The window for "AI is a novelty" closed a while ago. Enterprises are buying, consumers are paying, and every category has incumbents. That does not mean opportunity is gone — it means you have to be specific. A two-person team cannot out-resource OpenAI or Microsoft, but it can out-focus them. Here is where focus pays off.
1. Vertical SaaS with AI
The most durable AI businesses being built right now are not horizontal AI platforms — they are narrow SaaS products that automate one painful workflow in one industry. Legal document review, medical coding, construction bid analysis, insurance underwriting: these are tasks that eat hours of expensive professional time, follow predictable patterns, and have high enough value per transaction that customers will pay real money to automate them. The category works because domain specificity creates both product fit and a natural barrier to entry.
What a small team can build: pick one workflow you understand deeply, ideally through prior work experience, and build a product that handles the full loop. Not just extraction or summarisation — the review, the flagging, the output in the format the professional actually needs. Medical coding tools, for example, are landing enterprise contracts because they slot directly into billing workflows and reduce claim rejection rates in a measurable way. That measurability is what makes the sale.
The defensibility is not in the model. It is in training data specific to your vertical, fine-tuned classification logic, compliance coverage, and integrations with the legacy systems professionals already use. By the time a competitor gets there, you have months of customer feedback baked in. Horizontal AI tools will never prioritise a single industry's edge cases the way a focused team will.
2. AI Consulting
Demand for people who can actually implement AI — not just talk about it — is outpacing supply. Most medium-sized businesses know they want to use AI and have no idea how to start. They are not going to hire a team of ML engineers; they want a small firm or individual to come in, diagnose where AI can save time, and build something that works. This is a real business right now and the barrier to entry is genuine expertise plus the ability to sell trust.
The trap is staying in project mode forever. Consulting is a great cash-flow business early on, but the ceiling is low if you are billing hours. The path to scale is to turn repeatable solutions into productised offers — a fixed-scope AI audit, a standard RAG deployment package, an automation playbook for a specific industry. Once you have built the same thing three times, templatise it and charge accordingly.
Competition is growing fast: every freelancer who completed an AI course is calling themselves a consultant. The differentiation is depth. Specialise in one sector — legal, manufacturing, real estate, whatever you know — and become the person who has seen every variation of that sector's AI problems. Sector credibility travels through tight networks, and referrals do most of the selling once you have a few good case studies.
3. Content Creation Tools
The content tooling market is enormous and brutally competitive at the top. Jasper, Copy.ai, Runway, ElevenLabs, and a dozen others are well capitalised and moving fast. A small team competing for the same generic "AI writing assistant" space will lose. The opportunity is narrower: tools built for specific content types, specific audiences, or specific platforms that the horizontal players never handle well enough.
What works: AI tools for podcast production (automated show notes, chapter markers, clip selection), real estate listing copy generators that pull from MLS data and local knowledge, technical documentation assistants that maintain style consistency across a whole engineering org, multilingual social content for brands expanding into new markets. These are all narrow enough that the generic tools produce mediocre output and specific enough that paying customers exist.
Defensibility in this space comes from workflow integration, not generation quality. If your tool lives inside the platform your customers already work in — their CMS, their DAM, their podcast hosting dashboard — the switching cost is real. Generation quality between foundation models is converging; the experience around the generation is where you can still differentiate. Build the best workflow, not the best model.
4. Conversational AI
Customer-facing chatbots have had a bad decade of being bad. That is finally changing. The products that are landing in 2026 are not trying to replace a human for every query — they are handling the high-volume, low-complexity interactions that eat support team hours, and escalating cleanly when they cannot. Order status, return policies, account changes, appointment booking: most of this is straightforward and fully automatable. Businesses are paying for it.
Internal knowledge assistants are the less obvious and often easier sell. A law firm with fifteen years of contracts, memos, and precedents sitting in a shared drive; a manufacturing company with procedure manuals nobody reads; a sales team that cannot find the right case study when they need it — these are organisations that will pay for a system that lets employees ask questions in plain language and get real answers. The value is immediate and measurable.
The competitive risk here is that both Salesforce and Microsoft are pushing hard into this space, and their distribution is formidable. You win by going to market through a vertical channel they do not cover, or by offering something they cannot: true white-glove implementation, deep integration with an industry-specific platform, or pricing that works for a 50-person company instead of a 5000-person enterprise.
5. AI-Enhanced Analytics
Most businesses sit on data they cannot query. Their analysts know SQL; their executives do not. The gap between "we have the data" and "we can answer questions about the data" is still enormous, and natural language interfaces to data are a genuine solution. Products like this are landing at companies that previously would have needed a data team to answer a business question. The ROI is easy to demonstrate.
Automated reporting is the adjacent opportunity. Compiling weekly performance decks, monthly board reports, and daily operational summaries is mechanical work that consumes analyst time. Tools that pull from existing data sources, apply business logic, and produce formatted reports on a schedule are not glamorous, but they solve a real, recurring pain point. If you are also handling distribution — email, Slack, PDF — you own a workflow.
The hard part is connecting to the data. Every company's data stack is different: Snowflake, BigQuery, Postgres, Salesforce, spreadsheets, all of the above. Building reliable connectors and handling schema variation is real engineering work. Teams that invest in this infrastructure layer end up with a durable technical advantage over competitors who only built the LLM layer on top.
6. Education and Tutoring
Personalised learning is the long-promised application of technology in education, and AI finally makes it tractable. Not AI that replaces a teacher, but AI that gives every student the equivalent of a patient, infinitely available tutor who adapts to their pace and gaps. Products targeting specific subjects or certification paths — bar exam prep, medical licensing, coding bootcamps, language learning — have paying customers who are highly motivated and can quantify the value of passing an exam.
