What you pay for when you build an AI system, the factors that move the cost most, how common project types compare, a build or buy decision tree, what model usage costs at current list prices and how to get a reliable estimate.
“How much does an AI project cost?” is a fair question with an honest answer most buyers find unhelpful: it depends on what the system has to do. A chatbot that answers questions from twenty help articles and an agent that processes insurance claims across five systems are both called AI development, yet the effort behind them differs by an order of magnitude.
This guide explains what you pay for, which factors move the cost most, how common project types compare, and how to get an estimate you can trust. It covers the cost to build. The monthly cost of running models in production has its own guide on keeping LLM costs predictable, and chatbot budgets are broken down in our chatbot development cost guide. If you are still choosing which process to automate first, start with our roadmap for adopting AI in a mid-sized business.
Key takeaways
- Every AI system has three budgets: build, run and improve. Compare options on the total over two or three years, not on the launch price.
- Scope, data readiness and integrations move the build cost more than the choice of model. Data preparation is the most underestimated item.
- Most business projects start with a hosted model plus retrieval from your own data; custom or fine-tuned models are rarely the right first step.
- Model prices per token vary widely: at Anthropic's list prices on 13 Nov 2026, the same workload cost $5.50 on the smallest current model and $440 on the largest.
- A short paid discovery or proof of concept turns unknowns into facts and is the cheapest way to make the main estimate reliable.
What you are actually paying for
AI projects carry three budgets, and teams often plan only the first:
- Build: discovery, data preparation, design, development, integration with your systems, testing and launch.
- Run: model usage, hosting, vector search, storage, monitoring tools and licences, paid every month.
- Improve: evaluating quality, updating knowledge and prompts, fixing edge cases, and adapting when models, policies or products change.
A cheap build with high running costs can cost more over two years than a careful build that keeps usage under control. A system that nobody maintains loses accuracy as your products and policies change, and users stop trusting it. Budget for all three from the start.
Prices for the underlying technology have fallen fast. Stanford HAI’s AI Index 2025 reports that the cost of running a system that performs at the level of GPT-3.5 dropped more than 280-fold between November 2022 and October 2024, and that hardware costs declined by 30% a year (checked 13 Nov 2026). Cheaper models lower the run budget; they do little for the build budget, which is mostly people’s time spent on data, integration and testing.
The main cost drivers
1. Scope and type of solution
The biggest driver is what the system must do. Answering questions from a fixed set of documents is simpler than taking actions in other systems. One language, one channel and one team is simpler than many. Each extra integration, user role and edge case adds design, development and testing. Writing down what the system must never do is as important as what it should do, because each “never” needs a control and a test.
2. Data readiness
An AI system is only as good as the data and documents behind it. If your knowledge is current, consistent and in one place, preparation is quick. If it is spread across shared drives, old PDFs, email threads and people’s heads, a large part of the project is collecting, cleaning and structuring it. For assistants that answer from your documents, retrieval-augmented generation is the usual method, and its quality depends almost entirely on that preparation.
3. Model approach
There are three broad approaches, with very different cost profiles:
- Hosted model APIs from the major providers: little upfront work and no infrastructure to manage. You pay per use, so cost grows with volume.
- Open-weight models you host yourself: more set-up and operations work, and you pay for servers or GPUs, but you control where data goes and costs are more predictable at high, steady volume.
- Fine-tuned or custom models: need labelled data, experiments and evaluation. They pay off for narrow, high-volume tasks where general models fall short, and rarely for a first project.
Most business projects today start with a hosted model and move only when volume, privacy or quality demands it. Our comparison of which LLM to choose sets out vendor prices, context sizes and data residency options.
4. Integration with your systems
An assistant that reads a few documents is self-contained. One that checks order status, updates a CRM or creates tickets needs secure connections to those systems, error handling, permissions and logs of every action. Older systems without APIs may need custom connectors or RPA. Integrations are where estimates most often go wrong, because the quality of the other system is outside the developer’s control.
