The AI business that fails most reliably is the one that was easiest to start.
Wrap a foundation model API in a clean interface, charge a monthly subscription, and you have a product in a weekend. You also have a product with no moat, whose core capability belongs to someone else, competing against a model provider who can ship your entire feature set as a checkbox in their next release. Many did exactly that, and many of those businesses no longer exist.
The failure was not the execution. It was the structure. The value sat in the model, the model was rented, and the landlord was also a competitor.
Understanding why that structure fails is the whole basis for choosing a better one. The businesses on this list share a property:
the hard part is something a model provider cannot easily acquire.
Usually that means permissioned data, regulatory burden, workflow depth, or relationships. AI is the engine. It is almost never the moat.
A word on epistemic status before the list. Everything here about current market structure is inference from observable patterns. Everything about the future is judgement, not prediction. Private company revenue figures circulate widely and are reported rather than audited, so where they appear they are labelled as reported. Anyone who tells you with confidence which AI businesses will win in five years is guessing with more conviction than the evidence supports, including this article. What follows is reasoning you can inspect and disagree with.
How These Were Evaluated
Each opportunity was assessed against the same criteria: the severity of the customer’s pain, demonstrated willingness to pay, market size, competitive density, technical difficulty, defensibility, revenue model quality, customer acquisition difficulty, dependence on third-party models, regulatory exposure, operational complexity, likely margins, whether AI is genuinely necessary, and the risk of commoditisation.
Two of those deserve emphasis because they eliminate most ideas immediately.
Whether AI is genuinely necessary.
If a deterministic system solves the problem more reliably, AI is a liability rather than a feature. Adding a language model to a task with fully specifiable rules adds cost, latency, and a failure mode that did not previously exist.
Whether the opportunity survives the model providers improving.
Assume the underlying models get significantly better and cheaper.
Businesses whose value proposition is “we make the model usable” get compressed by that. Businesses whose value proposition is “we have the data, permissions, and workflow integration the model needs” get stronger.
The Ranking
1. Vertical AI for a regulated professional workflow
What it sells Software that performs a specific, high-value professional task inside one industry, built around that industry’s actual rules, formats, and compliance requirements.
Who buys Firms in legal, healthcare administration, insurance, accounting, or financial services where the work is expensive, rule-bound, and currently done by qualified humans.
The problem it solves Highly paid professionals spend large fractions of their time on structured document work. The cost is enormous and the work is genuinely repetitive, but it cannot be handled by generic tools because the rules are specific, the formats are idiosyncratic, and the consequences of errors are severe.
Revenue Subscription with usage components. Increasingly, outcome-linked pricing appears in this category, which is possible precisely because the ROI is measurable.
MVP One document type, one jurisdiction, one firm size. Deliberately narrow. A tool that handles one specific filing correctly for one segment beats a platform that handles everything approximately.
Skills Deep domain expertise is the binding constraint, not engineering. The best founders here are usually practitioners who learned to build, or technical founders in genuine partnership with practitioners. Domain knowledge cannot be researched quickly enough to substitute.
Risks Regulatory exposure is real and varies by jurisdiction. Professional liability requires clear positioning of the tool as assistive. Sales cycles in these industries are long and conservative.
Defensibility Strong and compounding. Permissioned access to a firm’s document archive, accumulated corrections, and embedded compliance knowledge are things a general model provider cannot replicate without the relationships. Switching costs grow as the tool becomes embedded in workflow. Competitiveness: High in the largest verticals, where well-funded companies have reportedly reached substantial revenue. Considerably lower in narrower niches, which is where a small team should look.
Suitable for: Small team with genuine domain expertise, or funded startup. Difficult for a solo generalist founder.
2. AI implementation and evaluation consulting
What it sells Assessment, design, and building of AI systems for businesses that have budget and no internal capability, with a specific focus on making systems measurable.
Who buys Mid-market companies, roughly 50 to 1,000 employees, with real process costs and no AI team.
The problem it solves The gap between wanting AI and deploying it usefully is enormous, and it is mostly not a modelling problem. Gartner predicted in June 2025 that over 40% of agentic AI projects would be cancelled by end of 2027, citing costs, unclear value, and inadequate risk controls. McKinsey’s State of AI survey found that in any given business function, no more than 10% of organisations reported scaling agents. That gap is a services market.
Revenue Project fees, then retained monitoring and maintenance. The retained portion matters, because AI systems degrade silently and need ongoing evaluation. This is also what turns a consultancy from a treadmill into a business with recurring revenue.
MVP One paid engagement. Genuinely, that is the entire MVP. Do it, document the outcome, repeat in the same industry.
Skills Business process analysis first, technical implementation second. The differentiating skill is knowing which processes should not be automated, because that judgement is what clients cannot get from vendors.
Risks Services businesses do not scale like software. Client concentration is dangerous. The market will get more competitive as capability spreads.
