This chapter offers a pragmatic checklist to decide when AI will help and when it will hinder. It organizes seven questions across three situations: building an AI-based product, using conversational AI as a personal assistant, and constructing software that relies on large language models under the hood. The guidance pushes beyond hype, urging a customer-centric lens, a sober look at failure modes like hallucinations and opacity, and a fit-for-purpose mindset that matches tools to problems—especially around whether natural language is truly central to the task.
For product decisions, the first test is whether you are putting AI before customer needs; pursuing trendy tech without a clear problem risks poor product–market fit. Next, assess whether hallucinations are acceptable for the use case; low-stakes contexts can tolerate occasional errors, while high-stakes or transactional settings cannot—sometimes reframing the tool as assistive rather than authoritative helps. Finally, understand how much explainability stakeholders require; when clarity matters, simpler models or partial explainability techniques may be preferable, acknowledging that state-of-the-art deep models remain largely black boxes.
For using conversational AI as an assistant, AI shines when tasks can be described succinctly and outputs can be validated easily; otherwise, crafting prompts and checking results may cost as much as doing the work yourself. Ask whether someone has already solved the problem with a well-tested library or tool, which can be faster and more reliable than prompting from scratch. Also clarify what “excellent” looks like for the task—expert standards in writing or design, for instance, may exceed what generic AI can deliver. When building LLM-centered “wrappers,” reserve this approach for products where users actually interact through natural language; if not, traditional APIs and algorithms often yield faster, more accurate, and structured results. Overall, use AI deliberately, guided by customer value, tolerance for error, explainability needs, task structure, available reuse, and the role of natural language.
Summary
Adopt an entrepreneurship mindset: Make sure you understand the customer’s problem and find the best way to solve it, which may or may not be AI. Try not to decide you’ll use AI before knowing what you’ll use it for.
Evaluate whether hallucinations are okay for your users. In some cases, such as the translation of hotel reviews, users are not too fussy about them. If hallucinations are a deal breaker, you may need to rethink the task.
If your client wants to understand how AI generates its outputs, you may find yourself in quite a pickle, as the most powerful AI cannot be understood in detail. You might need to use a less performant but more explainable model or a technique to try to explain some aspects of the model such as the finding the relative importance of input features.
If you use AI as an assistant, you can quickly ask yourself, “Can I easily describe the task and validate the output?” If so, you get good bang for your prompt, which is when AI is most effective. If you can’t describe the task succinctly or it takes a lot of effort to validate the output, using AI may not be very useful.
When using AI as an assistant, make sure you’re not trying to reinvent the wheel. If you’re performing a common task, perhaps there’s a well-documented and proven way of doing the job. For example, when coding a certain task, there may already exist an open-source software that does exactly what you need, and it might be easier and more reliable to use it instead of trying to do the same using an AI assistant.
Before having AI do a certain job, try to understand what doing that job in an excellent way entails. It might be more complicated than it seems. You can then decide whether an AI assistant will help you do the job to a satisfactory level depending on your needs.
Some people are building apps which are LLM wrappers—the job is done by an LLM under the hood, which is managed by the app. This is most powerful when you’re dealing with a natural-language problem, either because the user inputs free-form text or the system must generate humanly readable text. If your problem doesn’t involve natural language, maybe using old-school algorithms instead of building LLM wrappers will be a better choice.
FAQ
What’s the quick checklist to decide whether AI fits my use case?Think through seven prompts: For products—1) Are you solving a real customer need (not AI-first)? 2) Are hallucinations acceptable or manageable? 3) Do stakeholders require explainable outputs? For using conversational AI as an assistant—4) Can you describe the task succinctly? 5) Can you validate outputs easily? 6) Has someone already built a reliable tool or library for this? 7) What does “excellent” actually look like for the task? For LLM wrappers—8) Will users interact with your product via natural language? If yes across the relevant items, AI is likelier to help; if not, consider alternatives.What does “putting AI before the customer” mean, and why is it risky?It’s starting with “we need AI” and then hunting for a use case—driven by hype, FOMO, or fundraising optics—rather than beginning with customer problems. This often yields solutions nobody needs, weak product‑market fit, and wasted effort. Adopt a customer‑centric, lean approach: validate the problem with quick prototypes and only choose AI if it’s the best way to solve a validated need.When are AI hallucinations acceptable—and when are they deal‑breakers?They’re acceptable if users value breadth or speed over perfect accuracy and can tolerate occasional errors (e.g., auto‑translated reviews). They’re a deal‑breaker when errors damage trust, safety, cost, or compliance (e.g., billing or critical operations). You can sometimes reframe: position AI as a “second pair of eyes” or draft generator, not an authoritative source, so users expect to verify outputs.What if my stakeholders need to understand how outputs are generated?If explainability is necessary, consider simpler, more interpretable models (even at some accuracy cost) or reframe the work as analysis rather than prediction. Post‑hoc XAI tools (e.g., feature‑importance methods like SHAP) can provide partial insight, but won’t fully “open the black box.” Set expectations early about what can and can’t be explained.When does conversational AI give “good bang for my prompt”?When you can specify the task briefly and still get a useful output, and when you can quickly check whether the result is correct. Examples include summarizing code, small utility functions, refactors, or language conversions. If the prompt requires exhaustive, step‑by‑step specs or complex domain rules, the prompting overhead can rival doing the task directly.How can I validate AI outputs efficiently to reduce risk?Favor tasks where you already have expertise, quick tests, or clear acceptance criteria—so you can spot issues fast (e.g., run the code, compare against known results, lint, unit test). If validation is slow, ambiguous, or outside your skill set (e.g., unfamiliar language or intricate business logic), the time you save generating may be lost verifying and debugging.Should I use AI or search for an existing tool/library first?Check whether the task is common and likely solved already. Mature libraries and frameworks are usually faster, more reliable, and community‑tested. AI can produce verbose, hard‑to‑verify code and may hallucinate APIs. A quick search plus reading concise docs often beats prompting for cookie‑cutter tasks.When is it a good idea to build an “LLM wrapper” app?Primarily when users will interact via natural language (free‑form input and/or prose output). LLMs shine at interpreting and producing text. If your app has structured inputs/outputs, traditional pipelines (APIs, databases, algorithms) often outperform LLMs on accuracy, latency, and reliability, and avoid issues like outdated or malformed responses.How do I set the quality bar before using AI on creative or complex tasks?Research what “excellent” entails—structure, pacing, originality, competitive context, domain nuances. For writing or design, quality hinges on more than grammar or visuals; it involves storytelling, positioning, and differentiation. Decide whether an AI‑assisted draft is “good enough” for your stage, or whether the task warrants specialist expertise.What should I do if leadership wants “something with AI” but the use case is unclear?Pause and run a customer‑first discovery: clarify target users, pains, and success metrics; prototype the simplest solution (AI or not) to validate demand; test acceptance of errors and explainability needs; and only then choose AI if it clearly improves outcomes. This avoids the “AI hammer” trap and increases your odds of shipping something valuable.
pro $24.99 per month
access to all Manning books, MEAPs, liveVideos, liveProjects, and audiobooks!
The AI Pocket Book ebook for free
