Key takeaways
- One AI feature in an existing app takes about 150 to 400 hours; a production AI product with retrieval, integrations and evaluation takes about 500 to 1,400 hours.
- At Eastern Europe rates that is roughly $5,250 to $32,000 for a feature and $17,500 to $112,000 for a product; US onshore rates about double it.
- Data preparation, evaluation and guardrails are what separate a demo from a production system, and they often take a third of the hours.
- Model choice drives running cost: the same monthly workload can cost $32.50 on a small model or $1,300 on a top model at November 2026 list prices.
- Most teams should start with hosted model APIs; fine-tuning or training a custom model only pays off with stable tasks, clean labeled data and high volume.
AI development costs about $5,000 to $64,000 for a single AI feature added to an existing app, and about $17,500 to $224,000 for a production AI product with retrieval over your data, integrations, evaluation and monitoring, depending on where the team is based. High-volume document or media pipelines and custom models cost more, and every AI system adds a monthly bill for model usage on top.
We are engineers who build these systems: in-product assistants, transcription pipelines, AI content tools and AI wired into CRM and ERP workflows. The numbers below are planning ranges from that work combined with market hourly rates. They are not a quote, and they are not what any of our clients paid. Use them to size a budget and to check whether a proposal on your desk covers the parts that make AI work in production.
Cost at a glance
| Project size | Typical scope | Effort (hours) | At Eastern Europe rates | At US onshore rates | Calendar time |
|---|---|---|---|---|---|
| AI feature in an existing app | One LLM task such as summaries, extraction, classification or drafting, using a hosted model API | 150-400 | $5,250-$32,000 | $12,000-$64,000 | 3-7 weeks (2 engineers) |
| Production AI product | Assistant or AI workflow with retrieval or integrations, evaluation set, guardrails, monitoring | 500-1,400 | $17,500-$112,000 | $40,000-$224,000 | 6-16 weeks (3 engineers) |
| Document or media pipeline at scale | OCR, audio or video processing, queues, batch jobs, model routing, review tools | 900-2,200 | $31,500-$176,000 | $72,000-$352,000 | 10-25 weeks (3 engineers) |
| Custom model platform | Fine-tuned or trained models, data labeling, serving infrastructure, several AI features | 1,800-4,000 | $63,000-$320,000 | $144,000-$640,000 | 15-34 weeks (4 engineers) |
Costs are hours multiplied by market rate bands. 2026 outsourcing rate surveys put Eastern Europe at about $35 to $80 per hour, and North America at $80 to $180 or more; we use $80 to $160 for the US band. The low end of each range assumes the low hours at the low rate, the high end the high hours at the high rate. Calendar time assumes about 30 productive hours per engineer per week.
What drives the cost
1. How many AI tasks, and how well-defined they are. “Summarize this support ticket” is one task with a clear output. “An assistant that answers anything about our platform and writes code” is several tasks plus a router that decides which one applies. Each task needs its own prompts, test cases and failure handling.
2. Your data. Most of the surprise hours hide here. If the model needs your documents, tickets or records, someone has to get access, clean duplicates and outdated content, split it into useful chunks and keep it in sync. Scanned PDFs, mixed languages and audio add preprocessing. Data audit and preparation alone is 40 to 160 hours in our estimator.
3. Model and API choice. Hosted APIs from OpenAI, Anthropic and Google cost nothing to set up and charge per token. Choosing a model is a cost decision twice: at build time (stronger models need fewer workarounds) and every month after (see running costs below). Many production systems route easy requests to a small model and hard ones to a large one, which takes extra engineering but pays back quickly at volume.
4. Evaluation. A demo works on the examples its builder tried. A production system needs a scored test set of real inputs with expected outputs, run automatically whenever a prompt or model changes. Without it you cannot tell whether a cheaper model is good enough or whether a fix broke something else. Budget 60 to 160 hours. Our post on validating LLM-generated code shows what this looks like when the output must actually run.
5. Guardrails and access control. Users will paste personal data, try to jailbreak the prompt and ask questions outside scope. Internal tools must not show a user documents they are not allowed to see. Input and output checks, PII handling and permission-aware retrieval are real work, typically 40 to 120 hours.
