Does your v1 actually need AI?

The decision

Your pitch deck says AI because every deck does. The real question arrives at scoping, when "add AI" stops being a slide and starts changing your quote, your timeline, and your failure modes. Get it wrong in one direction and you've bolted AI onto a product that didn't need it — a permanent per-use cost and a brand-new way to be wrong, forever. Get it wrong in the other direction and you've shipped a commodity in a category where the model is the value. I run production AI systems at serious scale in my day job, and the most useful thing that experience has taught me is how often the right answer is "less AI than the deck says."

The questions that actually determine it

If the AI part were a very fast human intern, what would it be doing?

This is the gate for everything else. Walk to the exact moment in your product where AI would act, and imagine a very fast intern sitting there instead. If the intern would be looking things up or applying fixed rules — matching a pincode to a delivery zone, checking a form for missing fields, filtering a list — that is plain software wearing an AI costume, and building it with a model makes it slower, costlier, and occasionally wrong. If the intern would be reading, judging, drafting, or summarizing — things where the input varies and the output requires interpretation — you have a real AI candidate.

Is being wrong sometimes acceptable in this exact spot?

Models are probabilistic. A model that's right 95% of the time is wrong one time in twenty, indefinitely, and no prompt eliminates that. So ask about the specific spot: what happens when this output is wrong? A clumsy draft the user edits anyway — fine. A wrong price, a wrong dosage, a wrong legal clause — not fine. Where errors are unacceptable you have exactly two honest moves: add a human review step to the design, or move the AI somewhere lower-stakes. Pretending the model won't err is not a third option.

Is the AI the product, a feature, or garnish?

Strip the AI out entirely. Is there still a product someone pays for? If yes, AI is a feature — one line in the build quote and one line in the monthly bill. If no, AI is the product, and everything gets harder: the build is bigger, the evaluation burden is real, and your differentiation question is urgent. And if the AI is there so the deck can say "AI-powered" while the intern test says lookup-and-rules — that's garnish. Cut it. Garnish costs real money per use and adds a failure mode to a flow that had none.

What does each interaction cost at 1, 100, and 10,000 users?

AI is probably the first product cost you've had that scales with usage rather than headcount. A chat-style feature typically costs $0.005–0.05 per interaction; per active user that lands around $0.10–2.00 a month depending on usage depth. At 100 users, noise. At 10,000 users on the expensive end, $20,000 a month — a payroll-sized line item you signed up for one prompt at a time. Run this math at three tiers before anything is built, and know one lever dominates it: a cheap-tier model is 10–30x cheaper per call than a frontier model, and most v1 features don't need the frontier.

Your options, with honest costs and risks

Plain software — no model

Rules, search, and forms doing the job. Build cost: whatever your MVP quote already says, no delta. Run cost: effectively zero per use. Risk: only that you misdiagnosed and the task really did need judgment. If the intern test said lookup-and-rules, this is not the fallback option — it's the correct one, and "no AI in v1" is a success case, not a failure to be apologized for.

The AI feature

One model call in one flow: summarize this document, draft this reply, classify this ticket. Build delta on an MVP quote: +₹1.5–6 lakh ($2,000–8,000) for one well-scoped model-backed flow. Run cost: the per-interaction math above. Risks: quality drift, prompt maintenance, and creeping scope — one feature becoming five is how a $50 bill becomes $2,000. This is the right shape for most v1s that pass the intern test.

The AI workflow or agent

Multi-step, tool-using, model-driven: read the inbox, look up the account, draft and send. Build cost: 3–5x the single-feature delta — call it ₹4.5–30 lakh ($6,000–40,000) on top of the base build. Run cost multiplies too, because one user action triggers many model calls. Risks compound: each step can be wrong, so end-to-end reliability is the product of every step's reliability, and debugging needs specialists. Almost never the right v1 shape. Earn your way here from a working single feature.

The wrapper trap

A thin UI over one prompt. As a two-week MVP to test whether anyone wants the output — completely valid; some real companies started exactly here. As a defensible product — it isn't one. Anyone can rebuild it in a weekend, including the model vendors. A wrapper is a valid experiment and a weak asset: ship it to learn, but your moat must be forming somewhere else — your data, your workflow, your distribution — while it runs.

The bill, worked once

One feature at ~10 interactions per active user per month, mid-range pricing: at 1 user, pennies. At 100 users, roughly $10–200/month (₹800–17,000). At 10,000 users, $1,000–20,000/month (₹80,000–17 lakh). The spread within each tier is mostly model choice — that 10–30x cheap-versus-frontier gap — which is why the model decision is a budgeting decision, not just a technical one.

What I'd recommend

Run the intern test first, and let it gate everything. Lookup-and-rules: build plain software and put "AI" nowhere near the spec. Judgment: proceed — to the smallest shape that passes. That's almost always one AI feature, not a workflow, and never an agent in v1.

Before paying anyone to build it, prototype the prompt yourself in a chat window. Paste ten real examples of your actual input into ChatGPT or Claude and look at the output. If the raw model can't do the task with your data in front of you, no agency deliverable will fix that — and you'll have learned it for free instead of for ₹4 lakh.

Decide the acceptable monthly cost per user before choosing a model, not after the first invoice. Set the ceiling — say, $0.30 per active user per month — and pick the cheapest model that clears your quality bar under it. Start cheap-tier and upgrade on evidence; the 10–30x spread means the frontier model has to be dramatically better to be worth it, and in most v1 flows it isn't.

And where being wrong is unacceptable: human review in the loop, or move the AI down-stakes. Every time.

When this doesn't apply

• Regulated decisions. Medical advice, lending, hiring, insurance: model output feeding a regulated decision needs a compliance framework designed by someone who knows that regime. The intern test still helps you scope; it does not make the output legal.
• The model is the moat. If your entire pitch is that your model does something others can't, you're in specialist territory — evaluation pipelines, data strategy, possibly training. This guide covers using AI, not being an AI company.
• Hard real-time or on-device constraints. Sub-100ms responses, offline operation, on-device inference: the architecture is dictated by the constraint, and API-based cost math above doesn't transfer.
• You already have ML engineers. The build-delta economics assume you're buying the work from an agency. In-house specialists change the math, though the intern test and the bill ceiling still apply.