What's the point of building a business now?

Are we all screwed?

At the end of January, Anthropic released some plugins for Claude: legal, sales, marketing, data. On paper it was a small launch.

Within a few days Thomson Reuters had closed down about 16% in a single day, RELX was on course for its biggest one-day drop since 1988 and LegalZoom had lost a fifth of its value. Bloomberg put the total across software, financial services and asset management stocks at $285 billion.

Three weeks later Anthropic put out a blog post about using Claude Code to modernise old COBOL systems, and IBM dropped 13%. That was its worst day since 2000, and about $31 billion of value went with it.

If markets react like that to a blog post, I understand why so many founders ask me whether there's any point building something right now. I've asked myself the same thing more than once this year, so I went and looked at the numbers.

In the first half of 2026, 86 cents of every US venture dollar went to AI. Rounds of $100 million or more took 87.5% of the total, which left every smaller round in the country, from seed to late stage, sharing the other 12.5%. Anthropic's $65 billion round on its own was about 16% of what US VCs invested in those six months.

Total funding went up in 2025 while the number of deals fell. Startups on Carta closed 4,859 rounds, the fewest in at least six years and 41% below 2021.

Growth is harder to hold on to as well. ChartMogul tracked 6,525 software companies and found that roughly 1 in 10 reach $10 million ARR within ten years of first revenue. A company growing 65% one year typically grows 28% the next, and median growth for private SaaS companies fell to 22% last year.

Public investors have priced this in. In February the median public software company traded at 3.6x its next twelve months of revenue, the lowest in over ten years, and it kept falling to 2.9x by June. By September it was back around 4x, while the top five traded at 31x to 40x. A handful of winners get rewarded and almost everyone else gets very little.

The stat I didn't expect was on new businesses. Americans filed a record 5.67 million business applications in 2025, 62% more than in 2019, but applications from businesses that look likely to hire staff dipped slightly on 2024. All of last year's increase came from businesses that are less likely to employ anyone.

Chegg sold homework help. In 2023 it lost almost half its value in a day after its CEO said ChatGPT was hurting new customer growth, and it has since cut 45% of its staff. Jasper sold AI marketing copy. After ChatGPT arrived as a cheap alternative, it cut its internal valuation by about a fifth and its revenue forecast by at least 30%, according to reports at the time. DocuSign fell 12% after OpenAI showed off an internal contract tool in a blog post, JFrog fell almost 25% on the day Claude Code Security launched, and Figma fell almost 7% the day Claude Design came out. Insurance brokers lost 9 to 12% in a single day in February, the week after an app for comparing car cover launched inside ChatGPT.

What strikes me is how small most of these launches were: a plugin, a demo, a blog post, an app someone else built. Very few were new models. The businesses on the receiving end mostly sold packaged information or a tool built around one job, and both are now cheap to copy.

Nikita Bier said as much on X at the weekend... Technical moats are mostly over, he wrote, and what's left is proprietary data, network effects and a physical component. Ben Thompson wrote in September that "models are relatively substitutable". Put the two together and the software you build on top of a model won't protect you for long, because the next team can copy it.

Software isn't dead, to be clear. A lot of February was panic, and Figma beat its numbers and raised guidance a few weeks after its drop.

I mostly agree with Nikita's list, but I'd add to it, and most of what I'd add is stuff services businesses have had for decades.

The data that matters is the data you collect by doing the work. Anyone can buy or scrape a dataset, but nobody else sees the corrections your team makes on every job. Bessemer described it in July as "every delivered unit of work doubles as a training signal". Each contract reviewed or deal run makes the system a bit better in a way a competitor can't copy.

The physical side is on Nikita's list, and it's real. microagi works on factory floors and Monumental on building sites, and neither can be shipped through an API.

Nikita doesn't mention accountability. When the work goes wrong, clients want someone insured they can hold responsible. AI can't be sued or struck off. A firm can, and clients pay for that.

Licences come with it. Garfield became the first AI-driven law firm authorised by the UK's solicitors regulator in May 2025. Lawhive took the other route and bought Woodstock, a regulated firm with 50+ lawyers. Getting a licence takes months of paperwork, and a model release doesn't change that.

Owning the result matters too. Julien Bek at Sequoia wrote that "if you sell the tool, you're in a race against the model." If you sell the finished work, a better model just makes your delivery cheaper.

And then brand, which used to be the soft one. NFX, the VC firm best known for writing about network effects, now says brand "has become paramount".

Go back through that list with a law firm or an M&A advisory in mind and they tick most of the boxes: relationships, a licence, someone accountable and years of data from real work.

Their margins have always been the problem. Traditional services firms tend to run at 10 to 30% gross margin, according to a Village Global write-up of Emergence partner Jake Saper's thesis, while Emergence says AI-native ones can reach 50% or more.

Investors notice the difference. When I compared valuations a few weeks ago, legal AI companies had raised at roughly 39 to 56 times annual recurring revenue at their latest rounds, while Knights, the listed UK consolidator, was buying law firms for around 1x revenue. A firm that keeps the protection services already have and gets its margins closer to software should land somewhere between those numbers.

Bek also wrote in that Sequoia essay that the next $1T company will be "a software company masquerading as a services firm". I agree, and my bet is it gets built from scratch.

I'd rather raise the obvious objection myself. If AI makes the work cheaper, clients will expect to pay less, and some already do. KPMG reportedly pushed its own auditor, Grant Thornton, to pass on AI savings, and the audit fee for KPMG International came down about 14%. Clio found that 75% of solo law firms already offer flat fees alongside hourly rates.

Firms that add AI and keep billing by the hour will mostly hand those savings straight to their clients.

The ones I'd back have changed how they charge. With Sierra, if its AI doesn't resolve a customer's issue, in most cases there's no charge. Crosby charges a fixed rate per document, and Garfield asks £7.50 for a letter before action. If you price per result, the time AI saves goes to your margin.

Lower prices also bring in work that never used to happen. Plenty of contracts never see a lawyer today because the fee isn't worth it, and at a low enough price they will. Bessemer's test for these companies asks two things: does demand grow when AI collapses the price, and can the company "keep the surplus rather than competing it away"?

I don't think so, though it's harder than it was a few years ago.

I'm biased, because this is the bet we've made at Eilla. Our AI does a lot of the heavy lifting on a deal, experienced M&A advisors check the work and run the process, and most of what we earn comes when a deal closes.

If you're building something right now, ask yourself what you'd still have if the next model release could do everything your product does. Data from doing the work, a licence, someone accountable, a physical presence, relationships people trust... any of those gives you something to build on. If the honest answer is mostly code, I'd start on one of those this week.

Which one are you building? Reply to this or tell me in the comments, I read all of them.