Every pitch deck in 2026 has “AI-powered” somewhere on slide one. Investors expect it. Users are curious about it. But here’s what most founders won’t admit: half the AI features out there don’t actually need to exist.
Before you bolt AI onto your product, let’s figure out what’s worth building and what’s just noise. This post is the framework I use when a founder asks whether to add AI, what to buy versus build, and how to find out cheaply whether it works.
Start with the problem, not the model
The most reliable way to waste a quarter is to start from “we should use AI” and look for a place to put it. Start from a user problem and ask what the cheapest thing is that solves it.
AI is a good fit when the input is unstructured (free text, documents, images, audio), when the correct answer is fuzzy rather than exact, or when the volume makes handling it by hand impractical. It is a poor fit when the answer is a lookup, a calculation, or a rule. A deterministic system is cheaper, faster, testable, and never confidently wrong. If a rule handles ninety percent of cases, put the rule first and send only the remainder to a model.
What tends to be worth building
High return: support and internal automation
This is the easiest win. AI-assisted support (ticket routing, drafted replies, answers grounded in your own docs) has a clear, measurable return: fewer tickets handled by hand, faster responses. The same applies to internal work: summarising calls, extracting fields from documents, triaging inbound requests.
If you’re handling more than 50 support tickets a week, this should be on your roadmap. Tools like Intercom and Zendesk already ship AI layers, and for many teams that is the right answer. A custom build starts to make sense when your documentation, product, or workflow is specific enough that a generic layer keeps getting it wrong.
Good if you have data: recommendations and personalization
Recommendation engines work when there is enough behavior to learn from. Netflix, Spotify, and Amazon all run on this, and they all have enormous amounts of data.
If you’re a pre-launch startup with 200 users, a recommendation engine is premature. Focus on getting more users first. Once you have real usage data, personalization can meaningfully improve retention and engagement, and you will know which signals matter.
Expensive but sometimes necessary: custom models
Training your own model costs real compute, months of effort, and a team that knows what it is doing. It only makes sense if AI is the core of your product, not a feature of it. Medical diagnosis tools, fraud detection, and proprietary data analysis are the usual examples. Otherwise, use existing APIs and models.
Fast to build: LLM wrappers
Building on top of OpenAI, Anthropic, or other LLM providers is the fastest path to shipping an AI feature. You can go from idea to working prototype in days.
The catch is that a low barrier to entry means high competition. If your entire product is a thin wrapper around a hosted model, you are one provider update away from obsolescence. Use LLMs to test market fit quickly, then build defensible value on top: proprietary data, unique workflows, domain expertise.
Buy, wrap, or build
Most startups face the same three-way choice. The honest comparison:
| Option | Use it when | The catch |
|---|---|---|
| Buy a product with AI built in | The problem is common and your needs are standard | You inherit its limits, pricing, and roadmap |
| Wrap a hosted model in your own product | You need it specific to your users and want it live in weeks | Easy to copy; quality depends on your prompts, data, and evaluation |
| Build a custom model | AI is the product, or a hard requirement rules out hosted models | Slow, expensive, and needs data and expertise you may not have yet |
The usual path is to buy or wrap first, learn what users actually do with it, and only move down the table when a specific, measured limitation forces you to.
Prove it before you build it
A prototype that impresses in a demo is not evidence that a feature works. Four things separate the two.
Write the evaluation set first. Collect a few dozen real inputs and write down the output you would consider correct for each. Run them automatically whenever the prompt or model changes. This is the single most useful habit in AI work: without it every change is an opinion, and every model upgrade is a gamble. It also gives you a concrete answer to “is it good enough to ship?”
Decide what a wrong answer costs. A wrong product suggestion is a shrug. A wrong summary in a legal or medical workflow is a liability. The cost of failure determines how much checking, human review, and fallback logic the feature needs, and often whether it should exist at all.
Measure cost per request from day one. Hosted models bill per use, and a feature that is cheap at a hundred requests can dominate your margin at a hundred thousand. Log tokens and cost per request now. Route simple requests to smaller models, reserve the larger ones for hard cases, cache repeated work, and cap output length.
Plan for the model changing underneath you. Providers update and retire models. Keep the model name in configuration, keep your evaluation set runnable, and avoid prompts that only work because of one version’s quirks.
Data: the decision you cannot undo later
Where your data goes is an architecture choice, and it is much cheaper to make before the first request than after. Know what the provider retains and whether it trains on your inputs, keep sensitive fields out of prompts where you can, and log what was sent so you can answer a customer’s question about it.
If you handle regulated data, such as health information, the bar is higher: you generally need an agreement with the vendor and a deliberate boundary around what leaves your environment. I wrote up how to think about that in sending PHI to LLM APIs.
The decision framework
Before adding AI to your product, ask four questions:
- Does this solve a real user problem? Not “would this be cool”, but does it solve a pain point your users actually have?
- Can you measure the impact? If you can’t define what success looks like, you can’t justify the investment, and you can’t tell when it breaks.
- What’s the simplest version? Start with an API call and a basic integration. Prove it works before building anything custom.
- What happens when it is wrong? If the answer is “something bad”, design the review step and the fallback before the feature.
Don’t build AI for the sake of AI
The best AI features are invisible. Users don’t care that it’s “AI-powered.” They care that it works, it’s fast, and it saves them time. If a simple rule-based system solves the problem just as well, use that instead.
Thinking about adding AI to your product? Request a technical triage: I’ll help you figure out what’s worth building and what’s not. Learn more about my AI & intelligent automation services, or see what the first ninety days look like on the fractional CTO for AI startups page.