Honest cost drivers for chatbots, document AI, and mobile AI apps — without a fake public price list.
There is no honest one-line price
Anyone publishing “AI app: $X” without seeing your data and integrations is selling a template. Real cost moves with four things: how messy the input is, how many systems you must connect, whether you need iOS/Android/web, and how much a human must still approve.
On our enquiry form we ask for budget bands ($5K–$10K, $10K–$25K, $25K–$50K, $50K+) so we can say quickly if we are the wrong size. That is more useful than a fake calculator.
What usually sits in a first version
A focused chatbot or RAG assistant over a known document set, with a simple admin way to add content, is a smaller build than a multi-tenant SaaS with billing.
A mobile app with photo upload, accounts, and an LLM backend is larger: store listings, auth, and offline behavior cost time even when the model call is simple.
Computer vision (detection, coaching, traffic) needs data, evaluation, and a place for false positives. Budget the review UI, not just the model.
What inflates cost without adding value
Training your own foundation model. Fine-tuning before you have a workflow. Building a “platform” for one process. Hiring a theatre of data scientists for a job an API plus good retrieval can do.
We would rather ship a narrow production slice you can run than a strategy deck. If buy-a-SaaS-tool is cheaper and 80% fit, we will say that on the call.
How to get a real range
Bring: the process paragraph, sample files or screens (redact them), and your budget band and timeline. We reply with fit, a rough shape of v1, and whether you should wait. That is the start of an estimate — not a binding quote until scope is written down.