The most important AI decision a founder makes is often the one before any code is written: should this product have AI in it at all? Market pressure, investor expectations, and the speed at which AI tooling has improved all push toward adding it. The honest answer is sometimes yes, sometimes no, and occasionally a much narrower yes than the initial instinct suggests. This framework helps you distinguish between those cases before you commit budget to a build.
The real question to answer first
Before evaluating models, providers, or architecture, ask a simpler question: does AI remove a genuine problem for the user, or does it add capability that nobody asked for?
The most reliable signal is workflow friction. Look at the parts of your product — or the manual process it replaces — where users currently hesitate, stall, export data into a spreadsheet, or depend heavily on one knowledgeable person. Those points of friction, interpretation, and decision fatigue are where AI tends to create real value. If you cannot identify at least one such point, the case for AI is probably marketing, not product.
A second, harder question follows immediately: does the problem need intelligence, or does it need simplification? Some workflows that feel like AI candidates are actually just poorly designed. Adding AI to a confusing flow makes it a confusing flow with a chatbot on top. If the process should be eliminated or redesigned, AI will not save it — it will obscure it.
Four genuine use cases where AI earns its place
There are four patterns where AI consistently adds real product value rather than superficial novelty.
Personalisation at scale
When a product needs to adapt its behaviour to each user — recommendations, content sequencing, dynamic pricing, or tailored guidance — and the user base is large enough that manual curation is impossible, AI creates a genuine structural advantage. The keyword is scale: personalisation for fifty users is a filter, not an AI problem.
Unstructured data interpretation
Documents, emails, contracts, call transcripts, images, and free-text inputs carry meaning that traditional software cannot extract without significant manual work. AI is genuinely good at interpreting this class of data: classifying it, summarising it, extracting structured fields from it, and surfacing the parts that matter. A product in legal, finance, property, or operations that currently makes users read through long documents to find a number has a legitimate AI use case.
Automation of expert judgment
Some decisions require knowledge that most users lack — not just information retrieval, but interpretation. Waaza, Wall & Fifth's marine financing intelligence product, is a working example: it takes the ambiguity of vessel finance enquiries and turns them into structured, explainable paths, doing work that would otherwise require a broker conversation. The AI replaces a dependency on expert access, not just a manual task.
Conversational interfaces to complex information
When a product holds a large, structured knowledge base and users need answers that span multiple records or require synthesis — compliance requirements, product specifications, policy interpretation — a well-grounded conversational interface can genuinely outperform a search box. The critical word is grounded: the AI needs to answer from your data, not from general training. Without retrieval, this use case collapses into confident guessing. For what the grounding layer actually involves technically, the practical guide to LLM integration covers retrieval, prompt construction, and response handling end to end.
Three additions that rarely justify the engineering
Equally important is knowing where AI is routinely oversold. These three additions appear constantly in briefs and rarely deliver proportionate value.
Chatbots on simple FAQ content
A chatbot built on a shallow FAQ creates engineering cost, latency, and hallucination risk for a problem that a well-structured help page solves faster and more reliably. If the content fits on ten pages and users have a clear question, a search bar and good information architecture beat a conversational layer every time. The chatbot adds a layer of apparent sophistication over a content problem that has a simpler fix.
AI-generated product or listing descriptions
Generating descriptions at scale feels like a compelling use case until the outputs start sounding identical. For commodity content with no proprietary signal — product names, dimensions, categories — AI generation saves a little time and produces generic copy. For content where voice, specificity, and trust matter (premium goods, professional services, property), the output usually needs editing that erodes the time saving. This is a reasonable productivity tool for internal use, not a differentiated product feature.
Recommendation engines on thin catalogues
Recommendation systems require enough historical interaction data to surface meaningful patterns. A catalogue of three hundred items with six months of usage data is not a recommendation problem — it is a curation and merchandising problem. An AI recommendation layer on a thin catalogue often produces results that are random in practice and sophisticated-sounding in the pitch. Solve the merchandising problem first; a recommendation engine can come later if the data justifies it.
How to test whether AI belongs in your product
Before briefing a studio, answer three questions concretely. Vague answers are a signal to pause.
- Which specific user workflow does AI improve, and what does the improvement look like? Name the workflow, the current friction, and the measurable change the AI would produce. If you cannot describe this without using the word "smarter", the use case is not defined yet.
- What accuracy level is good enough to be better than the current process? An AI feature does not need to be perfect — it needs to outperform the alternative. A model that classifies documents correctly 85% of the time may still be far better than a manual process with a high error rate. Define the threshold before the build, not after.
- Do you have, or can you create, the data the AI needs to be useful? General-purpose models answer general questions. If your product requires domain-specific answers, you need proprietary data for retrieval, fine-tuning, or grounding. A model without a data layer is a starting point, not a product. For more on the difference between a genuine AI product and a thin wrapper, the guide to building defensible AI products covers the distinction in detail.
