Competitive Analysis for AI Startups – How to Find Where You Can Actually Win
Competitive analysis for an AI startup is the work of identifying every way a buyer could get the same job done today (rival products, platform features, in-house builds and plain manual effort) and then judging which of your advantages will still matter once the underlying models get cheaper and more alike. The evidence says that second part is where most analyses go wrong: model performance is converging fast, so the useful output is a short list of defensible positions and the decisions that follow from them, not a grid of logos and feature ticks.
Why AI markets need a different kind of analysis
Classic strategy tools still apply. Michael Porter's Five Competitive Forces (Harvard Business Review, January 2008) argues that competition for profit goes beyond established rivals to include customers, suppliers, potential entrants and substitute products. For an AI company, two of those forces carry unusual weight.
The first is suppliers. Most AI startups rent their core capability from a handful of model providers, and those providers also sell directly to the same customers. The second is new entrants. When the cost of a capable model collapses, the barrier to copying a thin product collapses with it.
The numbers behind this are stark. Stanford HAI's 2025 AI Index reports that the inference cost for a system performing at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024. The same report found that the score gap between the top model and the tenth-ranked model shrank from 11.9% to 5.4% in a year, with the top two separated by 0.7%, and that open-weight models narrowed their deficit against closed models from 8% to 1.7% on some benchmarks. If your pitch rests on access to a better model, the market is eroding that claim every quarter.
Start from the job, then list every alternative
Clayton Christensen and his co-authors set out the "jobs to be done" lens in the Harvard Business Review article Know Your Customers' "Jobs to Be Done" (September 2016). Its value for AI founders is that it forces the question away from technology labels. "AI for legal" describes a technology. "Get a first-pass risk summary of a supplier contract to a partner before Thursday's call" describes a job, with a buyer, a deadline and a cost of failure.
Once the job is written down, list the alternatives in four groups:
| Group | Who it is | The question it raises |
|---|---|---|
| Direct rivals | Startups and incumbents doing the same job for the same user | Why would a customer switch now? |
| Platform features | Model providers or large software vendors adding the job as a feature | What happens when this becomes a free checkbox? |
| Internal build | The customer's own engineering team using model APIs | Can you be cheaper and better than their build? |
| Non-AI substitutes | Spreadsheets, junior staff, outsourced services, doing nothing | Is the change worth the effort at all? |
The internal build line deserves more attention than it usually gets. Menlo Ventures' 2024 State of Generative AI in the Enterprise, based on a survey of 600 US IT decision-makers at companies with 50 or more staff, found that 47% of generative AI solutions were developed in-house and 53% bought from vendors. Nearly half the time, then, your real competitor is the customer's own team.
Keep the list short enough to act on. If research turns up forty names, group them and pick the five that come up in actual sales conversations.
Gather evidence that goes beyond the demo
Demos show the best case. The more useful evidence sits around them.
- Pricing pages and terms. Note the pricing unit (seat, usage, outcome), minimum commitments and what is gated to enterprise tiers.
- Documentation and changelogs. These reveal integration depth, data handling and how fast a team really ships.
- Hiring. Job adverts show where a competitor is investing, for example evaluation engineers or sales staff in a new region.
- Company filings. For UK private companies, the free Companies House search shows registered details, current and resigned officers, filed document images, charges, previous names and insolvency information. Filed accounts will not reveal everything about a small company, but they do show whether a rival is as big as its marketing implies.
- Customer conversations. Ask prospects what they tried before, what they pay for now and what made them stop. Lost deals are the most honest competitive data you will get.
Write down where each fact came from and when you checked it. Competitive claims age quickly, and a wrong one in a sales deck damages trust.
Compare workflow depth and evaluation quality
After the first good answer, a lot still has to happen. Who reviews the output? How are errors caught? What is stored, and who can see it? How does the result reach the system of record? A product that only returns text can be replaced in an afternoon. A product that carries approvals, audit history and integrations is much harder to rip out.
