Guide
Custom AI MVP Cost and Timeline in India
How Agentic Academy Labs scopes discovery vs production in India: timelines, cost drivers, and why model spend is a separate line from the build.
Published 2026-08-08 · Updated 2026-08-14
Founders in India often ask Agentic Academy Labs for a single rupee number for 'an AI MVP.' We will not invent a fake exact price. Discovery and production are different jobs, and model tokens are a monthly operating cost, not part of the engineering invoice. From Sikar we scope both in writing after we have seen your documents, channels, and compliance constraints. The ranges below are planning tools. The quote you should trust is the one that follows discovery.
Discovery vs production (do not mix the invoices)
Discovery is one to two weeks: data map, success metrics, eval questions, integration list, and a written estimate. It is cheap compared with building the wrong Slack bot. Production is the build: auth, Next.js admin, retrieval, monitoring, handoff docs. A proof of concept sits in between and is allowed to be ugly. Teams that skip discovery usually pay for a second production start. Interns on our bench are not a substitute for a scoped discovery; they join after the corpus and metrics exist.
Typical calendar, not a promise
| Phase | Calendar | What you should hold in your hand | What it is not |
|---|---|---|---|
| Discovery | 1-2 weeks | Scope, eval set sketch, data map, estimate | A chatbot on production traffic |
| Proof of concept | 2-4 weeks | Thin slice on real docs, logged answers | Admin, SSO, or five channels |
| Production MVP | 4-12 weeks after discovery | Auth, admin, eval loop, rollback notes | A fully staffed AI ops team |
| Hardening | Ongoing | Cost caps, quality regressions, runbooks | Included by default in a PoC fee |
Separate model spend from build spend
- Build cost is people time: engineering, design, and review. It is mostly front-loaded.
- Model spend is tokens, embeddings, and sometimes fine-tune jobs. It starts small in week 1 and can outrun the build once employees actually use Slack.
- Hosting is Next.js (often Vercel), Postgres, file storage, and a vector index. Idle spend is easy to miss after demo day.
- We ask clients to put model usage on its own budget line with a monthly cap and an alert. Mixing it into 'project cost' hides the bill until finance notices.
Cost drivers we measure before quoting
| Driver | Lower effort | Higher effort |
|---|---|---|
| Documents | Searchable exports, one source of truth | Scans, conflicting Drive folders, no OCR |
| Channels | One web chat in Next.js | Slack plus WhatsApp plus CRM tools |
| Admin | Engineers upload via SQL | Non-engineers need review queues |
| Compliance | Internal tool, no PII | PII redaction, audit logs, private networking |
| Eval rigor | Ten founder questions | Golden set, automated regression, human raters |
| Mobile | Web only | Flutter app sharing the same Bun API |
What week 1 of discovery looks like
- List systems and owners. If no one owns the handbook, RAG will cite ghosts.
- Pull a sample of real questions from support or Slack. Invented FAQs make cheap prototypes that fail in week four.
- Score data readiness: can we parse it, is it current, is it tenant-safe?
- Agree one success metric (example: 'correct citation on 30 golden questions') instead of 'make it smart.'
- Write the estimate with two columns: build (time-boxed) and run-rate (models plus hosting).
Planning mistakes we push back on
- Asking for an exact INR quote from a one-line WhatsApp description. We will give drivers and a discovery plan instead.
- Budgeting only GPU or API credits and forgetting admin UI, eval, and handoff.
- Fine-tuning in the MVP because a competitor mentioned it. Read RAG vs fine-tuning for startups.
- Treating intern hours as a way to skip production observability. Interns help with eval sets and upload UIs; they do not replace a rollback plan.
We run fixed-scope MVPs after discovery, retainers when the roadmap will move, and venture-studio pairing when a founder is co-building with Agentic Academy Labs from Sikar or remote. For a six-week production slice that included admin and Slack, see the Bangalore fintech RAG case study. Architecture notes live in Next.js AI app architecture. Start from custom AI development or schedule a call.
Frequently asked questions
- How much does a custom AI MVP cost in India?
- It depends on data readiness, channels, admin needs, and compliance. Treat published rupee figures on the internet as marketing. Get a written estimate after discovery, with build and model spend as separate lines.
- What is a typical timeline?
- Discovery 1-2 weeks, proof of concept 2-4 weeks, production MVP often 4-12 weeks after discovery. Hardening is ongoing and is not the same phase as the PoC.
- Why separate model spend from the build?
- Engineering is mostly a project cost. Tokens scale with usage and prompt size. If you blend them, finance cannot tell a slow product from an expensive model.
- Fixed price or retainer?
- Fixed-scope after a clear discovery. Retainer when the backlog will change. Contact us or schedule a call and we will say which one fits.
- Does internship labor change the quote?
- Interns at Agentic Academy Labs help with eval questions, document cleanup, and admin screens. Production responsibility, security, and estimates still sit with the delivery team.