The cost and complexity premium — and when it is worth paying
AI adds cost at three points that standard software does not: inference (the ongoing cost of calling a model on every relevant user action), the retrieval and data infrastructure that makes the model useful, and the evaluation work required to catch regressions when model behaviour changes.
| Factor | Standard software | AI-powered software |
|---|---|---|
| Compute cost per user action | Minimal, scales predictably | Inference cost per call; must be modelled against revenue |
| Data requirements | Structured inputs, defined schema | Needs grounding data; quality and coverage matter |
| Output reliability | Deterministic for a given input | Probabilistic; requires testing, evaluation and guardrails |
| Model maintenance | No equivalent | Model updates can change behaviour; regression testing needed |
| Build complexity | Lower for equivalent features | Higher; integration, prompt design, retrieval stack, response handling |
The premium is worth paying when AI creates a structural advantage that standard software cannot replicate: when it interprets data no rule set could handle, automates judgment that would otherwise require an expert, or delivers personalisation that a static system cannot approach. It is not worth paying when the feature could be implemented as a filter, a search index, or a well-designed static interface.
What a defensible AI feature looks like
An AI feature is defensible when a competitor cannot reproduce it by calling the same API. The differentiator is almost never the model itself — it is what surrounds the model.
Defensibility comes from one or more of: proprietary data that cannot be replicated (transaction history, domain-specific documents, user behaviour over time); retrieval logic tuned to a specific domain; workflows that embed AI into a user journey in a way that is genuinely hard to reverse-engineer; or a data flywheel where each new user makes the system more useful for the next one.
A feature that a competitor can copy in a week by adding the same model to a similar interface is not a product strategy. It is a demo.
Hypothetical example: two property platforms
Consider two property platforms — both adding AI-assisted search (this is a hypothetical illustration). Platform A adds a natural language search bar that calls a general model and reranks results. Platform B builds a retrieval layer over five years of their own lettings data, including rejected applications and landlord preferences, so that the AI can surface listings matching nuanced criteria no filter could express. Both use similar underlying models. Platform A's feature can be copied in days. Platform B's cannot be copied without years of proprietary data.
How to brief a studio on an AI requirement
When the decision to build AI is sound, the brief that reaches a studio should specify the workflow being changed, not the technology being used. Lead with the user problem and the current friction, not with the model or the architecture. A good studio will challenge both the feature and the approach — that challenge is part of the value.
Leave open the questions of model choice, retrieval architecture, and exact output format. Those are implementation decisions best made once the workflow is understood. Fix the requirement: the specific journey, the input, the expected output, and the definition of success. A brief that specifies a model and a retrieval architecture before defining the user problem is specifying a solution to an unconfirmed question.
For founders at the stage of deciding what to build before any studio conversation, the Wall & Fifth MVP service includes scoping as a first step — the goal is to define the right thing to build before fixing a price, which applies equally to AI features as to any other product decision.
Frequently asked questions
Can I add AI to my product after launch rather than building it in from the start?
Yes, and in many cases this is the right order. Launching without AI, gathering real usage data, and then identifying the specific workflows where AI would create measurable improvement is a more grounded approach than speculating before any users exist. The risk of building AI in from day one is that you are solving a workflow problem you have not yet observed at scale. The exception is when AI is the core product premise — in that case, shipping without it is shipping something different.
What data do I need before building an AI feature?
The honest answer depends on the feature. For a retrieval-augmented assistant, you need a proprietary document or data corpus worth retrieving from — the size matters less than the quality and relevance. For a recommendation system, you need sufficient interaction history to surface non-obvious patterns; a thin catalogue with sparse data will not produce useful recommendations. For classification tasks, many work well with a well-prompted general model and no training data at all, though fine-tuning on labelled examples improves precision for domain-specific classifications.
Is an AI feature a risk to my product if the model behaves unexpectedly?
Yes, and this risk is underestimated in early-stage planning. A model update, a change in the API's behaviour, or an edge-case input can change outputs in ways that affect users. Production AI products need evaluation pipelines — ways to test whether outputs have changed meaningfully when something upstream shifts. This is one of the meaningful differences between a demo and a shipped product. Building with a studio experienced in production AI rather than proof-of-concept work reduces, but does not eliminate, this risk.
How do inference costs affect a SaaS pricing model?
Inference cost — what you pay the model provider for each API call — scales with usage in a way that fixed-cost SaaS does not. A user who triggers ten AI calls per session costs ten times more to serve on that dimension than a user who triggers one, regardless of what they pay. This creates a meaningful unit economics question: at your planned subscription price, at what usage level does the margin invert? This calculation must be done before the build, not after the first billing cycle. Usage caps, tiered plans, and per-call metering are all structural responses to this problem.
Should I build my AI feature or use an off-the-shelf product?
The useful question is not build versus buy in isolation — it is whether AI is your competitive advantage or a feature. If the AI capability is what makes customers choose your product specifically, it should be built with enough custom logic and proprietary data to be defensible. If AI is table stakes for the category and your differentiation lies elsewhere, a well-integrated commercial product may be faster, cheaper, and sufficient. Many products end up in a sensible middle: buying commodity infrastructure (transcription, embeddings, vector search) and building the domain-specific layer on top.
When is the right time to start building AI into a product?
Readiness has two dimensions: the clarity of the use case and the maturity of the underlying data. If you have identified a specific, high-friction workflow and have proprietary data that would make AI genuinely better than the manual alternative, the right time is now — because a competitor with those same assets who moves first creates a data advantage that compounds over time. If either element is missing, the right time is after you have addressed it, not before. Shipping AI prematurely produces a product that is expensive to maintain and easy for a better-prepared competitor to surpass.
Back to Building AI Products: A Practical Guide for Founders Who Want More Than a Wrapper