Defensibility Weak structurally, moderate in practice through reputation and specialisation. Choose one industry and become the obvious choice within it. Competitiveness: Crowded at the generic end, thin at the specialist end. “AI consultant” is saturated. “The person who automates claims processing for regional insurance brokers” is not.
Suitable for: Solo founders and small teams. This is the most realistic starting point for most people reading this , and it has an underrated property: it generates the domain knowledge and the customer relationships that make opportunity #1 possible later. Several strong vertical software businesses started as services businesses that noticed the same problem repeatedly.
3. AI-enabled business process outsourcing
What it sells The completed work, not the software. The customer buys an outcome, such as processed invoices or qualified leads, and never sees the tooling.
Who buys Businesses currently paying for outsourced back-office work or employing people to do it.
The problem it solves Most companies want the output, not a platform to administer. Selling completed work removes the adoption problem entirely, which is the single largest cause of software failure in mid-market businesses.
Revenue Per unit processed or monthly retainer. Margins improve as automation deepens, which is the structural advantage: you can start mostly human and automate incrementally without the customer noticing or caring.
MVP Do the work manually for the first few clients. Automate the parts that repeat. This has the lowest risk profile of anything on this list because you validate demand before building anything.
Skills Operational discipline above all. Quality control matters more than engineering elegance.
Risks People-dependent early, which caps growth. Margin compression if competitors automate faster. Requires genuine operational rigour, which is a different temperament from product building.
Defensibility Moderate. Switching costs are real once you hold the process. Accumulated exception-handling knowledge becomes a genuine asset. Competitiveness: Moderate, and the incumbent outsourcing firms are slower to adapt than they should be.
Suitable for: Solo founders and small teams, particularly those with operational rather than engineering backgrounds. Underrated relative to the attention it receives.
4. AI for compliance and regulatory operations
What it sells Systems that monitor obligations, detect breaches, prepare documentation, and maintain audit trails within a specific regulatory regime.
Who buys Regulated businesses in finance, healthcare, data protection, employment, and increasingly AI governance itself.
The problem it solves Compliance is expensive, mandatory, and growing. Failures carry fines rather than inconvenience. The work is largely reading documents against rules, which is well suited to the technology, and the budget already exists as a cost line rather than needing to be created.
Revenue Annual subscription. Retention in this category is unusually strong because switching a compliance system mid-cycle is genuinely painful.
MVP One regulation, one industry, one company size.
Skills Regulatory expertise is essential and not researchable at the required depth. Partnership with a compliance professional is usually necessary.
Risks Regulations change and you must track them. Liability positioning must be careful: assistive, never advisory. Long enterprise sales cycles.
Defensibility Strong. Encoded regulatory knowledge, audit history, and the sheer inconvenience of switching create durable retention. Competitiveness: Moderate and rising, but fragmented by jurisdiction and regime, which leaves genuine gaps.
Suitable for: Small team with regulatory expertise. Not a solo generalist opportunity.
5. AI voice agents for appointment-driven local businesses
What it sells A system answering calls, booking appointments, and handling routine enquiries for businesses that lose revenue to unanswered phones.
Who buys Clinics, trades, veterinary practices, salons, restaurants, property managers. Any business where the phone rings while everyone is busy.
The problem it solves Missed calls are lost revenue with a directly calculable value. A dental practice knows what a new patient is worth. That makes the sales conversation unusually easy, because the ROI arithmetic is obvious to the buyer.
Revenue Monthly subscription plus usage, typically at price points a small business can approve without a committee.
MVP One vertical, one language, one region. The vertical focus matters more than it appears, because the value is in handling that industry’s specific enquiries correctly.
Skills Integration engineering, particularly with booking systems, plus telephony familiarity. Sales capability matters as much as technical skill because this is a high-volume, low-value-per-customer motion.
Risks Voice quality expectations are unforgiving. Booking errors are highly visible. The platform layer is commoditising quickly, which pushes value toward vertical specialisation and distribution.
Defensibility Weak at the technology layer, moderate through vertical integrations and distribution relationships. Partnering with practice management software vendors is often the real strategy. Competitiveness: High and rising. Differentiation must come from vertical depth, not from the voice technology itself.
Suitable for: Solo technical founders and small teams. Reachable, but do not expect the technology to be the advantage.
6. AI evaluation and monitoring infrastructure
What it sells Tooling that tests AI systems against known cases, tracks quality over time, detects regressions, and alerts when behaviour changes.
Who buys Companies running AI in production, which is a growing but still limited population.
The problem it solves AI systems degrade silently. They produce fluent, plausible, wrong output that looks identical to correct output. Most organisations deploying AI have no systematic way to detect this, which is a direct cause of the failure rates the industry reports.
Revenue Subscription, usage-based, or as part of a services engagement.
