AI agent costs become comparable once you’ve defined the work the agent should do. Drafting an answer from a document needs a different scope from changing an order in your administration system. This guide separates development, usage and maintenance. Our guide to AI automation helps you choose a suitable task first; the explanation of AI agents and chatbots covers their different ways of working.
What makes up the cost of an AI agent?
The total includes building and introducing the application, paying for the services it uses and keeping it useful over time. Comparing model bills alone gives an incomplete picture. You might spend little on tokens while staff spend a great deal of time correcting wrongly processed requests.
Start with the outcome you need delivered. Will a member of staff review a draft, or should the system save a change independently? The latter needs checks for permissions, duplicate actions and failures. That difference belongs in the quote, even if both suppliers call their proposal an agent.
| Cost item | What should the budget describe? | What should you check? |
|---|---|---|
| Preparation | Process, data, permissions and test material | Which assumptions have been investigated? |
| Development and integrations | Screens, model calls, source retrieval and actions | Which systems and operations are included? |
| Introduction | Testing, instruction and the move to daily use | Who evaluates the results? |
| Usage | Models, storage, hosting and external services | Which volumes and rates are assumed? |
| Maintenance | Incidents, updates, checks and changes | What’s covered by the contract? |
| Internal work | Source upkeep, approval and corrections | Whose time is needed after delivery? |
Which choices change the development price?
An agent answering from your documents must retrieve suitable sources and respect their access permissions. The work includes making that information available and keeping it current. A folder of current, clearly named documents needs different preparation from scans, duplicates and files with unclear access rules. Our guide to AI agents using your own data and RAG explains the technical choices.
Discuss integrations separately. Can the application only read data, or can it change a customer record? Is a test environment available? How will you find out who approved a change? Ask the developer to describe a difficult case, such as an action that succeeds but whose confirmation never arrives. Their answer should clarify whether the price covers a working demonstration or also its behaviour when things go wrong.
The channel matters too. Voice adds telephony and speech processing, for example, while a chat interface in an existing portal may require work on access permissions. That doesn’t produce a fixed ranking from cheapest to most expensive. Compare the actual requirements.
A useful quote also identifies unknowns. If nobody has checked whether the accounting package permits the intended change, include an investigation step. A fixed price based on an untested assumption can still lead to a scope discussion later.
How do you calculate model usage costs?
For the Claude API, Anthropic’s pricing page lists Haiku 4.5 at a base rate of $1 per million input tokens and $5 per million output tokens. Checked on 24 September 2026. These are US dollars; check the current rates and billing conditions for your chosen platform.
The following is a hypothetical usage scenario, not a business average or a TopDevs quote. We assume 1,000 tasks per month and exactly two model calls per task. Each call contains 2,000 input tokens in total and produces 500 output tokens. Tokens are the units of text processed by the model; they aren’t the same as words.
| Calculation | Quantity | Cost at the base rate |
|---|---|---|
| Input: 1,000 × 2 × 2,000 | 4 million tokens | $4 |
| Output: 1,000 × 2 × 500 | 1 million tokens | $5 |
| Total for these model calls | 2,000 calls | $9 |
That’s only the token bill under these assumptions. It excludes hosting, development, maintenance, taxes and any paid tools. The calculation uses no caching, batch discounts or other rate adjustments. It also isn’t a recommendation that this model will perform well enough for your task.
A real measurement must include all input: instructions, attached documents, conversation history and any tool results. A retry counts too. Record the number of tasks completed usefully alongside the model bill. A cheap call that keeps requiring review or correction can prove expensive for the business.
How do you compare a subscription, a platform and custom development?
Have each option perform the same work. Use the same inputs, permissions and review requirements. Otherwise, you’re comparing the subscription for a standalone assistant with the full development of a business process.
An existing product may suit a task it already covers. Include setup, user licences, usage limits and any required extensions. With a workflow-building platform, also establish who maintains the flow and how failures become visible. Our comparison of no-code and custom automation explores that choice further.
Custom development puts the agreed functions within the development scope. It doesn’t make external services free. You may still pay for models, hosting, an email service or access to a connected package. With TopDevs, the custom code is and remains yours. Also agree which documentation, settings and access are needed to transfer maintenance.
What belongs in the first-year budget?
Add the one-off investment to recurring costs over your chosen period. Account for the start date: twelve months of use differs from a calendar year in which the application only starts in November. Use the same currency throughout and treat VAT consistently.
Use this structure to fill in your own figures:
- One-off work: investigation, development, data preparation, testing, instruction and introduction.
- Recurring costs: model usage, hosting, storage, other services and agreed maintenance.
- Internal work: source upkeep, review, incidents and training new users.
- Uncertain items: extra source formats, new integrations or changing usage volumes.
Give each item an owner and an assumption. A supplier might budget for hosting while your team takes responsibility for cleaning the data. Don’t leave that internal work out of the total just because it doesn’t generate an external invoice. Our custom software budgeting guide explores the distinction between investment and running costs.
Then calculate a higher-usage scenario. Some costs stay flat until a subscription threshold; others grow with each processing operation. Ask what happens at a limit: do tasks stop, join a queue or keep adding to the bill? Agree on alerts and responsibility before the application goes live.
When does an agent pay for itself?
That requires evidence from your own process. Measure time per task before starting and include review, corrections and exceptions during the trial. Compare like-for-like work. A test containing only short, routine questions won’t predict what happens with difficult cases.
Time saved doesn’t automatically become money you no longer spend. Your team might use it to clear a backlog, answer sooner or increase capacity. Describe the benefit you expect and how you’ll recognise it. Calculating a payback period using an invented hourly value and unproven time savings doesn’t make the budget more reliable.
Also consider whether the task would be simpler without a model. A fixed rule for a known order number needn’t become an open-ended search. Make the process smaller first, then choose the model or platform to run it.
How do you actually get started?
Describe one task from input to intended result. Include representative examples, a failure and a situation where the agent must do nothing. Note expected volume, who maintains the data and who may approve an action. That gives a supplier enough context for a focused estimate with the remaining unknowns identified separately.
If you want to build a custom AI agent, we’ll discuss that scope first. A useful quote ties its price to the agreed work and explains the ongoing commitments after delivery. Book a call to look at your process and budget together.
