Shimira IT
Most content about software cost gives you a range — "$50k to $150k" — and stops. That's useless when you're trying to decide whether to start. What actually helps is seeing where the money goes, phase by phase, on a real build.
So here's one. This is an anonymized teardown of a real AI product we scoped: a consumer platform that generates personalized, illustrated output from user-supplied inputs, sold direct to consumers, with a human review step before anything ships to the customer. We've stripped identifying details and widened the exact figures into ranges, but the structure, the proportions, and the lessons are real.
The headline
- Full build: roughly $90k–$125k
- Lean launch: roughly $70k–$90k
- Timeline: ~6 months to launch
- The one number that mattered most: the per-unit AI generation cost, which wasn't known until we measured it — and which drives the entire business model
Where the money went, by phase
| Phase | Share of budget | What it bought |
|---|---|---|
| Discovery & technical spec | ~5% | Workflow mapping, architecture, the itemized plan itself |
| AI feasibility sprint | ~7% | Proving the hard thing worked, before the main spend |
| UI/UX design | ~10% | The customer flow and the operator panel, prototyped |
| Core platform | ~22% | Accounts, the builder, storage, history |
| AI generation pipeline | ~26% | The actual engine — the largest single slice |
| Payments & tax | ~7% | Checkout, receipts, the revision-purchase flow |
| Logistics / fulfillment | ~6% | The physical delivery integration |
| Operator admin panel | ~8% | The human-review and dispatch tooling |
| Notifications | ~3% | Transactional email |
| QA & validation | ~7% | Testing, including real physical proofs |
| Launch & stabilization | ~3% | Deploy, monitor, fix |
Two things in that table surprise most first-time buyers.
The AI engine is the biggest slice but not a majority. People assume an "AI product" is mostly the AI. It isn't. The generation pipeline was the single largest line at ~26%, but the un-glamorous platform around it — accounts, storage, payments, the operator panel, fulfillment — added up to more than half the budget combined. This is the rule, not the exception: in most AI products, the AI is the smaller engineering problem. The product around the AI is where the hours go.
A human sits in the loop, and that's a feature. ~8% went to an operator admin panel so a person reviews every output before it reaches the customer. For a consumer AI product where quality is the brand, that review step is cheaper than the reputational cost of shipping a bad result — and far cheaper than trying to make the AI perfect.
How this compares to a "typical" build
It's worth placing this against the generic ranges, because an AI product reshapes them. A conventional web platform of similar size spends roughly 45–55% of its budget on core development and relatively little on anything AI-specific. This build looks different in three ways:
- A feasibility phase exists at all. A standard CRUD app doesn't need one — you know a form can save a record. An AI product needs to prove the hard thing is even possible, so ~7% goes to de-risking before the main build.
- Design carries more weight for a consumer product (~10%) than for an internal tool, because the customer-facing flow is the product.
- Fulfillment and an operator panel exist (~14% combined) because this product ships something physical with a human check — costs a pure-software SaaS wouldn't carry at all.
The lesson isn't that these percentages are universal — they're not. It's that the shape of the budget follows the shape of the product, and anyone handing you a generic percentage breakdown without asking what you're building is guessing.
The phase that paid for itself: feasibility
The second phase — ~7% of the budget — was a dedicated sprint to prove the single hardest technical requirement worked, on real inputs, at the required quality, before committing the other 93%.
This is the most important structural decision in any AI build, and it's where we'd push back on any vendor who skips it. AI projects carry a risk ordinary software doesn't: sometimes the accuracy or consistency your product needs simply isn't achievable at acceptable cost, and you cannot know that from a demo. Spending ~7% to find out — with a real decision gate after it, where the client could stop, adjust, or proceed — is the difference between a measured bet and a six-figure hope. We write about this pattern in how much a SaaS MVP costs and choosing a software development company.
What a feasibility sprint actually produces is worth being concrete about: sample outputs from real user inputs (not cherry-picked demos), a measured quality rate you can hold to, the real per-unit generation cost, and a go/no-go recommendation with the evidence behind it. If the answer is "this doesn't work at the quality you need," you've learned it for ~7% of the budget instead of 100%.
The number that wasn't in the build cost at all
The build was a one-time cost. The decision that shaped the whole business was a recurring one: the AI generation cost per unit sold.
Every unit this product sells consumes model API calls — and for a generation-heavy product, that's real money per sale, not rounding error. We estimated it up front as a range and then measured it during the feasibility sprint, because it sets the floor on retail price and the ceiling on margin. A product that costs more to generate than customers will pay has no business model, no matter how good the app is.
This is the cost that build-cost estimates routinely omit and that founders routinely discover too late. For any AI product, model the per-unit inference cost before you model anything else. It's covered as a budget factor in our cost to fix an AI-built app and custom software development cost guide.
The ongoing costs after launch
The build was ~$90k–$125k. Running it is its own line:
- Per-unit AI generation: scales directly with sales — the margin-defining number above
- Infrastructure: hosting, storage, email — a few hundred dollars a month at launch, scaling with volume
- Maintenance: the industry-standard 15–25% of build cost per year
A buyer who budgets only the build and forgets the run is the most common way these projects get into trouble in year two. On a $100k build, that maintenance line alone is $15k–$25k a year — not optional spending, but the cost of the software continuing to work as dependencies age and the model landscape shifts underneath it.
Where a lean launch found its savings
The gap between the ~$90k–$125k full build and the ~$70k–$90k lean launch wasn't cut corners on the core — it was deferral of things that could safely come later. The revision-purchase flow, some of the fulfillment automation, and polish on the operator panel were pushed to post-launch, because none of them block the first real sale. The feasibility sprint, the generation pipeline, payments, and a basic operator review were not negotiable, because without them there's no product to launch. Knowing which is which — what blocks launch versus what can wait — is most of what a good discovery phase buys you.
What to take from this
- The AI is rarely the biggest cost. Budget for the whole product around it.
- De-risk the hard thing first, with a real gate. ~7% to know beats 100% to hope.
- Model per-unit inference cost before anything else on a generation-heavy product.
- Budget the run, not just the build — per-unit cost, infrastructure, and 15–25%/year maintenance.
- The budget's shape follows the product's shape — treat any generic percentage breakdown as a starting hypothesis, not an answer.
If you're scoping an AI product and want a teardown like this for your idea — itemized, phase by phase, with the feasibility step built in — tell us what you're building. You'll get a real breakdown within 48 hours, in the same format shown here.
This teardown is based on a real engagement. Identifying details have been removed and exact figures widened into ranges to protect the client's confidentiality and pricing. Proportions and structure are faithful to the actual build. Figures are illustrative of this project, not a quote for yours.