6. Integrations. An AI feature is only useful when it reads context from and writes results back to the systems people already use: CRM, ERP, helpdesk, CMS. Each integration needs auth, mapping, error handling and tests. One clean REST API is quick; a legacy system with no API is not.
7. Monitoring and operations. You need traces of what the model saw and returned, token spend per feature, latency, and alerts when quality drops. Without them, cost overruns and silent failures show up in the invoice or in customer complaints instead of on a dashboard.
To size your own scope, switch components on or off and pick a rate band.
The default selection is a typical production AI feature: discovery, data preparation, architecture, the LLM backend, a UI, integrations, evaluation, guardrails, monitoring, deployment and a pilot. That adds up to about 504 to 1,420 hours, which is the “production AI product” row in the table above.
Cost by component
| Component | Hours | Notes |
|---|---|---|
| Discovery and feasibility spike | 40-80 | Model trials on real examples before committing to a design |
| Data audit and preparation | 40-160 | Grows with messy, scanned or multilingual sources |
| Architecture and model selection | 24-60 | Providers, model routing, data residency, fallback |
| LLM feature backend | 80-200 | Prompts, structured output, tool calls, retries, rate limits |
| Retrieval over your data (RAG) | 80-200 | Ingestion, embeddings, search, citations, keeping the index fresh |
| Document or media pipeline | 120-320 | OCR, audio and video, chunking, queues and batch jobs |
| User interface | 60-160 | Chat or inline UI, streaming output, review and edit flows |
| Fine-tuning or custom model | 160-480 | Training data, runs, serving and comparison against API models |
| Integrations | 60-200 | Per system; legacy systems sit at the top of the range |
| Evaluation set and automated tests | 60-160 | The part most often missing from low quotes |
| Guardrails and access control | 40-120 | PII, prompt injection, permissions, output validation |
| Monitoring and cost tracking | 30-80 | Traces, spend per feature, quality alerts |
| Deployment and infrastructure | 30-100 | Environments, secrets, autoscaling, CI/CD |
| Pilot, tuning and handover | 40-100 | Real users, prompt tuning, documentation |
Running costs after launch
AI software has a running cost that ordinary software does not: every request to a hosted model is billed by tokens (roughly, pieces of words) in and out. List prices per million tokens on the official pricing pages in November 2026:
- Anthropic: Claude Opus 5.5 at $4 input and $20 output, Claude Sonnet 5.5 at $2 and $10, Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens. Anthropic’s pricing page also lists a 50% discount for batch processing and much cheaper cache hits for repeated prompt prefixes.
- OpenAI: gpt-6.1-sol at $2 input and $10 output, gpt-6-luna at $0.10 and $0.50, and text-embedding-3-small at $0.02 per million tokens. Transcription with gpt-4o-transcribe is estimated at $0.006 per minute and gpt-transcribe at $0.0045 per minute.
- Google: Gemini 3.1 Pro preview at $2 input and $12 output for prompts up to 200,000 tokens, Gemini 3.8 Flash at $0.75 and $3.75 through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027, and Gemini 3.1 Flash-Lite at $0.25 input and $1.50 output.
To make that concrete, take a feature that processes 50,000 documents a month at about 4,000 input tokens and 500 output tokens each. That is 200 million input and 25 million output tokens. At list prices it costs about $32.50 a month on Claude Haiku 5.5, $243.75 on Gemini 3.8 Flash, $650 on Claude Sonnet 5.5 or gpt-6.1-sol, and $1,300 on Claude Opus 5.5. The same work, a 40x spread. This is why the evaluation set matters: it tells you which of these models is good enough.
Audio is priced differently. Transcribing 10,000 hours of recordings (600,000 minutes) costs about $3,600 with gpt-4o-transcribe or $2,700 with gpt-transcribe at OpenAI’s published per-minute estimates, before any post-processing passes with a text model.
Beyond tokens, plan for:
- Hosting: the application, a database, a vector index if you use retrieval, queues and workers for batch jobs. For a typical business app this is usually a modest cloud bill; GPU servers only enter the picture if you self-host or fine-tune open models.
