tokens
• Frontier models --the most capable available: $10–$75+ per million tokens, depending on input vs. output

For most business use cases involving document summarization, email drafting, and Q&A, the token cost per individual interaction is a fraction of a cent. Costs scale with volume, not with any single transaction.

The Three Procurement Models

SaaS AI tools (subscription)

This is the familiar model: a per-user monthly fee for access to an AI-powered product. Microsoft 365 Copilot, Salesforce Einstein, HubSpot AI features, and most enterprise AI tools operate this way. You pay a fixed amount per seat regardless of how much each user actually uses the AI.

Best for: predictable budgets, non-technical teams, tools where AI is embedded in a workflow you already use.

Watch for: paying for seats where adoption is low, vendor lock-in, limited ability to customize or control model behavior.

API access (pay-per-use)

Direct API access lets you call AI models programmatically from your own applications and workflows. You pay based on actual usage --the number of tokens processed. Providers like OpenAI, Anthropic, and Google offer API access to their models. This is the foundation of most custom AI implementations.

Best for: developers building custom integrations, high-volume automated workflows, situations where you need precise control over model selection and behavior.

Watch for: costs that scale unexpectedly with usage spikes, the need for engineering resources to build and maintain integrations, prompt injection and other security considerations.

Build your own (self-hosted or fine-tuned)

Organizations with significant AI requirements sometimes choose to self-host open-source models (covered in the Hosting Your Own AI Server article) or fine-tune models on proprietary data. The upfront investment is higher, but the per-query cost is lower at scale, and data privacy is enhanced because nothing leaves your environment.

Best for: high-volume use cases where per-token costs at scale become significant, strict data privacy requirements, scenarios where model customization for a specific domain delivers meaningful accuracy improvements.

Watch for: significant upfront hardware and engineering investment, ongoing maintenance burden, the capability gap between open-source and frontier models.

Total Cost of Ownership: Beyond Token Prices

Token costs are often the smallest part of the real cost of an AI deployment. The full picture includes:

• Engineering time --building integrations, maintaining pipelines, debugging failures
• Prompt development and testing --iterating on prompts until output quality meets the bar
• Human review --staff time spent verifying AI output before it is used
• Infrastructure --hosting, vector databases, logging, and monitoring systems if self-deploying
• Training and change management --getting your team to adopt and use AI tools effectively
• Governance overhead --policy development, compliance reviews, ongoing auditing

A commonly observed pattern: organizations focus intensely on per-token pricing during procurement and discover that the real cost driver is engineering and maintenance. Get the full picture before committing.

How to Estimate Costs

For a usage-based API deployment, start with these inputs:

• Average tokens per interaction --how much text goes in, how much comes back. For most business queries, 500–2,000 tokens per interaction is a reasonable estimate.
• Interactions per day --how many times the AI will be called, across all users and automated processes
• Model selection --which model tier is needed for your quality requirements

Example: 100 users, 10 interactions per day each, 1,000 tokens average per interaction = 1 million tokens per day. At a mid-tier model cost of $3 per million tokens, that is roughly $90/day or $2,700/month. Compare that to 100 Microsoft 365 Copilot seats at $30/user/month = $3,000/month. The economics are often closer than they appear.

Controlling AI Spend

• Match the model to the task --do not use a frontier model for tasks where a budget model performs equally well. Summarizing a short email does not require your most expensive model.
• Set usage limits and alerts --most API providers allow spending caps and budget alerts. Use them. A runaway automated workflow can generate significant costs quickly.
• Cache common responses --if many users ask similar questions, caching the response avoids re-processing the same tokens repeatedly
• Audit low-value usage --periodically review what the AI is actually being used for. High-token, low-value interactions are a spending optimization opportunity.
• Negotiate at volume --enterprise API agreements often include volume discounts, dedicated infrastructure, and committed-use pricing that reduces per-token costs significantly compared to pay-as-you-go rates

“The organizations that control AI costs best are those that measure both what they spend and what they get. Usage without outcome measurement is just expense.”

← Previous Next: AI for Small Business →