The real cost of an AI agent: budgeting beyond the subscription

In 60 seconds: An agent’s monthly cost is neither its subscription nor its model bill. Use this formula: model + tools + channel + infrastructure + human supervision + maintenance + cost of failure. Measure every component per completed task, separate variable, fixed, and integration costs, and project a volume range. Text, images, audio, history, tool results, and retries change consumption. For a pilot, set limits, log usage, choose a fit-for-purpose model for each task, cache repeated context where appropriate, and compare results with a measured manual baseline.

A demo may answer one question for a few cents and still become an expensive operation. The model might make five attempts, call three services, read a long document, send a WhatsApp message, and finish in a human review queue. If the result is wrong, someone must correct it or support an affected customer. Asking “what does a token cost?” helps with one invoice line, but it does not approve the full budget.

A formula you can reuse

Calculate every component for the same period and currency:

Total cost = model + tools + channel + infrastructure + supervision + maintenance + failure

Then divide it by the outcome that matters:

Cost per correct outcome = total cost / completed and accepted tasks

“Task” needs an operational definition. It could be a confirmed appointment, a classified and reviewed invoice, or an order recorded without duplication. A generated answer is not always a resolved task.

Component	What to record	Examples
Model	input, output, cache, modality, model, and retries	text, images, audio, reasoning
Tools	calls and billable unit	search, OCR, maps, email, signing, database
Channel	messages, sessions, numbers, users, or licenses	WhatsApp, Teams, voice, SMS
Infrastructure	time, storage, and traffic	functions, queues, vectors, files, backups
Supervision	minutes and loaded labor cost	review, approval, exception, support
Maintenance	recurring hours	evaluations, prompt, connector, and permission updates
Failure	frequency by severity and recovery cost	rework, refund, noncompliance, lost sale

A spreadsheet can start with these seven rows. Precision comes from feeding each row with pilot telemetry, not from guessing more decimal places.

Separate variable, fixed, and integration costs

Variable costs grow with usage: tokens, searches, audio minutes, messages, storage, retries, and review minutes. Express them per task and per scenario: simple, typical, and exception.

Fixed costs remain when the agent does little work: minimum licenses, observability, environments, phone numbers, on-call coverage, security, and reserved operations time. Dividing them across a small number of tasks can make a low-volume pilot look expensive even when the technology works.

Integration cost is the initial and ongoing work required to connect systems: authentication, permissions, field mapping, tests, deployment, and migrations when an API changes. Amortize it over an explicit period. Do not hide it inside monthly consumption or call it a one-off payment when the surrounding systems will continue to change.

What increases model consumption

For text, input can include the user’s words, instructions, history, tool definitions, retrieved documents, and results returned by those tools. Output includes the response and, depending on the product, other billable tokens. Read actual usage fields instead of estimating from character counts.

Long context costs in two ways: it contains more input units and may be repeated on every turn. Sending the full manual twenty times is not a memory strategy. Retrieve only the relevant passages, summarize stable state, and measure how much context improves the answer.

Images become billable units according to the model, image count, size, or detail level. Audio may be metered by tokens, seconds, or minutes depending on the service. One voice note can also trigger transcription, a model response, speech synthesis, and storage. Do not collapse “multimodal” into one average rate.

Frequency multiplies everything above. An agent that checks a mailbox every minute, retries without a cap, or rereads the same file may spend more on activity without outcomes than on useful work. Record the trigger, attempt, repetition cause, and final result.

The model rate is only one line

Sources and pricing verified on September 16, 2026. The official OpenAI pricing table separates input, cached input, cache writes, and output, and distinguishes context and processing mode. The official Anthropic pricing table also publishes separate rates for input, output, and cache operations, plus charges for hosted tools. These pages are the current source of truth: copy their rates on the day you estimate and store model, region, and service mode beside each value.

