Guides
AI implementation cost for businesses
Prepared by: improveIT.pl
An audit, a first agent, RAG and the cost of running models.
AI implementation cost for a business: AI audit, a first agent or automation, RAG, API and maintenance costs, with prices excl. VAT and a pilot-first approach.
For a small or mid-sized company, AI is best started with an audit (USD 612.50–2,450 excl. VAT) or by automating one process (USD 962.50–2,625). An assistant that answers from company documents (RAG) costs USD 2,800–6,325. Monthly model and maintenance costs come on top and depend on the number of queries. Prices are converted from PLN at a fixed rate of 4 PLN = 1 USD and exclude VAT.
| Use case | Evaluate | Pilot boundary |
|---|---|---|
| Drafted reply | Correctness and human editing time | A person approves the result |
| Document knowledge | Retrieval, freshness and permissions | Agreed sources and question set |
| Classification | Categories and consequential errors | Test examples with expected outcomes |
| Tool-using agent | Permissions, retries and action records | Limited tools, spending limits and approvals |
Three common starting points
Start with one measurable process rather than a strategy for the whole company.
- AI audit: a list of processes to automate with estimated benefits and risks
- First agent or automation: one process, e.g. handling requests or invoices
- RAG: an assistant that answers from company documents
Running costs after implementation
AI also costs money while in use. Model fees grow with the number of queries and document length.
- Model API fees or tool licences
- Hosting for the knowledge base and integrations
- Monitoring answer quality and costs
- Keeping documents and permissions up to date
What the price depends on
Integrations and quality control cost the most, not the model itself.
- Number of systems to connect
- Volume and quality of documents
- Security requirements and where data is processed
- Need for human approval of decisions
- Number of users and queries
How to check whether it pays off
Before implementation, measure the current time and cost of the process. After the pilot, compare it with the cost of AI, including time spent checking results.
- Time to handle one request before and after
- Share of answers that need correction
- Monthly model cost
- Hours saved, converted to money
Risks worth assessing
AI works with company and customer data, so it needs the same rules as other systems.
- Personal data and GDPR
- Document permissions
- Wrong answers and how to detect them
- AI Act requirements for the use case
- Dependence on one model provider
Estimate AI around a workflow worth changing
Do not start with a model choice or a generic request for a company chatbot. Choose a repeated task with a defined input and result: drafting a reply, classifying a document or finding information in company knowledge. Describe how staff do it today, the exceptions they encounter and who checks the output. That lets you compare AI with a simpler automation or form improvement. Not every problem needs a language model. An estimate should include data assessment, representative tasks and success criteria. Otherwise it is easy to deliver an impressive demonstration that cannot be evaluated properly or safely integrated into daily work. Start with one use case rather than attempting to automate an entire department.
- A suitable candidate has a recognisable beginning and end, available data and an owner of the outcome.
- Define when the system may suggest an answer and when it may perform an approved action.
- Compare the task with a simple rule or lookup so AI has a justified role.
- Identify cases requiring refusal, uncertainty or escalation to a person.
Data and permissions often determine pilot effort
Company knowledge can be scattered across files, email, CRM records and outdated instructions. A model does not automatically reconcile conflicting information. Before the pilot, select sources, content owners, update methods and access rules. Where staff have different permissions, retrieval should enforce them before showing information. A prompt asking for confidentiality is not an access-control mechanism. The integration must technically restrict the permitted data and tools. Preparing sources is real project work and can involve more effort than invoking a model. Use a small, representative dataset approved for the initial test rather than copying everything because it is available.
- Separate information required for the task from accessible data that should not be processed.
- Define how removal or permission changes also affect indexes and auxiliary storage.
- For scanned documents, check extraction quality and OCR errors before judging model answers.
- Store keys in a suitable secrets mechanism and grant integration accounts only the permissions they need.
Model charges are only one part of AI cost
The budget includes data preparation, integration, testing, oversight and ongoing operation. Usage costs can include model calls, document processing, storage and retrieval. Provider prices and billing models change, so a token rate should not be presented as a permanent implementation price. Estimate using the selected model and representative tasks. Long context, retries and multi-step agents can cost more than a single response. Caching or a smaller model may help, but measure their effect while maintaining required quality. A cheap output that staff must spend time correcting can still make an expensive workflow. Set limits and alerts before opening the tool to a wider group.
- A practical model is task volume times average successful-task cost, plus infrastructure, care and human verification.
- Include failed calls, additional tools and retries rather than counting only the final answer.
- Separate one-off pilot preparation from monthly operation and subsequent development.
- Apply limits per account, project or workflow so one faulty run cannot consume the whole budget.
