AI can reduce manual work, improve decisions, and help your team move faster. It can also create an invoice that looks nothing like the budget you approved.
That surprise is token shock.
The problem is not always fraud, waste, or a billing error. Your AI system may simply be doing more work than you realize. A single request can trigger planning, document retrieval, tool calls, retries, quality checks, and multiple model responses.
You see one answer. The meter sees ten or twenty AI calls.
For small and mid-size businesses, this matters. AI usage increasingly behaves like variable cloud infrastructure: not a fixed software subscription. You need visibility, limits, and a cost model before your AI workflow becomes a financial liability.
WHAT ARE TOKENS?
Tokens are the small units of text that an AI model processes. A token may be a word, part of a word, punctuation, or a piece of code.
AI providers typically charge separately for:
- Input tokens the model reads
- Output tokens the model generates
- Cached tokens reused from previous requests
- Requests made by tools, agents, or background processes
Longer prompts cost more. Larger documents cost more. Detailed responses cost more. Repeated conversations cost more.
The important point is simple: AI cost is based on everything the system processes, not only what the user types.
WHY AI BILLS EXPLODE
1. AGENTIC WORKFLOWS MULTIPLY MODEL CALLS
A standard chatbot may process one question and return one response.
An AI agent may:
- Interpret the request
- Create a plan
- Search internal documents
- Call a business application
- Review the result
- Correct an error
- Generate a final answer
Each step can create additional token usage.
Research cited by Forrester suggests that a single agent may use about four times as many tokens as a normal chat interaction. Multi-agent workflows may use approximately 15 times as many.
The user still experiences one task. Your cloud bill reflects the entire workflow.
2. CONTEXT KEEPS GROWING
AI systems need context to provide useful answers. That context may include conversation history, customer records, product documentation, policies, or database results.
The longer the context, the more the model must read each time.
This creates a common cost problem: a workflow keeps sending the same large instructions or documents with every request. Nothing appears to have changed, but the input token count steadily increases.
A useful question is not just, “How many AI requests did we make?”
Ask instead:
> How many tokens did we use to complete one valuable business outcome?
That outcome might be a resolved support ticket, a qualified lead, a completed document review, or a generated report.
3. AGENTS CAN LOOP
An agent may retry when a tool fails. It may ask itself to improve an answer. It may repeat a search because the result does not meet its internal criteria.
Without controls, these loops can continue longer than expected.
Common causes include:
- No maximum number of retries
- No time limit for background tasks
- Poorly defined success criteria
- Tool failures that trigger repeated calls
- Multiple agents reviewing the same work
- A prompt that encourages unnecessary analysis
A workflow that normally costs a few cents can become a much larger expense when it loops.
4. MODEL CHOICES ARE NOT EQUAL
Frontier models are powerful. They are also more expensive.
Many businesses use a high-cost model for every task, including routine classification, summarization, extraction, and simple customer responses. That is similar to using a high-performance server for every small application.
A better approach is model routing:
- Use a smaller, lower-cost model for routine work
- Use a stronger model for complex reasoning
- Escalate only when the first model cannot complete the task
- Review whether the extra quality produces measurable business value
Recent reporting has shown that falling token prices do not automatically reduce total AI spending. Companies can still spend more because usage volume, context size, and agent complexity grow faster than unit prices fall.
5. HIDDEN CLOUD COSTS ADD UP
Tokens are only one part of the bill.
Your total AI cost may also include:
- Cloud compute
- Data storage
- Vector databases
- Data transfer
- Monitoring and observability
- Third-party APIs
- Security controls
- Backup and retention
- Human review
- Engineering maintenance
This is why AI cost control belongs in your broader cloud strategy, not only in an application team’s budget.
HOW TO MEASURE AI SPEND BEFORE IT SURPRISES YOU
You do not need an enterprise FinOps department to begin.
Start with a simple inventory. List every AI-enabled tool, workflow, department, and cloud account. Include embedded AI in software your team already uses. These costs are easy to miss because they may appear as credits, overages, or usage tiers rather than a separate AI invoice.
Then measure each important workflow.
Track:
- Model used
- Input tokens
- Output tokens
- Number of tool calls
- Number of retries
- Processing time
- Cost per completed task
- Business result
Run at least 20 to 50 realistic examples through each high-value workflow. Do not test only ideal prompts. Include long documents, incomplete data, failed integrations, and unusual requests.