5. Accuracy, safety and compliance
How bad is a wrong answer? For internal drafting, a reviewer catches mistakes. For customer-facing advice, regulated industries or automated decisions, you need stronger guardrails: evaluation sets, approval steps, audit logs, content filters, data residency and sometimes formal risk assessments. If you serve users in the EU, our guide to the EU AI Act for companies outside Europe explains which uses carry heavier obligations. Each control is justified, and each adds work.
6. User experience
A widget on an existing website is lighter than a new web or mobile app with accounts, dashboards and admin tools for your team to manage content and review conversations. Agents that act on someone’s behalf also need screens for approvals and for undoing mistakes.
7. Team and engagement model
Rates vary by region and seniority, and so does speed. A small senior team that has built similar systems is often cheaper overall than a larger, less experienced one. Fixed scope suits well-defined work; time and materials or a dedicated team suits work where the scope will be learned along the way, which is common with AI. If you want to add AI skills to your own team instead, you can hire AI developers by the month.
Want a second opinion on your plans?
Tell us what you are working on. A specialist will look at it and reply with practical next steps.
Get in touchTypical AI project types compared
The table compares common project types by what they involve and their relative effort. It shows effort rather than prices, because a price without your scope would be a guess, and two honest quotes for the same idea can differ widely on team size and rates.
| Project type | What it involves | Main cost drivers | Relative effort |
|---|---|---|---|
| Proof of concept | One use case tested on real data with a simple interface | Access to sample data, a test set of real cases | Low |
| Knowledge assistant or chatbot | Retrieval from your documents, chat interface, handover to people, content admin | State of the content, permissions, channels | Low to medium |
| Document processing | Extracting data from invoices, forms or contracts, validating it and sending it to your systems | Document variety, validation rules, target system | Medium |
| Generative AI feature in an existing product | Drafting, summarising or transforming content inside your app, with controls | Fit with the existing code base, quality checks | Medium |
| AI agent with tool access | Plans steps and acts in several systems, with permissions, approvals and logging | Number and quality of integrations, approval design | Medium to high |
| Voice agent | Phone calls with speech recognition and synthesis, transfer to people | Telephony, latency, languages and accents, consent | Medium to high |
| Predictive model | Forecasting or scoring from historical data, with a dashboard or integration | Volume and quality of history | Medium to high |
| Fine-tuned or custom model | Training data, experiments, evaluation, hosting and retraining | Labelled data, evaluation, ongoing retraining | High |
Source: AcmaSoft delivery practice. Relative effort for a small, experienced team on a clear scope; not a quote. Your project may sit outside these ranges in either direction.
To turn effort into money, multiply the team and the weeks by the partner’s rates. Our explainer on AI agents shows why agents sit higher on this table than assistants that only answer questions: every action needs a permission, a log and a way to undo it.
Build, buy or a mix
Before pricing a build, check whether you need one. If the need is common, such as meeting notes, writing help or a standard help desk bot, a SaaS product may cover it for a subscription. Custom work makes sense when the use case is specific to your business, needs deep integration, handles data you must control, or runs at a volume where per-seat or per-conversation pricing becomes expensive. Many projects combine both: hosted models and existing platforms for the commodity parts, custom work for your data, workflows and interface.
Build, buy or combine for this AI use case?
Question 1: Does a SaaS product already do this task well for businesses like yours?
Question 2: Can it connect to the systems and data it needs, under your data rules?
- YesBuy, and pilot it on your own cases
- NoGo to question 3
Question 3: Is the use case specific to your process, data or product?
- YesGo to question 4
- NoRe-check the market, or combine a SaaS tool with integration work
Question 4: Do you have the data and system access the use case needs?
- YesGo to question 5
- NoFix data and access first; build later
Question 5: Will usage be high enough that per-seat or per-use pricing becomes expensive, or is AI part of your own product?