Evaluate quality without promising perfect accuracy
Prepare a task set with expected results or explicit assessment criteria. Include routine work, incomplete data and cases where the correct answer is uncertainty. Compare alternatives on the same set. Evaluate correctness, usefulness, latency and human review effort. Fluent language alone is not enough, particularly when a model fills gaps convincingly. If the solution uses documents, check whether the answer follows the correct source rather than merely sharing similar words. A new model or instruction can improve one class of tasks and worsen another. Re-run a control set after changes so the product does not gradually drift away from the behaviour accepted in the pilot.
- Identify critical errors that block release rather than averaging them with many easy examples.
- For drafted answers, measure corrections; for classification, compare against agreed categories.
- Include document text that impersonates a system instruction to examine trust boundaries.
- Record dataset, model and configuration versions so tests can be reproduced and compared.
An action-taking agent needs narrower operational boundaries
An assistant can propose an answer for human approval. An agent connected to tools can change records or trigger workflows, increasing permission, testing and failure-handling requirements. Start with read-only access or approval-gated actions. Validate parameters in application code rather than relying only on prompts. Limit steps, duration and cost. Actions need an audit record and retry recognition to avoid performing the same operation twice. Expand autonomy after examining real exceptions. Full administrator access is not required to test one useful automation. A deliberately narrow capability is easier to verify, easier to support and usually more informative than a broad demonstration operating with unrestricted permissions.
- Separate read access from permission to write, delete or send messages.
- Require approval for materially consequential actions under the company’s agreed policy.
- Test limits, invalid tool output and unavailable external systems.
- Provide a stop mechanism and manual handling so automation failure does not block basic work.
A pilot should end with a decision, not just a demonstration
Decide beforehand how the result will be used. A pilot can support further investment, expose data-preparation work or show that a simpler approach is preferable. Each outcome is useful when criteria are clear. The improveIT starting scope covers one agreed use case, dataset and evaluation method; it does not automatically include every department. Production needs ownership, cost monitoring, source updates and issue handling. Measure savings across the complete process, including preparation, review and corrections. Do not claim a return on investment without observing customer conditions. A useful plan also states when continuing would not be justified, rather than making expansion the assumed result of every pilot.
- Compare task time before and after the pilot using similar cases and including verification.
- Select actual users and a feedback process instead of relying only on a presentation.
- Record production-entry conditions and the tasks that remain manual.
- To start a discussion, send the workflow, data types, existing tools and main difficulty; we can define a pilot that can be evaluated.
Choose RAG only when retrieval addresses the problem
Retrieval-augmented generation combines source retrieval with model-generated answers. It can help with questions over a larger document collection, but does not replace organised information or access control. For a specific order status, a direct system integration with a constrained answer format may be more appropriate. Document use introduces decisions about chunks, metadata, filtering and retrieval evaluation. Compare alternatives on examples before investing in a large pipeline. Make index maintenance, source updates and quality controls visible alongside model cost. Architecture should follow the task rather than a desire to include a fashionable acronym. Understanding which component failed also makes later support more efficient.
- For numbers and statuses, identify the authoritative source and a current lookup mechanism.
- For documents, preserve metadata and versions so answers do not rely on withdrawn instructions.
- Evaluate retrieving the right passage separately from answer correctness because failures have different causes.
- The related RAG and semantic-search guide covers source preparation and retrieval in more detail.
Primary sources
OpenAI: current API pricing · OpenAI: API deployment checklistFAQ
Can a small business implement AI?
Yes. It usually starts with one process, e.g. classifying requests or drafting documents, for a few hundred to a few thousand dollars excl. VAT.
How much does using AI cost per month?
It depends on the number of queries and the model. With modest traffic it is usually from a few to a few dozen dollars a month for the API, plus integration maintenance.
Will company data reach the model provider?
That depends on the service and configuration. You can choose models that do not train on customer data, processing in the EU, or a model running on your own server.
Where should we start?
By choosing one process that takes a lot of time and has clear correctness criteria. An AI audit helps identify it.
Do I need to provide all company data immediately?
No. The first pilot uses a limited dataset relevant to the task and agreed access rules. Define which information may be processed and how it will be removed after testing. Data unnecessary for evaluation should not be sent merely because it can be retrieved. If a sample is anonymised, check whether it still represents real cases. Full integration is a subsequent decision, not a default requirement for starting a discussion.
How do I know whether a cheaper model lowers total cost?
Compare models on the same tasks at the required quality level. Include calls, retries, tools and human verification time. A lower response price may not save money when answers need more corrections or repeated runs. The pilot should measure a successfully completed task, not only a token rate. Results apply to a particular process and dataset, so they should not automatically be generalised to every company use case.
What happens if the AI system produces an incorrect answer?
Design this case before launch. An output can require approval, follow another workflow or be escalated to a person. Data-changing actions need restricted permissions, validation and action records. Reported errors should improve the control test set. A warning that mistakes are possible is not sufficient by itself: the workflow should limit their impact and allow the task to continue safely.