MVP Tooling built for your own consulting clients, then productised. This is a natural extension of opportunity #2 rather than a standalone start.
Skills Genuinely technical. Requires understanding evaluation methodology, which is a specialist discipline.
Risks The main one is timing. The buyer population is smaller than the discourse suggests, and model providers and cloud platforms are building similar capabilities into their offerings.
Defensibility Moderate. Accumulated evaluation data and integration depth help. The category could be absorbed by platform vendors. Competitiveness: Increasingly crowded at the general end. Domain-specific evaluation, such as evaluating clinical or legal outputs correctly, is much thinner.
Suitable for: Technical founders. Ranked lower than its strategic importance because the market is early and the buyer population is thin. Being right too soon is a common way to fail.
7. AI-powered knowledge management for professional firms
What it sells Systems that make an organisation’s accumulated knowledge retrievable, with citations to source documents.
Who buys Professional services firms, engineering companies, and any organisation where expertise lives in documents and people’s heads.
The problem it solves Expertise is trapped. Employees ask colleagues questions answered in documents nobody can find, which costs two people’s time instead of one. Firms lose institutional knowledge when people leave.
Revenue Per-seat subscription, which is a well-understood buying motion in these firms.
MVP One firm, one knowledge domain, done properly with citations working reliably.
Skills Retrieval engineering, which is more difficult than it appears. Permission handling is critical and frequently botched: a retrieval system that ignores the underlying access controls will eventually surface something it should not, and in a professional firm that is a serious incident.
Risks Data quality determines everything. Firms with disorganised documentation get poor results and blame the tool. Answering from superseded documents is the characteristic failure.
Defensibility Moderate. Integration depth and accumulated tuning create switching costs. General-purpose enterprise search vendors compete directly. Competitiveness: Moderate and rising.
Suitable for: Small technical teams with patience for the unglamorous data work that determines success.
8. AI sales operations for mid-market B2B
What it sells Systems that research accounts, prepare briefings, maintain CRM hygiene, and draft communications, sold as sales productivity rather than as AI.
Who buys B2B companies with sales teams of roughly 5 to 50 people, large enough to feel the inefficiency and too small for internal tooling.
The problem it solves Sales staff spend a large minority of their time on research and administration rather than selling. The cost is directly calculable from salary, and the buyer is usually a sales leader with budget authority and no procurement committee.
Revenue Per-seat subscription.
MVP Pre-call briefing generation for one CRM platform. Narrow, immediately useful, and easy to demonstrate.
Skills Integration engineering plus real understanding of sales process.
Risks Serious commoditisation risk. CRM vendors are building these features natively, and a feature bundled free with the system of record beats a better standalone product most of the time. Depth in one industry’s sales process is the only durable answer.
Defensibility Weak to moderate. This is the ranking’s main constraint. Competitiveness: Very high.
Suitable for: Technical founders with sales domain knowledge, ideally targeting an industry vertical rather than sales teams generally.
9. Industry-specific AI training and enablement
What it sells Structured programmes teaching a specific profession to use AI effectively for their actual work, delivered as courses, workshops, or ongoing programmes.
Who buys Professional bodies, mid-sized firms, and individual professionals.
The problem it solves Generic AI training is abundant and mostly useless, because the gap is not “how do I prompt” but “how does this apply to my specific work, and where is it dangerous.” Research on enterprise AI consistently finds employees using unsanctioned tools while official initiatives stall, which is a training and governance failure as much as a technology one.
Revenue Course fees, corporate contracts, or subscription for continuously updated material.
MVP One workshop for one professional group, refined through delivery.
Skills Teaching ability and credibility within the profession. The credibility requirement is absolute: professionals will not learn from someone who does not understand their work.
Risks Content decays fast. Material written eighteen months ago is often actively misleading. This requires continuous rewriting, which is the hidden operating cost. Low barriers to entry mean crowding.
Defensibility Weak, resting almost entirely on reputation and relationships. Competitiveness: Very high in general AI training. Much lower where the trainer has genuine professional standing in a specific field.
Suitable for: Solo founders with existing professional credibility. The best non-technical opportunity on this list , and a viable route to building the relationships that lead to opportunities #1 or #2.
10. AI for a physical or field-based industry
What it sells Systems handling documentation, scheduling, quoting, and compliance for construction, logistics, manufacturing, agriculture, or field services.
Who buys Operators in industries with heavy paperwork burdens and historically poor software.
The problem it solves These industries carry substantial administrative overhead and have been underserved by software because they are difficult to sell to and operationally messy. The underserving is the opportunity.
Revenue Subscription, often with per-site or per-vehicle pricing.
MVP One document workflow, such as compliance reporting or quote generation, for one trade.
Skills Domain access matters more than technology. Without relationships in the industry, customer acquisition is very difficult.
