$2.4M. That is what a US healthtech company was on track to spend in Year 2 on cloud AI inference for a feature that was approved with a $90,000 development budget and no inference cost model. The feature was popular. Users queried it 12 times per session on average. At $0.02 per query and 400,000 monthly active users, the math was not in the budget because no one built the model before launch.

The board mandate to "add AI" is real and common. The cost model that should accompany it is not. This guide builds the full per-user cost framework - development, inference, and scale - so the number you present to the board is the number that actually lands.

Key findings
Development cost: $40K to $200K depending on AI feature type. One-time.
Cloud inference cost: $0.003 to $0.08 per query. Scales with every active user, every session.
On-device inference: zero marginal cost per query after development. Higher upfront development cost by $30K to $80K.
The cost cliff hits most enterprises between 50,000 and 150,000 MAU. Model it before launch, not after.

The two cost buckets

Every AI feature in a mobile app carries two separate cost categories. Most enterprise budgets model only the first.

Development cost. The one-time engineering cost to design, build, integrate, and launch the AI feature. This is the number that appears in budget requests. It covers prompt engineering, API integration or on-device model embedding, UI design for the AI interaction, testing, and App Store submission.

Inference cost. The ongoing cost of running the AI feature for every user, every session, every query. For cloud-based AI features, this is a per-query charge billed by the AI provider. For on-device features, inference cost is zero per query - the computation runs on the user's device at no marginal cost to you.

The second bucket is where enterprise AI feature budgets collapse. Development cost is finite and visible. Inference cost is infinite and invisible until the billing arrives. A feature used by 10,000 users costs almost nothing in inference. The same feature at 500,000 users costs between $18,000 and $2.4M per month depending on model choice and query volume.

Build both numbers before you approve the feature. Not the development cost in isolation.

On-device vs cloud inference cost

The choice between on-device and cloud AI fundamentally changes the cost structure of an AI feature. The two models do not trade off quality one-for-one - they trade off cost, privacy, capability, and device compatibility.

Cloud inference. The user's input goes to an external model API (OpenAI, Anthropic, Google, or a hosted open-source model). The model processes the input and returns a result. The enterprise pays per token or per query. Cost scales linearly with usage.

On-device inference. A quantized or distilled AI model runs directly on the user's phone using Apple's Core ML, Google's ML Kit, or a third-party framework like ONNX Runtime. Processing happens locally. There is no per-query cost. The model is bundled with or downloaded to the app.

The development cost differential matters. Integrating a cloud AI API adds $25,000 to $60,000 to a feature's development cost. Integrating an on-device model adds $60,000 to $140,000 - because model selection, quantization, device compatibility testing across the device matrix, and on-device performance optimization require significantly more engineering work.

The on-device premium pays for itself when query volume is high and the use case does not require the latest large-scale models. Smart text completion, document scanning, basic image classification, and local search all work well on-device. Complex reasoning, real-time retrieval from large knowledge bases, and generative tasks that require current information do not.

Cost per user at scale

The table below models monthly AI inference cost at four MAU levels for three representative AI feature types. Costs assume one AI interaction per user per day (30 interactions per MAU per month).

The 500,000 MAU row illustrates why the model choice matters as much as the feature itself. A conversational assistant at scale costs $2.16M per year in inference alone. A smart search feature using a small model costs $540,000. The same feature implemented on-device costs nothing in inference - though the app download is larger and the development cost was higher.

The cost cliff

Most enterprise AI feature budgets are modeled at current or near-term user scale. The cost cliff is the point at which the inference cost exceeds the budget that was authorized for the feature - typically discovered not during planning, but when a quarterly billing report arrives.

Three conditions create the cost cliff.

Higher query frequency than assumed. If users query the AI feature 8 times per session rather than the modeled 2 times, the inference cost is 4x the estimate. Features that are genuinely useful get used more than conservatively modeled. That is good news for the product and bad news for the inference budget.

