An AI agent costs more than the language-model requests it makes. The budget also covers understanding the process, connecting information, restricting actions and demonstrating that the application is useful. Prices become comparable only when the task, data and responsibilities are clear. This guide offers a budgeting method rather than OXIAE rates or a claimed market average.
Start with the smallest application that addresses a meaningful problem. Ask proposals to separate initial development, recurring services and work that remains with your employees. A lower implementation price can come with more internal review or fewer recovery options. Those consequences belong in the same business decision as the supplier’s invoice.
Scope is the first cost driver
An agent summarising enquiries has a different scope from one sending quotations independently. Summarisation needs input handling, an output format and assessment. Quotations also involve pricing rules, customer identity, validity periods, approval and transaction records. Break the assignment into individual operations and specify which ones are actually included in the first release.
Define the supported variation as well. One department, language and enquiry type is easier to delimit than several countries with different contractual arrangements. Expansion can still be sensible, but additional cases should be visible in the budget. A list of excluded situations prevents “suitable for customer service” from silently becoming responsibility for every conceivable customer problem.
Preparing information is a separate activity
Existing documents are not automatically usable knowledge sources. Duplicate versions, expired terms and unclear ownership need investigation. If nobody knows which information applies, a stronger model will not resolve that organisational problem. Budget for an inventory, source selection, any necessary conversion and responsibility for future updates as identifiable activities rather than invisible preparation.
Access boundaries matter too. Public product information differs from files restricted to individual customers. File count alone is a poor measure of complexity. Ten confidential documents with different permissions can need more design than hundreds of public instructions. Ask how access changes, deletion and source updates are handled, and which of those behaviours the proposal actually includes.
Integrations and actions define the technical reach
A demonstration may use a sample file; production often needs reliable connections to existing systems. Establish whether the required API exists, which subscription permits access, whether a test environment is available and who can answer functional questions. The connection must also cope with missing records, expired credentials and temporary outages rather than only the successful path.
Budget reading and writing separately. Retrieving a customer record differs from creating an appointment. Writes need further decisions about validation, repeated requests, changes made since a suggestion and recovery from an uncertain result. Ask the supplier to list supported operations for each connection. A proposal containing only system logos or product names does not establish a comparable scope.
Include outcome evaluation in the budget. This guide helps define that work: Testing an AI agent: turn fluent answers into useful evidence.
Usage needs scenarios rather than false precision
Model normal, quiet and busy periods. Count conversations alongside typical steps, retrieval operations, retries and the amount of information processed. A short conversation involving extensive internal searching can consume more resources than a longer exchange on a straightforward subject. Ask which usage measurements will actually be available so the assumptions can be checked later.
Use your own assumptions for an initial estimate. A fictional team expects 600 cases per month and investigates whether 400 fit the first task boundary. The other 200 continue through the existing process. Estimate model use, infrastructure and external services for the supported 400 cases separately. Insert current supplier prices once the proposed architecture and relevant billing categories are known.
Include human review in the business case
Do not count every automatically prepared reply as entirely saved handling time. An employee may still need to inspect sources, assess uncertainty and approve the action. Measure current handling time against the complete new handling time, including recovery. Released capacity only becomes a cash saving when the organisation actually avoids corresponding expenditure or changes its staffing costs.
For illustration, suppose 400 cases each save four minutes of entry but require one and a half minutes of review. That releases 1,000 minutes, approximately 16.7 hours a month. Then deduct expected exception handling and source maintenance. Applying an internal hourly rate produces a capacity value rather than guaranteed revenue or savings. A bounded pilot should test these assumptions with actual work.
Evaluation and handover are deliverables
Ask which examples will be used to assess the agent, who defines acceptable outcomes and how disagreements are resolved. A demonstration with five successful prompts is not a complete acceptance exercise. Budget for ordinary cases, missing information, conflicts and requests outside the agent’s authority. Employees also need time to review the resulting evidence with care.
OpenAI’s evaluation documentation supports application-specific tests. In procurement terms, a test collection and a repeatable comparison should be explicit deliverables. Specify whether correcting findings within the agreed scope is included. Ask for operating instructions, an access inventory and a handover exercise in which somebody other than the original builder stops and resumes the application successfully.
Separate upkeep from changed behaviour
Some recurring work keeps the existing application available: hosting, investigating alerts, renewing access and installing updates. Other work changes behaviour, such as supporting another department or adding a new permission. Keep these categories visible. Otherwise a maintenance subscription says little about which requests are included and which require another agreement or budget.
Consider supplier changes and exit as well. Can the organisation retain its source material, configuration and evaluation cases? Who responds when a model or external API changes? What work is required when the engagement ends? A workable handover plan reduces dependency. Include its cost in the comparison even when you expect to keep the same supplier for many years.
Compare proposals using one scope sheet
Create a table listing supported tasks, channels, languages, sources, connections, operations and exceptions. Assign an owner and a verifiable acceptance condition to each part. Put initial costs, fixed monthly costs, variable usage and expected internal hours alongside it. Ask each provider to identify differences explicitly so that a cheaper offer does not silently describe a smaller outcome.
Finally, agree on a decision point. Discovery or a pilot may show that a fixed workflow is sufficient, or that data quality needs attention first. A phased engagement makes that result useful: the investigation delivers a supported next decision. Specify which evidence the decision requires and which subsequent investment still needs its own assessment before work begins.
Turn the explanation into a project decision
Record your assumptions, unknowns and acceptance checks in the matching tool. Use the service guides to discuss delivery and scope.
Open the preparation tool ↗Questions about this topic
Can a project have a fixed price?
A fixed price can suit a bounded assignment with understood sources and connections. Where uncertainty is substantial, a paid discovery stage may be appropriate first. The commercial arrangement and amounts for your project belong in an agreed proposal; this guide is not an OXIAE price list.
Why are model charges only part of the budget?
The application also retrieves information, checks permissions, processes records and handles failures. Preparation, evaluation and ongoing support require work. A cheap model does not make an unclear business process or a complex integration straightforward by itself.
How do I assess the value of a pilot?
Measure supported cases, complete handling time, review effort and errors against the current process. Compare measured capacity with initial and recurring costs. Track improved quality or availability too, while keeping those benefits separate when they cannot reliably be expressed as money.
Continue with a related question
AI agent or chatbot: which approach fits your process?
Compare chatbots, fixed workflows and AI agents by task, permissions, reliability and operating effort, with a practical decision example.
AI agents & chatbotsTesting an AI agent: turn fluent answers into useful evidence
Build a practical AI-agent test set covering sources, actions, exceptions, handover and repeatable acceptance criteria.
AI agents & chatbotsOperating an AI agent: monitoring, changes and recovery
Plan AI-agent operation with useful alerts, usage limits, source ownership, regression checks, pausing and a workable recovery route.
n8n & process automationWhat does workflow automation cost? Budget beyond the initial build
Budget for workflow automation using discovery, integrations, testing, maintenance and review. Includes a transparent fictional calculation.
Sources and technical background
- OpenAI: Evaluation best practices ↗
Task-specific evaluation, representative cases, human calibration and continued regression testing.