- Monitoring tools: tracing and evaluation platforms, often a per-seat or per-event subscription.
- Maintenance: providers retire and reprice models, and each change means rerunning the evaluation set and adjusting prompts. Budget a few days of engineering per quarter at minimum, more if the system is central to your business.
Build, buy or combine
You do not always need custom AI development. Buy when a SaaS tool already does the job inside the software you use: AI writing in your CMS, meeting summaries in your video tool, an AI agent in your helpdesk. A subscription is cheaper than any build when the workflow is generic.
Combine when a good tool exists but must fit your data or structure. Our Umbraco packages are an example: umContentGenerator and umContentTranslator put OpenAI and Claude into the CMS editor and keep Block Grid structure intact, so a site can add AI content without a custom project. We describe the approach in AI content generation in Umbraco.
Build when the AI has to work with your proprietary data, your own systems or your own domain language, when you need control over which model sees which data, or when per-seat SaaS pricing across a large team costs more than running your own. If the system should take multi-step actions on its own rather than answer or transform content, read our breakdown of AI agent development cost. If what you need is a customer-facing or internal bot, our guide to chatbot development cost goes deeper on channels, retrieval and human handoff.
How to get an accurate estimate
A useful AI estimate depends less on a features list than on real examples. The steps we follow:
- Pick one workflow. Who uses it, what goes in, what should come out and how often. Volume drives both architecture and running cost.
- Collect 30 to 100 real examples with the answer a good human would give. These become the first evaluation set.
- Map data and systems. Where inputs live, who may see what, where results must be written back.
- Run a short spike. Try two or three models on the examples and measure accuracy, latency and cost per request. This turns guesses into numbers.
- Estimate by component, with hours per module, an expected monthly model bill and named risks.
A vendor who quotes a fixed price before seeing any examples is either padding for risk or leaving out evaluation.
From our work
For Lokad, a supply chain software company, we built an assistant inside its platform that answers questions and writes scripts in Envision, Lokad’s own language. Generated scripts are run against the real Envision runtime, and failures go back for another pass: 82% run on the first attempt and 94% by the second, and users reach a first working script about 70% faster. Most of the engineering went into the parts this article warns about: mode detection, testing against the runtime and memory on Lokad’s Azure event store. Read the Lokad case.
For an educational lecture archive, we built a Kubernetes pipeline that transcribes long recordings mixing Hebrew, Aramaic and English with Gemini on Vertex AI. Recordings over 20 minutes are split at natural pauses, and each chunk is routed to Gemini Pro, Flash or Flash-Lite. That model selection lowered AI and API costs by up to 60%, while the pipeline runs more than 300 jobs at once at 99% transcription precision. Read the transcription case or the technical write-up on transcribing long multilingual audio with Gemini.
We have also wired AI services into existing CRM, ERP and support tools: chatbots, sentiment analysis of feedback and fraud alerts on transactions. Customer response times improved by 40%. Read the automation case.
Where OSKI fits
OSKI is an engineering company with a team mostly in Ukraine, registered in Estonia, working in .NET, React and AI. We build AI features into existing products and new AI systems end to end, including the data, evaluation, integration and monitoring work. See our AI development services, integrations and custom software development. Send us a workflow and a handful of real examples, and we will come back with a component-level estimate and an expected monthly model cost.
Planning ranges for typical scopes, including design, QA and project management. Not a quote.
Standard-tier list prices from the Anthropic, Google and OpenAI pricing pages, November 2026, before batch or caching discounts.
- 01 Pick one workflow 1
Who uses it, what goes in, what comes out, how often.
- 02 Collect 30 to 100 real examples 2
With the answer a good human would give.
- 03 Map data and systems 3
Where inputs live, who may see what, where results go.
- 04 Run a short spike 4
Try two or three models on the examples and measure.
- 05 Estimate by component 5
Hours per module, model cost per month, risks named.
Estimate your AI project
Pick the components your AI feature or product needs. Hours include design, QA and project management. The typical scope below is a production LLM feature with integrations, evaluation and monitoring.