Faster user growth than planned. A mobile app that grows from 30,000 to 150,000 MAU in six months carries an AI inference bill that grows by the same factor. Growth plan revisions that do not trigger inference cost revisions are a budget gap waiting to happen.

Model selection drift. Engineering teams sometimes upgrade to a more capable model mid-development because the smaller model did not meet quality requirements. A model switch from $0.003 per query to $0.02 per query multiplies the monthly inference cost by 6.7x. Budget for the model you will actually ship, not the cheapest option in the initial spec.

The cost cliff defense is a scale scenario table built before development, not a post-launch cost review. Model the inference cost at your current MAU, your 6-month growth target, and a 2x growth scenario. Show all three numbers to the team and the budget committee. If the 2x scenario is unaffordable, decide now whether to use an on-device model, rate-limit the feature, or charge users for premium access.

Read more case studies at mobile.wednesday.is/work

What features justify the cost

Not every AI feature has a measurable return that justifies its inference cost. Three feature categories have the clearest and most defensible ROI at enterprise scale.

Support deflection. An in-app AI assistant that resolves user questions before they reach a human agent has a direct, measurable financial return. Enterprise support contacts cost $8 to $22 each, depending on channel and complexity. An AI assistant that deflects 25 to 35% of contacts against an inference cost of $0.012 per query generates a positive return at almost every scale above 20,000 MAU. The math is straightforward enough for a CFO to follow in a single slide.

Conversion improvement. AI-powered product recommendations, smart onboarding flows, and personalized content sequencing lift conversion rates in ways that are attributable and measurable through A/B testing. A 2-percentage-point lift in free-to-paid conversion at an average revenue of $40 per user, applied to 100,000 monthly app visits, generates $80,000 per month in incremental revenue against an inference cost that typically runs $5,000 to $25,000 per month for recommendation-type features.

Error and friction reduction. AI-powered form validation, document scanning, and data extraction reduce user abandonment at high-friction points. The return is measured in completed transactions rather than deflected support contacts. For financial services, logistics, and healthcare apps where incomplete transactions have direct revenue impact, friction reduction AI often generates the fastest payback period.

Features that do not justify the cost at mid-market enterprise scale: conversational agents that handle open-ended questions without a defined support scope, AI-generated content that users do not notice versus non-AI content, and personalization features in apps with user bases too small to generate meaningful training signal.

How to present AI ROI to a board

Board presentations on AI feature investment fail in one of two ways: too vague ("AI will improve the user experience") or too granular (inference cost tables that no one in the room can connect to business outcomes).

The format that holds up uses three numbers and one assumption.

Number 1: Development cost. One-time. Specific. "$95,000 to build and launch the in-app support assistant."

Number 2: Annual inference cost at target scale. Specific to the feature and the expected MAU. "$54,000 per year at 150,000 monthly active users using one query per user per session."

Number 3: The return. Denominated in dollars, not percentages. "At 30% deflection of our current 8,000 monthly support contacts at $12 average cost, annual support savings are $345,600."

The assumption. State the one number the ROI model is most sensitive to and what happens if it is wrong. "If deflection rate is 15% instead of 30%, annual savings are $172,800 and payback extends to 18 months instead of 9."

This structure lets a CFO stress-test the model in the room rather than asking for a follow-up analysis that delays the decision. A board that can challenge the assumption and still see a positive return approves the budget. A board that cannot follow the logic asks for more analysis.

One additional guardrail: do not attribute all improvement in a metric to the AI feature if other product changes shipped at the same time. A support deflection rate that improved by 28% in Q3 may reflect the new FAQ redesign that shipped simultaneously. Isolate the AI feature's contribution through holdback testing - a control group that does not see the AI feature - before claiming the full return in a board update.

---

Originally published at https://mobile.wednesday.is/writing/ai-feature-cost-per-user-modeling-guide-2026