For a concrete high-capability example, the guide to GPT-6 Astra for business connects the published API rate with three pilot workflows, computer-use risks, and the controls needed before write access.

This structure explains why two tasks with the same visible text may cost different amounts. One might include a long history, five tool schemas, and two searches; the other might use one short instruction. Compare the full usage record, not the model’s marketing name.

The channel has its own meter and rules

WhatsApp should not be budgeted as a free pipe. According to the official WhatsApp Business Platform pricing page, verified on September 16, 2026, Meta charges per delivered message and the rate depends on market and category. The page says service messages and utility messages sent in response to users are free, and a user’s message opens a 24-hour customer-service window. Use the rate card for the pilot’s country and category, then add any platform provider fee separately.

In Teams, cost may sit in existing licenses, Power Automate, a premium connector, or the service running the agent. The US official Power Automate pricing page, checked on September 16, 2026, listed Premium at USD 15 per user/month, Process at USD 150 per bot/month, and Hosted Process at USD 215 per bot/month, paid yearly. Microsoft notes that actual pricing may vary by market, contract, and tax. Validate the tenant and connectors: owning one license does not mean every integration is included.

Manual interface, API, or visual automation

There is no universal winner. They are different operating models.

Option	Fits when	Commonly hidden cost
Manual interface	a person explores, drafts, or decides at low volume	labor, copy and paste, plan limits, and weak traceability
API	the process needs volume, control, tests, and telemetry	development, infrastructure, security, on-call work, and maintenance
Visual automation	triggers and connectors are clear and Operations can maintain them	premium licenses, runs, hard-to-debug branches, and connector dependency

A subscription may be economical for twenty assisted tasks per month. An API may win with thousands of repeatable events and a team that already operates services. A visual tool can shorten a Microsoft 365 integration. For a deeper decision framework, read Copilot and Power Automate or custom development.

How to budget a pilot without false precision

1. Choose one bounded task. Define its trigger, correct output, exception, and owner. If the priority is still unclear, use the framework in What should a business automate first?.
2. Measure the baseline. Across a representative sample, record volume, human minutes, waiting, corrections, incidents, and outcomes. Include the happy path and difficult cases.
3. Create three scenarios. Project low, expected, and high volume. For each, estimate the mix of simple, typical, and exception cases.
4. Set technical limits. Cap context, output, tool calls, retries, runtime, and daily spend. Stop or route to review when a limit is exceeded.
5. Route by task. Use a lower-cost model for extraction or classification when it meets the measured threshold, and reserve more capable models for cases that need them. Evaluation makes the decision, not the label.
6. Use useful repetition. Cache stable content when policy, latency, and the usage pattern support it. Batch non-urgent work if the provider offers a suitable batch service.
7. Instrument the full cost. For every task, store the model, tokens or equivalent unit, cache, tools, channel, infrastructure, human time, outcome, and failure reason. Alert on spend and abnormal behavior.
8. Review weekly. Look for unnecessary context, tools that add no value, retries, volume drift, integration failures, and tasks that should remain manual.

Project the month with ranges:

Monthly budget = fixed costs + amortized integration + Σ(volume by type × variable cost by type) + incident reserve

The reserve is not a universal percentage. Base it on observed failures and recovery times, then update it when the process, provider, or volume changes.

Measure savings against a real baseline

Savings are not “hours AI could save.” They are the difference between comparable periods:

Net savings = baseline process cost - total assisted-process cost

Keep the task definition, required quality, and period consistent. If the agent processes more cases but creates more corrections, a gross time metric can hide the problem. Track cost per correct outcome, cycle time, backlog, exception rate, and failure severity.

Also record benefits that do not reduce labor directly, such as shorter waits, better traceability, or after-hours capacity. Name them separately and do not turn them into currency without evidence.

A useful pilot does not promise a generic ROI. It produces enough evidence to scale, redesign, or stop the automation. Kiia can help map that baseline, instrument the pilot, and choose the simplest combination the team can operate.