Your dashboard does not need to be complicated. At minimum, show:
- Current month spend
- Daily burn rate
- Spend by team
- Spend by model
- Spend by workflow
- Projected month-end cost
- Cost per business outcome
If you cannot identify which workflow is driving the bill, you are not ready to scale it.
HOW TO CONTROL TOKEN COSTS
SET HARD LIMITS
Create limits at several levels:
- Maximum tokens per request
- Maximum cost per session
- Maximum retries per task
- Maximum runtime for an agent
- Daily or monthly budget per user
- Separate limits for development and production
A limit should stop or pause a workflow before it creates an unexpected charge. A warning after the money is already spent is not a control.
USE AN API GATEWAY OR PROXY
A gateway can inspect a request before it reaches the model. It can check the user, workflow, model, estimated cost, and remaining budget.
It can then:
- Approve the request
- Route it to a less expensive model
- Require human approval
- Reduce the context
- Reject the request
- Record the decision for later review
This moves governance upstream. You control cost before tokens are consumed.
REDUCE UNNECESSARY CONTEXT
Do not send your entire knowledge base with every request.
Use focused retrieval. Return only the documents or data needed for the task. Remove duplicate instructions. Summarize long conversations. Cache stable information when appropriate.
Better context design often improves both cost and response quality.
CONTROL AGENT LOOPS
Define what “done” means.
Set a maximum number of steps. Stop the workflow when the required information is found. Require human approval before expensive or irreversible actions.
You should also log every tool call. If an agent repeatedly calls the same system, you need to know why.
ROUTE TASKS TO THE RIGHT MODEL
Do not use the most expensive model by default.
Test whether a smaller model can handle:
- Classification
- Data extraction
- Simple summaries
- Draft responses
- Basic routing
- Format conversion
Reserve more capable models for work where accuracy, reasoning, or complexity justifies the cost.
WHAT SHOULD YOU BUDGET?
Vendor pricing changes frequently, so we recommend budgeting by usage rather than relying on a permanent price sheet.
For planning purposes:
- Simple AI tasks may cost cents per completed task
- Single-agent workflows may cost several times more
- Multi-agent workflows can reach dollars or more per complex task
- A poorly controlled workflow can exceed its expected cost by 10x or more
For a small business, a controlled AI pilot may require approximately $500 to $2,500 per month in cloud and model usage, depending on volume and workflow complexity. A production deployment with multiple departments may range from $2,500 to $15,000 or more per month.
Those are planning ranges, not promises. Your actual cost depends on users, model selection, context size, integrations, data volume, and automation frequency.
For consulting support, a focused AI cost and architecture assessment commonly takes one to two weeks. A reasonable planning range is $2,500 to $7,500, depending on the number of workflows and cloud environments involved. Implementation projects may take four to twelve weeks and range from $10,000 to $50,000 or more when production integrations, security, and governance are included.
We will tell you when your requirements fall outside a practical scope. Transparent expectations are more useful than an artificially low estimate.
A PRACTICAL SMB CHECKLIST
Use this checklist before expanding AI usage:
- Inventory every AI tool and workflow
- Identify which systems use agents or background processing
- Measure tokens and cost per business outcome
- Set per-user, per-workflow, and monthly limits
- Add retry and runtime controls
- Route simple work to lower-cost models
- Reduce repeated context
- Monitor cloud storage, compute, and data transfer
- Review AI spend monthly
- Tie continued investment to measurable results
HOW FIVE 9 CAN HELP
AI should solve a business problem. It should not become an uncontrolled experiment.
Our artificial intelligence consulting approach starts with the use case, data, and expected outcome. We help you determine where AI adds value and where a simpler solution is better.
Our IT consulting services can help you:
- Assess current AI and cloud usage
- Build a practical cost model
- Design usage monitoring
- Configure model and workflow guardrails
- Improve cloud architecture
- Establish governance policies
- Train your internal team
- Document the system for long-term support
We also support broader digital transformation consulting services, including process redesign, cloud modernization, system integration, and capability building.
Knowledge transfer is part of the work. We want your team to understand what we implemented, why it works, and how to manage it after the engagement.
START WITH AN HONEST CONVERSATION
You do not need a large AI program to get control of your costs.
Start with one or two important workflows. Measure actual usage. Set limits. Prove the business value. Then scale carefully.
If your AI or cloud bill is growing faster than expected, contact Five 9 for a no-pressure conversation. We will review the situation, explain the likely cost drivers, and tell you what we would address first.
Your AI budget should support growth( not surprise you after the invoice arrives.)