Evaluation is the second comparison worth making. Ask how each competitor measures quality on customer-like data, and whether it can show results when it changes the underlying model. Public benchmark scores say little about a narrow business task, and given the convergence Stanford reports, they say less every year. A competitor with a disciplined evaluation process can swap in cheaper models quickly and pass the savings on. That makes it a more serious threat than one with a flashier demo.
Judge which moats are real
Investors have been blunt about this for years. In January 2023, Matt Bornstein, Guido Appenzeller and Martin Casado of Andreessen Horowitz wrote in Who Owns the Generative AI Platform? that "there don't appear, today, to be any systemic moats in generative AI." They observed that many application companies were "relatively undifferentiated, since they rely on similar underlying AI models", with gross margins that were sometimes as high as 90% but more often 50 to 60%, driven largely by inference costs.
A few months later, a leaked internal Google memo, published by SemiAnalysis on 4 May 2023, put it more sharply: "We have no moat. And neither does OpenAI." The anonymous author argued that open models were catching up faster than the big labs could keep secrets.
Neither piece means defensibility is impossible. It means the moat rarely lives in the model. When you score yourself and each competitor, test these candidates:
- Proprietary, refreshed data tied to the workflow, such as expert corrections captured every day, which a rival cannot buy.
- Workflow embedding, where templates, approvals and history make switching painful.
- Distribution, such as a reseller channel, a marketplace listing or an existing customer base that lowers acquisition cost.
- Trust and compliance evidence in regulated sectors, which takes time to accumulate.
Track pricing, platform risk and switching costs
Model prices move constantly, and the spread between tiers is wide. On Anthropic's Claude pricing page, for example, Opus 5.5 is listed at $4 per million input tokens and $20 per million output tokens, while Sonnet 5.5 is $2 and $10, with a 50% saving for batch processing (checked October 2026). A competitor that routes routine work to cheaper tiers or batch jobs can undercut you without a better product. Compare cost per completed task, not cost per token.
Switching behaviour among enterprise model buyers is also instructive. Menlo Ventures' 2025 mid-year LLM market update found that 66% of builders upgraded to newer models from their existing provider and only 11% switched provider entirely. Its estimates put Anthropic at 32% of enterprise usage, ahead of OpenAI and Google (20%). The authors also noted that builders "consistently choose frontier models over cheaper, faster alternatives". Customers stick with what works once it is embedded, which helps incumbents and should shape how you plan to displace one.
Platform risk belongs in the same section. The UK Competition and Markets Authority's initial report on AI foundation models, published on 18 September 2023, examined how these markets could develop. Alongside it the CMA proposed seven principles: accountability, access, diversity, choice, flexibility, fair dealing and transparency. For a founder, the practical reading is simple. Note which of your suppliers also compete with you, and keep a tested fallback model.
Turn findings into decisions
An analysis that changes nothing is a waste of a week. End with three short lists: where you are clearly differentiated, where you are behind, and what you will stop building. Give each item an owner and a date.
Narrowing is often the right call. A small company competing with large platforms across every use case usually loses, while one that owns a single painful workflow, with proof from real customers, has a credible path to expand. Sequoia partner David Cahn, writing in AI's $600B Question in June 2024, argued that founders "will benefit both from lower costs and from learnings accrued during this period of experimentation". Cheaper infrastructure lowers your costs too. The question is whether you turn that into deeper workflow and better data faster than the next team.
Keep the final document to a page of findings, one comparison table and a list of actions. For more writing on running AI businesses and operations, see George Bregman's site.
Frequently asked questions
How do you analyse AI competitors?
Define the customer's job first, then list direct rivals, platform features, internal builds and non-AI substitutes. Compare workflow depth, evaluation discipline, data rights, distribution, pricing units and switching costs, using dated evidence from pricing pages, documentation, filings and customer conversations.
What makes a moat for an AI company?
Rarely the model alone. Industry analysis since 2023 has questioned whether generative AI has any systemic moats, so look for combinations of proprietary data that keeps refreshing, embedded workflow, distribution and accumulated customer trust.
How often should the analysis be updated?
There is no standard interval. Refresh it when a competitor changes pricing, when a model provider launches a feature in your segment, before a fundraise, and before you change your own pricing. Keep a running log of competitor changes between full reviews.