Risks Slow sales cycles, price sensitivity, low technology adoption, and buyers who are legitimately sceptical of software vendors. Field conditions impose real constraints, including poor connectivity.
Defensibility Moderate to strong once embedded, because these customers switch software rarely. Competitiveness: Low, which is the attraction. The barrier is access, not technology.
Suitable for: Founders with existing industry relationships. Ranked last on general accessibility rather than on opportunity quality. For someone with a background in one of these industries, it may be the best option on the entire list.
The Verdict
| If this is you | Start here |
|---|---|
| Best for beginners | AI-enabled business process outsourcing (#3). You validate demand by doing the work manually before building anything, which eliminates the most common way these businesses fail. |
| Best for non-technical founders | Industry-specific training and enablement (#9), with implementation consulting (#2) close behind if you can partner with a technical builder. Both trade on domain credibility rather than engineering. |
| Best for developers | Vertical AI for a regulated professional workflow (#1), provided you find a genuine domain partner. Evaluation infrastructure (#6) is intellectually the most interesting and commercially the earliest. |
| Best for recurring revenue | Compliance and regulatory operations (#4). Mandatory spending, painful switching, annual contracts. |
| Best for high-ticket services | Implementation consulting (#2), specialised into one industry. |
| Best long-term SaaS opportunity | Vertical AI for a regulated professional workflow (#1). It has the clearest path to durable defensibility, because permissioned data and embedded compliance knowledge are assets model providers cannot buy at any price without the customer relationships. |
| Most overhyped and best avoided | The general-purpose “AI agency” selling chatbots and automation to any business that will listen. No specialisation, no defensibility, no accumulating asset, competing on price against everyone with the same idea. Gartner’s observation about “agent washing,” and its estimate that only around 130 of thousands of vendors positioning themselves as agentic were genuinely so, describes a market where undifferentiated positioning is the norm. The identical work, focused on one industry, becomes a real business. The specialisation is the entire difference. |
The Pattern Worth Extracting
Look at what the higher-ranked opportunities have in common.
They are boring. Compliance documentation, invoice processing, regulatory filings. Nobody demonstrates these at a conference.
They solve expensive problems that already have budget attached. You are not creating a new line item. You are competing with an existing cost, which is a much easier sale.
The AI is not the product.
In every case, the durable asset is something else: domain knowledge, permissioned data, workflow integration, regulatory expertise, or relationships. The model is rented infrastructure, and rented infrastructure is never a moat.
They are narrow. Every MVP described above is deliberately smaller than feels ambitious, because narrow scope is what allows a small team to be genuinely better than a large one at something specific.
There is a corollary. If you are choosing between an idea you find exciting and an idea where you have unusual access to a specific industry’s problems, the access is worth more than the excitement. Domain knowledge takes years to acquire and cannot be shortcut. Technical capability can be hired, partnered for, or learned.
The single most reliable starting move is not on the list as an entry, because it is a method rather than a business: find work that expensive people do repetitively in an industry you can reach, do that work for a few paying customers, automate the parts that repeat, and let the product emerge from the operation. It is slower than launching a product. It is considerably more likely to produce one that survives.
Frequently Asked Questions
Do I need to be technical to start an AI business?
No, but you need genuine access to something: a domain, a customer base, or a technical partner. Non-technical founders should look at training, consulting, or process outsourcing before software.
How much capital is required?
The services-led options on this list can start with very little beyond time. Vertical software targeting regulated industries realistically needs funding or a long runway, because sales cycles are measured in quarters.
What about competing with OpenAI, Google, and Anthropic?
Do not compete on model capability. Build where the constraint is access to data, workflows, and relationships that model providers do not have. Note the genuine risk that these companies move further into vertical applications over time, which is an argument for depth rather than breadth.
Is it too late?
For thin wrappers around model APIs, yes, and it was late a while ago. For applying AI to specific industry problems, adoption data suggests most of the market has not started. McKinsey found no more than 10% of organisations scaling agents in any single business function, which describes an early market rather than a saturated one.
Sources and Further Reading
Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (press release, 25 June 2025). Includes the “agent washing” analysis and the estimate of genuine agentic vendors.
McKinsey & Company, The State of AI in 2025 (5 November 2025). Adoption and scaling figures by function.
McKinsey & Company, Building the Foundations for Agentic AI at Scale . Data as the primary constraint.
Industry reporting on vertical AI funding and revenue through 2025 and 2026. Private company revenue figures in this category are reported rather than audited and should be treated accordingly.
Andreessen Horowitz’s published thesis on vertical software and AI, referenced here as a framework for how vertical categories have historically expanded revenue per customer.
The opportunity is rarely in building a better model. It is in knowing one industry deeply enough to see the work nobody else can see.
The GosAI
If you are weighing one of these and would like an honest read on what it would actually take to build, get in touch with The GosAI.