Assessment and curriculum tools are the B2B side. Schools, bootcamps, and corporate L&D teams need to build, administer, and grade assessments, generate curriculum materials, and track learner progress. This is paperwork-heavy work that AI handles well. A small team building for one of these buyer types — corporate training, K-12 curriculum development, bootcamp instructors — can move fast and charge enterprise prices once they have a handful of institutional customers.
The honest challenge: engagement and retention are hard. Consumer education apps have notoriously high churn because the intrinsic motivation to keep studying fades. B2B customers are stickier. If you are building for consumers, your product needs to be genuinely better at keeping people engaged than the apps they already have, and that requires serious product work beyond the AI layer.
7. Developer Tools
The tooling around AI development is still immature and filling fast. Prompt management, model evaluation, observability, deployment infrastructure, cost monitoring — every team building AI products needs these things, and most are either building them in-house or stitching together multiple incomplete tools. A focused product that solves one of these problems cleanly will find willing buyers among the growing population of AI-first engineering teams.
Model evaluation and testing is an underserved area with high demand. As AI products move into production, teams need systematic ways to know whether a change to their prompt or model is an improvement or a regression. This is not a solved problem. Tools that let you define evaluation criteria, run your prompts against test sets, and track quality over time fill a gap that matters more as AI systems become load-bearing. Companies that have already shipped AI to customers will pay for this.
The risk in developer tools is that the platform providers — OpenAI, Anthropic, AWS, Azure — may ship native versions of whatever you build. Build in a way that is model-agnostic and integrates with multiple providers, so you are not dependent on one platform's product roadmap. If your tool works equally well whether the team is on GPT-5 or Claude 4, you are harder to displace.
8. E-commerce AI
E-commerce operations involve enormous amounts of repetitive, structured work: writing product listings, setting prices, personalising recommendations, generating ad copy, handling customer questions. AI handles all of these and the ROI is measurable in conversion rates, return rates, and hours saved. Large platforms like Shopify and Amazon have their own AI features, but their one-size-fits-all tools leave room for products that go deeper for specific merchant types.
Auto-generated product listings are the immediate opportunity: a merchant importing a thousand SKUs from a supplier does not want to write a thousand descriptions. AI that takes a product data feed and produces SEO-optimised, brand-consistent copy at scale saves real time and real money. Dynamic pricing tools that adjust to inventory levels, competitor prices, and demand signals are landing at mid-market retailers who cannot afford the enterprise solutions but need the capability.
Defensibility comes from integrations and merchant trust. If your tool is embedded in Shopify's admin, or connects directly to a specific ERP system, or handles a specific product category — fashion, electronics, grocery — with appropriate domain knowledge, you are hard to replace. Merchants who have connected their catalog and trained the system on their brand voice are not going to switch for a marginally better generation model.
9. AI in Finance
Financial services are drowning in compliance, documentation, and risk assessment work that is rule-based enough for AI but complex enough that it consumes expensive analyst and legal time. AML transaction monitoring, KYC document processing, loan underwriting, regulatory change monitoring — these are areas where automation has high business value and where the regulatory environment actually creates a moat: you have to understand compliance deeply to build in this space, and that knowledge is not easily replicated.
Fraud detection deserves a dedicated mention. The pattern-recognition task of identifying fraudulent transactions, synthetic identities, and account takeovers is exactly what modern ML handles well. There are strong incumbents here, but the landscape is fragmented by geography, payment type, and industry vertical. A tool tuned for a specific fraud surface — marketplace seller fraud, insurance claim manipulation, crypto wash trading — can outperform a general solution on the metrics that matter to that buyer.
The regulatory environment is the highest barrier to entry and the strongest moat once crossed. Building AI products that touch financial data requires SOC 2, data residency compliance, audit trails, and in some jurisdictions, model explainability. Teams that build this infrastructure properly can charge accordingly and face fewer competitors willing to do the same work. If the compliance overhead is your main concern, target the infrastructure buyers — banks and fintechs who will pay for compliant AI components they can embed in their own systems.
Strategic Advice: How to Actually Win
The model is not the product. In 2024 you could build a passable business by wrapping GPT-4 before most customers had access to it. That window is gone. Every category above has competitors already doing the AI part. What separates durable businesses from feature demos is distribution, workflow ownership, and domain depth. Get those three right and the underlying model is a commodity input.
Distribution comes first. A better product with no distribution loses to a worse product with a sales channel, a community, or a platform integration. Before you write a line of code, figure out how you will reach ten paying customers. If you cannot explain that clearly, the product is not ready to build. In practice this means going to market through an existing channel — a vertical community, a partner platform, an industry event — rather than trying to build an audience from zero.
Go deep in one domain rather than broad across many. The temptation to add more verticals, more use cases, more integrations is real and nearly always wrong in the first two years. Customers in a specific industry talk to each other. If you solve the problem better than anyone else for, say, independent insurance adjusters, you will get referrals from every independent insurance adjuster who sees the demo. That network is more valuable than marginal feature breadth.
Own the workflow, not just the AI layer. The businesses with pricing power are the ones whose product is the system of record — the tool where work gets done, where outputs live, where the team logs in every day. If you are a feature that sits on top of another tool, you are always one platform update away from being deprecated. If your product is where the work happens, you own the relationship. Build toward that.
Charge for outcomes, not tokens. Token-based pricing made sense when AI was a novelty; customers are now sophisticated enough to understand that they are buying time saved, errors reduced, or revenue generated. Price to that outcome. A tool that saves a paralegal three hours a week on document review is worth hundreds of dollars a month, regardless of what it costs to run the model. Outcome pricing aligns your incentives with your customers' and creates the margin you need to build a real business.