Where should AI actually fit in my product?
Where should AI fit in your product? A non-technical founder's test for separating AI that earns its cost from AI theater, and what to check before you build it.
Put AI only where it earns its place: a specific, repetitive task where speed matters, a human stays in the loop, and you can point to the cost it cuts or the experience it improves. Everything else is AI theater — features added because competitors mention AI, not because users need them. The question is never "should we use AI," but "which one task, and what does it move?"
When does AI actually belong in a product?
AI belongs on repetitive, high-volume tasks that tolerate imperfect answers and keep a human in the loop: support triage, drafting, summarization, classification, and search. Tools like OpenAI's ChatGPT, Anthropic's Claude, and GitHub Copilot excel there because speed and coverage matter more than perfect precision. The narrower the task, the more reliably AI pays off.
The strongest AI features are unglamorous. Intercom's Fin and Zendesk's assistants deflect routine support tickets; a summarization step saves a user from reading a long thread; a classifier routes work automatically. Each targets one task, has a fallback when the model is wrong, and moves a measurable number — response time, deflection rate, hours saved. That is the shape of AI that survives contact with real users.
Where does AI usually fail to earn its cost?
AI fails when it is added for image rather than utility: a chatbot bolted onto a site nobody asks questions on, a generic recommendation engine users do not trust, or a headline feature chasing hype. These raise build and running cost, add failure modes, and get ignored. Unused software is pure cost, and AI theater is expensive to run.
The trap is competitive mimicry. A founder sees rivals announce AI, so a feature gets funded to keep pace, not to help anyone. It demos well and then sits idle, because it never mapped to a real job. Worse, generative features that occasionally produce wrong or embarrassing output can damage trust more than the feature ever added. Absence of AI is not a weakness; misplaced AI is.
How do you tell AI leverage from AI theater?
Run one test: name the task, the metric it moves, and the recurring cost per user. Leverage passes all three; theater fails at least one. If you cannot say which task gets faster, which number improves, and what each use costs to run at scale, you are buying the appearance of AI rather than its benefit.
| Signal | AI that earns its place | AI theater |
|---|---|---|
| Task | One specific, repetitive job | Vague "smart" everything |
| Metric moved | Named (deflection, hours, speed) | None you can point to |
| When wrong | Human catches it; low stakes | User sees it; trust drops |
| Adoption | Staff or users actually use it | Sits idle after launch |
| Cost | Known per-use, worth it | Unmeasured, recurring |
MIT Sloan Management Review and other researchers have repeatedly found that most enterprise AI initiatives fail to reach production or return value, usually because of weak problem selection rather than weak technology. The models are rarely the bottleneck. The decision about where to point them is.
What should a non-technical founder check before adding AI?
Before building, confirm the data exists, the recurring per-use cost is acceptable, a fallback handles wrong answers, and someone will actually adopt the feature. Also confirm the model connects to your real workflow rather than standing alone. Skipping these checks is how founders fund AI that looks impressive in a demo and collapses in production.
- Data readiness. AI needs relevant, accessible data; without it, even a good idea stalls.
- Running cost. Model calls recur per use — estimate cost per active user, not just the build.
- Failure handling. Decide what happens when the output is wrong before you ship it.
- Adoption. If staff or users will not change behavior to use it, the feature is dead on arrival.
How do you add AI without wasting the build?
Add AI as a scoped decision, not a theme: pick one task, define the metric and cost, ship a narrow version, measure adoption, then expand only if it earns it. An independent review keeps AI aligned to the product instead of the hype cycle, catching theater before it consumes budget a non-technical owner cannot easily judge.
This is exactly where an outside view pays for itself, because vendor and internal incentives both push toward more AI, not the right AI. A short, fixed-scope Mobile Product Roadmap decides where AI genuinely helps your product, where it does not, and in what order to build, before you commit real money to a model.
Key Takeaways
- Ask "which one task and what does it move," never "should we use AI" in the abstract.
- AI earns its place on repetitive, high-volume tasks that tolerate imperfect answers with a human in the loop.
- AI theater — features added to look modern — raises cost, adds failure modes, and gets ignored.
- Use the three-part test: name the task, the metric moved, and the recurring per-user cost.
- Most enterprise AI efforts fail on problem selection, not technology, per MIT Sloan research.
- Check data readiness, running cost, failure handling, and adoption before building anything.
FAQ
- Does my product actually need AI?
- Most products do not need AI to be valuable; they need AI only where it removes real friction. If you cannot name the specific task AI speeds up, the money it saves, or the experience it improves, you do not need it yet. Solve the core product first.
- What is AI theater?
- AI theater is adding AI to look modern rather than to help users — a chatbot nobody uses, a generic recommendation nobody trusts, a feature added because competitors mention AI. It raises cost and complexity without moving a number you care about, and users quietly ignore it.
- Where does AI create the most value in a product?
- AI creates the most value on repetitive, high-volume tasks with tolerance for imperfect answers: support triage, drafting, summarization, search, and classification. These are places where speed matters more than precision and a human stays in the loop. Narrow, well-scoped uses beat broad, ambitious ones.
- How much does adding AI to a product cost?
- Beyond development, AI adds ongoing per-use costs for model calls, plus monitoring, evaluation, and guardrails that many founders overlook. A feature cheap to prototype can be expensive to run at scale. Estimate the recurring cost per user before committing, not just the build price.
Giacomo Balli is an independent mobile product advisor in San Francisco. He helps non-technical founders decide where AI genuinely improves a product and where it only adds cost, so budget goes to leverage rather than hype.
Disclosure: Giacomo Balli provides independent advisory services and does not resell AI tools or development work, take equity, or earn vendor commissions. Product names are examples, not endorsements.