Goldman says AI will touch 5% of tasks. Two investors take the report apart

No Priors skips a guest this week. Sarah Guo and Elad Gil sit down to "shoot the breeze," and Guo climbs onto the soapbox immediately. Everyone in their circle is reading a Goldman Sachs report whose title, as she summarizes it, is something like calling the top on AI. For obvious reasons she does not like it. She still wants to decompose it, and she tells listeners to skim the thing themselves.

Two interviews carry the note. Daron Acemoglu, an MIT professor, argues that AI will affect fewer than 5% of all tasks and that the trillion dollars of capex going into training is a waste, because the systems cannot solve the complex problems they are being asked to solve. Jim Covello, Goldman's head of global equity research, draws a contrast with the internet. The internet, he says, disrupted something expensive from the start. AI is an expensive solution on day one. Any efficiency it creates will be competed away, so no company really keeps the gain.

Gil takes Acemoglu first. The claim that more data will not make a customer-support rep better is, he says, an argument about a technology the author does not understand, and it is objectively at odds with the last several years of scale and data improving model output. A lot of commentators, he adds — not necessarily these two authors; he has not read the full report — are still in the old AI world. They imagine munging data and running a fancy regression. Transformer-style models are scale-dependent and do a different kind of work. In healthcare he still hears "data is the new oil." Labeled data can be extremely useful. The model and the application decide whether anything happens.

Guo agrees and calls it a "this time it's different" case, plus a failure to look at the ground. There are already cost-effective things to do. Prosumer products are democratizing capability, and even healthcare — a famously slow industry — has companies going from zero to $5 or $10 million of run-rate in year one. The second hidden assumption in the bear case is that the technology will not get much better, or much cheaper, very fast. "The willingness to predict 10 years into the future of insignificant improvement is ludicrous when literally all of the people working on this tech are unwilling to predict two years."

Covello's picture has a third flaw. Treating the way companies apply AI as uniform is, Guo says, like saying everyone will use the internet, so no firm can gain. The internet produced Amazon and it produced Borders. You want to be Amazon, or Klarna, or whoever is actually rewriting a cost structure. Gil pulls an older bottle off the shelf. IBM's CEO in the internet era called the new companies "fireflies before the storm," the storm being IBM and its peers. The trillion-dollar names that followed were Amazon and Google. "This is an old story. It's old wine in new bottles."

The prediction Gil will make with confidence is the opposite of a top. Two years out there will be even more widespread applications, because this is early. Large enterprises, he says, think about AI as three questions: which vendors can I buy from, how do I adapt internal tools, and what do I do to external customer products. Almost everything so far has been the first bucket — buying Decagon, Harvey, and their peers. The product side is still the AI-native prosumer apps, ChatGPT and Perplexity. The big wave of enterprise adoption has not happened. Planning cycles and ordinary friction have not even begun to clear.

What is already open, he argues, is new markets, for three reasons. The capabilities span biology, robotics, language, image, and video, and one general model now does work that used to need a pile of bespoke mini-models. Anyone on earth can hit an API; there is no need to build MLOps or train a model. And organizations have a mandate to "do AI," the way they once had a mandate to become an e-company. Buying behavior has shifted. Products that would have been unthinkable without that mandate are suddenly in the room.

His team is walking the map market by market. U.S. software spend is something like half a trillion dollars a year. Headcount cost in services industries that AI could transform — legal, accounting, sales — is about $5 trillion. For each vertical they ask whether a company already exists, whether something should be incubated, or whether the right move is an AI-driven buyout that takes over the asset and changes cost structure or leverage per employee. For the first time, two plays that are usually bad ideas start to make sense: incubation, and buying the company in order to rewire it.

Guo states the buyout premise. In some industries technology will not land fast, and in some cases the incumbents have an incentive not to adopt it. Taking control shortcuts adoption, shortcuts change management, and lets you rework how a slice of the organization functions. Klarna did that to itself on support: OpenAI plus custom workflows, 700 fewer people, 24/7 coverage, almost 20 languages, higher NPS than the human reps, faster replies, fewer repeat incidents. In a lot of industries cost is leverage on time. Keep the same employees and make them ten times more effective. A lot of so-called white-collar work is what people call email jobs — copy-paste between spreadsheets, inbox, CRM — and pieces of that are now automatable.

The case against is a skills mismatch. Identifying assets, running services companies, and living through the operational intensity of acquisitions is not what a software-heavy founding team usually knows how to do. Getting the right operators, Guo says, is the actual question. Gil puts the last cycle on the table. "Tech-enabled buyouts" mostly failed. A PE buyer rolled up companies, painted on a thin software veneer, bought cash-flow assets on a private-equity multiple, and raised on a tech multiple. Early investors and founders did well. Later investors and employees did poorly, because the market eventually priced the thing as PE and the multiple came down.