- YesBuild, starting with a proof of concept
- NoCombine: hosted models and platforms, custom integration and interface
Our article on building, buying or partnering for AI goes through this decision with a three-year cost comparison.
Run costs in brief
Hosted models are priced per token, a unit of text roughly the size of a short word, with input (what you send) and output (what the model writes) priced separately. A simple estimate is: monthly usage cost = requests per month x tokens per request x the price per token.
Model choice changes that figure dramatically. The chart applies Anthropic’s published list prices to one workload, to show the spread within a single vendor’s current range.
How the figures are calculated: 10,000 requests of 3,000 input tokens is 30 million input tokens, and 500 output tokens each is 5 million output tokens. For Claude Sonnet 5.5, 30 x $2 + 5 x $10 = $110. The same pricing page lists a 50 percent discount for requests sent through the Batch API, which suits work that does not need an instant answer. Prices change often, so check the current page before you budget.
Usage is rarely the whole run cost. Add hosting for the application, vector search, logging and monitoring, and the hours spent reviewing quality. Caching, routing easy requests to smaller models, and sending less context are the main levers; the run cost guide works through a full monthly bill.
Costs after launch that buyers forget
- Evaluation: a test set of real cases, scored before every release and after any model or prompt change.
- Content upkeep: products, prices and policies change, so the knowledge behind the system must change too.
- Model changes: providers retire older model versions, which means re-testing and sometimes adjusting prompts.
- Review time: someone needs to read samples of conversations or outputs each week and turn failures into fixes.
- Security and compliance: access reviews, penetration tests, data retention and audits recur, and in regulated sectors they are not optional.
How to keep AI costs under control
Cost controls that work from the first sprint
- Start with one use case, one team and one measure of success; expand once it works
- Fix the data first: current, consistent knowledge lowers both build effort and wrong answers
- Use the smallest model that passes your test set, and route only hard requests to larger ones
- Send the few relevant passages, not whole documents, so each request carries less context
- Cache answers that do not change per user
- Use plain workflow automation where fixed rules are enough; it is cheaper to run than a model
- Set usage limits and spending alerts so a bug or misuse cannot run up a surprise bill
- Review cost per task each month alongside quality, not separately
Rules-based steps are often better handled by workflow automation than by a model, and our guide to automating business processes explains where each fits.
How to get a reliable estimate
Step 1: Write down the use case
One paragraph: who uses it, what it does, what it must never do and how you will measure success.
Step 2: Collect samples
Real documents, tickets, forms or records the system will work with, including the messy ones.
Step 3: List the systems
Every system it must read from or write to, with who owns each one and how it can be accessed.
Step 4: Agree the quality bar
What counts as a correct answer or action, and who reviews outputs before and after launch.
Step 5: Run a short discovery or proof of concept
Test the riskiest part on real data before committing to the full build; it turns unknowns into facts.
How we price AI projects
We do not publish a price list, for the reason given at the top: without your scope, any figure would be a guess. We price in two stages. A short discovery or proof of concept is fixed in scope and price, and ends with a working test on your data, a measured result against your test set, and a written estimate for the full build with its assumptions listed. The full build is then priced as a fixed scope or a monthly team, depending on how well the scope is known.
Our AI consultants run the discovery, our generative AI and LLM development team builds assistants and AI features, and our AI agent development team builds systems that act in your tools. We have shipped AI in our own products too: AI agents and a copilot in Acmafone CRM, and AI voice calls in Hindi and regional languages in attpro CollectOS.
Want an estimate you can rely on?
- Start with our free assessment of your AI readiness. You get a written view of your use cases, the data and integration work each one needs, and a recommended first project with what it would take to prove it.
Get new guides by email
Articles and guides from our team, and company news. Unsubscribe at any time.
About AcmaSoft
AcmaSoft designs and builds websites, mobile apps and custom software, and runs SEO and digital marketing for its clients.
- Years of experience
- 17+
- Clients served
- 650+
- Projects delivered
- 2,000+
- Products shipped
- 300+


