The short answer: at Oysty a focused AI prototype starts at USD 3,000, a production assistant with evaluation, permissions and integrations usually lands between USD 8,000 and USD 25,000, and an AI-powered mobile app is quoted after scoping because device testing and store releases add work. Those are our published bands, not a market average — quotes from other studios will differ.
What the three budget tiers actually buy
“An AI app” covers everything from a single document question box to a mobile product with accounts, billing and an admin console. The price difference is mostly about how much engineering sits around the model.
| Stage | What ships | Typical band | Timeline |
|---|---|---|---|
| Prototype | One workflow over a small document set. No accounts, billing or admin. | USD 3,000–5,000 | 3–5 weeks |
| Production assistant | Retrieval, permissions, an evaluation set, logging, integrations and a human review step. | USD 8,000–25,000 | 6–10 weeks |
| AI-powered product | Mobile or web clients, accounts, billing, admin, monitoring and store release. | Scoped individually | 10+ weeks |
The production tier is where most of the value sits. A demo can answer a question; a product has to answer it for the right person, cite where the answer came from, refuse when the source is missing, and leave a record of what happened.
What drives the price
- Data readiness. Clean, text-searchable documents cost less than scanned PDFs, shared-drive sprawl or records that live in three systems. We can clean it; it is simply work.
- Integrations. A standalone assistant is cheaper than one wired into your CRM, help desk, WhatsApp and billing. Each integration adds authentication, error handling and permissions.
- Accuracy work. Building a test set, tuning retrieval, adding reranking and measuring precision takes longer than connecting an API. It is also the difference between a tool people trust and one they abandon.
- Interface and workflow. Chat is not always the right surface. An extraction screen with a side-by-side document view costs more to design than a text box.
- Privacy and compliance. On-premise inference, no-training API settings and audit logs add setup work, especially when the data cannot leave your infrastructure.
- Maintenance. Models change, documents change and prompts drift. A small ongoing budget keeps quality from decaying silently.
The costs that are not in the build quote
Every AI system has running costs after launch. They fall into three buckets:
- Model usage. Providers charge per token, so cost scales with questions asked and document length. On-device or self-hosted models trade that bill for infrastructure and engineering.
- Hosting and storage. Vector indexes, file storage, background jobs and logs need somewhere to live. A prototype can run on almost nothing; a team-wide tool cannot.
- Human review. If the output has commercial or legal weight, someone still approves it. The automation saves time on the first draft, not the judgement.
How to compare three AI app quotes
The lowest number is not automatically the best value. Ask every vendor the same five questions:
- Which single workflow is in scope, and which are explicitly out?
- How will accuracy be tested before go-live, and against whose documents?
- Where does the data go, which provider sees it, and is training disabled?
- What happens when the model is unsure or wrong — refusal, fallback, escalation?
- Who owns the code, prompts, indexes and evaluation set at handover?
A quote that cannot answer those questions is a guess with a logo on it.
What we would do with a USD 3,000 budget
Build a question-answering assistant over about 200 pages of your own material — a policy manual, a set of past proposals, a support knowledge base. Measure it on 40 real questions, including questions the documents cannot answer. If it cites correctly and declines cleanly, you have evidence for the next stage. If it does not, you have saved yourself a much larger invoice.
If that sounds like the right starting point, read how we build retrieval-augmented assistants or look at the Private AI Assistant case study, which runs entirely on a phone.
Mayur Saptal
CEO at Oysty India. Writes about AI delivery, scoping and the unglamorous part of shipping software. You can check the public client reviews on Upwork.