This time the leverage is not a veneer. One deal Gil backed is driven by a software founder who is hiring a PE team, the reverse of the old stack. You need both DNAs. Financial engineering without the technology is just a roll-up with a new name. What makes the idea resonate for Guo is a sentence she has heard from friends running tech-eating services companies in fragmented industries. In a small room they say the automation works. Distribution is the problem. You are buying the customer relationships. You are buying the whole thing.

Guo's historical test for incubation is narrow. It makes sense when you can see an opportunity or a customer set that others cannot, and when the mix of expertise a company needs is unlikely to assemble on its own. That combination is the alpha. Gil is blunter. Incubations are usually a terrible idea. Most firms that try them do not do them well. The shops that work tend to be vertically specialized — they really understand DTC or healthcare — and they have proprietary customers, a captive buyer, or a product idea they already know will land.

The startup ecosystem works, he says, because something like 5,000 founders explore markets and technologies in parallel and a handful, maybe three to ten a year, actually matter. Snowflake, incubated by Sutter Hill Ventures, is the famous counterexample, not the mean. What is odd in AI is the split screen. A dozen companies crowd into the same product. Other places that are obvious to build are empty. Large enterprises also want to adopt, so there is customer pull that usually does not exist. Gil is working on a handful of incubations he would normally refuse.

The last one he did, about a year earlier, was Braintrust with Ankur Goyal, whose previous company Figma had acquired. Evals, a prompt playground, and a few related products, rolled together, because the two of them kept asking what every enterprise needs in order to adopt AI. He is now looking hard at healthcare, building a founding team against large enterprise assets that would be instant customers, and he has two more areas he lacks bandwidth for: services, and a large-scale model aimed at a specific application. Guo's first incubation, years ago, was The Wake — last-generation machine learning on network data for security. She had just finished a healthcare incubation when they recorded. The mismatch she names is simple. The people who understand the technology and the people who understand the domain are rarely the same people. Pair them and the empty markets are the opportunity.

On public and late private books, Guo's short is specific. Mid-to-late-stage private companies and some small publics that lack the speed and the founder will to change, even when they can see the writing on the wall, are the ones she would want to fade. Not every company. Some founders are attacking the opportunity. Most will not. Anyone holding those names, she says, should rationalize the book. Gil thinks less in tickers than in durability. Ten years ago the hold-forever set was FANG — Facebook, Alphabet, Netflix, Amazon — later joined by Nvidia and rebranded the Magnificent Seven. The first basket is the next wave of multi-hundred-billion or trillion-dollar compounders. He names SpaceX and Stripe as the kind of company that might keep compounding for a decade.

The second basket is big tech that still gets a lift. He would not be surprised if Apple earned another iPhone replacement cycle just for AI: a new chip, smaller models that run locally, upgrade the phone. The third is what he calls AI-durable companies — railroads, certain SaaS products — where adding AI does not matter. If the model does not improve the product, AI is not a competitive factor, which is its own kind of moat. The fourth is an AI index besides Nvidia. Scale is one name he floats. Those four segments — a subset of big tech, the next Magnificent Seven, what survives AI, and the pickaxe — are the frame he would use to build the book.

Staffing is the last turn. Guo has not seen founder quality like this in a while, especially among second-time founders, and the metric they keep chasing is revenue per employee. The one-person billion-dollar company is a slogan; the serious version is a small team and more technical leverage, which is what startups already do when they never have enough money. Gil thinks that world is still a couple of years out. Coding copilots help. They are not a 5x. When coding agents work at scale, cost structures move. A six-to-eight-engineer seed company has almost nothing to cut. Guo agrees: this generation of models is not smart or coherent enough, and you will not eliminate headcount in the first ten people. You look for pockets of workflow that already sit at minimum viable quality. When she asks founders which functions shrink first, the answer is two to three years, and then SDRs and support, once the company is hundreds of people.

Velocity is the new constraint. Experienced go-to-market leaders still matter, and they arrive with a big-company pace. For AI companies that already have market pull, Guo would keep the product cadence and hunt only the veterans who can match or raise it. Gil's structural read is simpler. Big companies suck, and AI lets a startup punt on becoming one for longer. He also notes a product shift. At Midjourney, HeyGen, Ideogram, Suno, Udio, and Pika, the model is most of the experience, and the aesthetic of that model — the output, the data — comes from the founder, the way Ivan Zhao's Notion or Stewart Butterfield's Slack once did. AI companies, more than ordinary software, are founder-driven in the product and the weights.