Marketplace products
Attribute AI cost per completed transaction, and be strict about the word completed. A buyer who runs six AI-assisted searches and abandons four, then retries twice and books once, has consumed the cost of thirteen operations against one revenue event. Matching and fraud screening consume tokens on listings that never transact, and content moderation does the same.
Review steps compound this. If a human checks flagged output before it goes live, that labor attaches to attempts rather than to conversions. Divide total monthly AI spend by completed transactions and compare it against your take rate. If the ratio moves the wrong way as volume grows, the problem is the funnel.
Internal workflow apps
There's no price to compare against, so compare against labor. Measure cost per completed task and add the exception review that a human still performs. Then set that total against the time the task used to take and the errors it used to produce. Throughput matters as much as hours saved, because faster cycle time in claims or onboarding has a value your finance team can already quantify.
Back-office work is where the returns have actually shown up. MIT's report documented $2–10M in annual savings from replacing outsourced support and document review, alongside a 30% reduction in external agency spend for content work. For internal AI in app development, the honest denominator is the fully loaded hour you're displacing, which includes the review hour you're adding back.
Build a worked estimate
Start with the request math and use published rates. Take a support-answer feature at 200,000 requests a month that averages 3,600 input tokens and 400 output tokens. The input count includes the system prompt and retrieved passages, and it includes 1.4 tool calls per request. On GPT-5, listed at $1.25 per million input and $10.00 per million output tokens, that's 720M input tokens at $900 and 80M output tokens at $800. Model spend lands near $1,700 a month.
Then add everything that isn't the model:
-
Embeddings and ingestion for the retrieval corpus, plus the monthly refresh as source documents change
-
Vector storage and reads, which at Pinecone's rates means paying the $50 floor until query volume grows
-
Observability at roughly $293 a month for 480,000 LLM spans, hosting for the application layer, and support handling for AI-specific tickets
Human review is the line that dominates. Reviewing 5% of 200,000 outputs at two minutes each is 333 hours, or about $8,300 a month at a $25 loaded hourly rate. Total run rate comes to roughly $10,400 a month, which is $0.052 per request, and the model itself is only 16% of it. AWS makes the same argument in its cost guidance. It advises teams to count initial development, ongoing operation, and management expense, because the inference line is only part of the cost. Any cost estimate for AI in app development that stops at tokens is off by a multiple.
Test cost scenarios
One number is a forecast nobody should sign. Run the expected case you just built and a high-usage case, then run a failure case where things go wrong at once. The point is knowing which variable breaks you first.
Vary the inputs that actually move the bill. Double the request volume and the picture barely changes, because human review scales with it and model spend was never the constraint. Now double context length instead and push retrieval depth from five passages to twelve. Raise the retry rate from 2% to 8% and shift 30% of traffic from the cheap model to the frontier one. Model spend can triple while request count stays flat, which is the scenario most AI app development cost models never test.
The failure case is worth writing down explicitly. A retrieval corpus that goes stale drives retries up and pass rates down, so your AI model evaluation metrics degrade, and your review queue grows at the same time. Enterprise spending on generative AI hit $37 billion in 2025, up 3.2x year over year, and a good share of that growth is workloads doing more work per interaction rather than serving more users.
Control production spend
Caching is the first lever because it's the cheapest to implement. OpenAI discounts reused prompt prefixes by up to 90% on cached input tokens, which matters enormously when your system instructions and tool definitions repeat on every call. For anything that doesn't need an immediate answer, the Batch API runs at 50% of standard pricing with turnaround under 24 hours.
After that, the levers are architectural. Route simple classification and extraction to a small model and reserve the expensive one for requests that genuinely need it. Cap context length so a pasted 40-page PDF can't quietly cost twenty times a normal request. Set per-tenant quotas, and attribute spend per feature so you can tell which part of the product is consuming the budget.
None of this works without measurement from real traffic. Clean test prompts don't retry and don't paste in irrelevant context. They also don't argue with the model for six turns. Production traffic does all three, which is why AI in app development budgets should be rebuilt on measured pilot data before launch rather than carried forward from prototype assumptions.
Review before each gate
Before funding the next stage, confirm five things:
-
Quality: pass rates from a current evaluation run, with failure cases reviewed by name
-
Security and data handling: reviewed, documented, and signed off
-
Unit economics: measured cost per account, unit, transaction, or task
-
Operating capacity: named owners for review, on-call, and corpus maintenance
-
Refreshed provider rates, since published prices change often enough to invalidate old math
Bring us your measured pilot data, and we'll help you build the model behind it. Our team works with product leaders on evaluation design and unit economics. We also work on budgeting for AI in app development, so reach out before your next funding gate.
