HOME›BLOG›WHICH STARTUPS ARE AT RISK FROM AI LABS?
GTM NOTES6 MIN READNOV 16, 2026

Which startups are at risk from AI labs?

37 types of startup business, each rated durable, contested or exposed to AI labs by 2029, with the reason for each rating.

BADideas.fund

We rated 37 types of startup business on how likely it is that AI labs will do the same job by 2029. This page lists every type with its level and the reason for it. What it means for founders raising money, and how the market treats each level, is in Are AI startups at risk from the big labs?

How we rated each type

For each type we went through the same questions. What does the best company of this type become if it works? What is the one threat that would end most companies of this type, who brings it and when? Does that threat end them, or does the company hold something that stays hard to copy, such as a licence, a physical presence, its own data, its customers or the legal risk it carries? And how much money is already in the field, so that a new company arrives late? Only then did we give each type a level.

  • Durable: what a winner holds stays hard to copy through 2029. AI cuts its costs more than it cuts its advantage.
  • Contested: the type splits in two, or turns on something not yet decided. The line that splits it is given.
  • Exposed: by 2029 the labs or funded leaders do the core job well enough that most new companies lose their pricing or never reach customers.

Durable: 6 types

  • Software for small businesses in one trade. The system a small business runs on, with its payments, in one scattered industry. The labs do not build for such small customers, and small businesses change slowly.
  • Care delivery and eldercare. Care needs hands, licences and trust in the home. AI cuts the cost of organising it, not the need for people.
  • Insurance. A company that carries the risk, holds the licence and earns its own claims data has what AI cannot copy.
  • Manufacturing and industrial operations. The winner holds plant data, links into old machines and a presence on the factory floor. AI makes it cheaper to serve, not cheaper to copy.
  • Energy, grid and climate. Power stays scarce, and AI raises demand for it. Grid connections, licences and physical assets are hard to copy.
  • Defence, drones and dual use. Government buying, security clearances, tested hardware and field data are not things a lab copies. European defence budgets are growing.

Contested: 13 types, and the line that splits each

  • Data layer for AI. Owning data that only operations or experts produce holds. Tools that only move data are copied by databases and clouds.
  • AI chips and edge devices. Chip designs are physical and slow to copy. AI gadgets for consumers lose to phones and to hardware the labs make.
  • AI-native cybersecurity. Security that carries accountability, or runs in sovereign or industrial settings, holds. Detection sold as software is bundled by the big platforms.
  • Recruiting and HR. Payroll and employing people across countries hold. Sourcing, screening and interview tools are exposed to the big HR platforms and the labs.
  • Tax, accounting and audit. The ledger and the filing link to the tax office hold. A copilot that makes accountants faster does not.
  • Clinical AI. Certified diagnostic tools with medical approval and outcome data hold. Note-taking for doctors is taken by the hospital software vendors.
  • AI for science, drug discovery and biotech. Owning lab data or the drugs themselves holds, but needs more capital than early investors bring. Software-only design is absorbed by the labs.
  • Lending and banking. A lender that carries the credit risk and earns repayment data holds. Scoring and bank software becomes a feature.
  • Payments, identity and fraud. Licensed payment rails and fraud networks with their own transaction data hold. Software-only identity checks become a commodity.
  • Construction and real estate. Companies inside the building site, the rent flow or the materials trade hold. Document and planning tools for builders are bought by the big construction software firms.
  • Agriculture and food. Companies inside the trade and the supply of inputs hold. Farm data software is absorbed by the big equipment and seed companies.
  • Robotics. Companies that run their own robots in one industrial job and sell the result hold. Robot software and generic hardware are squeezed by DeepMind, Nvidia and cheap Chinese hardware.
  • AI-native services. Services that hold a licence or a signature, such as audit or tax filing, keep their margin as AI cuts labour costs. Services that only sell output, such as agencies, do not.

Exposed: 18 types

  • Model infrastructure. Owning power and chips is a capital game already won by large players. Serving and tuning models is done by open source and the clouds.
  • Coding agents and developer tools. The labs sell coding as their own product, so a new company competes with its model supplier on the supplier's best use.
  • Agent infrastructure. Every cloud and lab already ships the tools agents run on, often free.
  • Evals, observability and guardrails. Testing and monitoring AI became a feature of the clouds and data platforms within two years.
  • Security for AI and agents. The first wave of startups was bought by the security platforms, which now own the capability.
  • Sales agents. The agent ships inside the sales software the seller already uses.
  • Customer support and voice agents. Contact centre platforms ship the agent themselves, and a few funded leaders hold the independent places.
  • Marketing and content. Content stopped being scarce, and the ad platforms create and target inside their own tools.
  • Meeting notes and personal memory. Note-taking is already free inside the video call tools and the assistants.
  • Knowledge work, search and research. Searching and analysing a company's own documents is what the office suites and the labs already ship.
  • Finance back office and procurement. Paying bills, reconciling and closing the books are the tasks agents run inside the finance software companies already use.
  • Legal. Legal software is split between two funded leaders and the labs selling direct. The licensed law firm itself is the durable part.
  • Commerce and online store operations. Catalogue, pricing and store tools are what the big commerce platforms bundle first.
  • Education. Tutoring and exam help are free inside the assistants. Only certificates and school system contracts hold.
  • Government, compliance and regulation tools. Most of this work is compliance paperwork, which agents do well, and the labs sell to governments directly.
  • Logistics, mobility and fleets. Quoting, documents and tracking are back office work agents do inside the systems the big forwarders and Amazon already run.
  • Consumer AI apps. The assistant people already use does the same job as a feature, so most new apps never reach users.
  • Creative tools and media generation. Generating images, video and voice is a lab capability that gets cheaper every year, and the leaders are already funded.

Ask a follow up

Send the deck. We read every one.

Yes or no in five business days, with the reason either way.

↑
SEND YOUR DECK (PDF, MAX 20MB)
Yes or no in five business days, with the reason.
New guides, data and reports, once a month.