Why Becoming AI Native Is Harder Than It Looks

I round up the most relevant AI-in-finance news, the deals being done, who's rolling out what, and what's actually working on the front lines.

Moonshot released Kimi K3 on Thursday, 2.8 trillion parameters, and by Saturday it had stopped selling subscriptions because it ran out of GPUs. It now sits fourth on Artificial Analysis's index, a point ahead of Claude Opus 4.8, and it's joint first on the agentic banking benchmark at roughly half Opus's cost per task.

Meanwhile L.E.K. asked 100 US buyout professionals how good they are at this. One in five rated their own AI expertise as advanced. The same group reported an average 28% productivity gain from it.

Elsewhere: Stripe and Advent bidding $53bn for PayPal, BofA appointing an AI adoption chief 28 months after Morgan Stanley did, record quarters at Goldman and JPMorgan, and PE software deal value falling off a cliff.

But first, my take on why becoming AI native is so much harder than it looks. The first few weeks feel like progress, and that's precisely the problem.

In today's Acquisition Intelligence:

From The Trenches

Why Becoming AI Native Is Harder Than It Looks

The most useful question I ask on a call is also the simplest: what have you actually done?

The answer, more often than you'd think, is that they got Claude for the whole team. Great, I say. Are you doing better deals now? Bigger ones? Faster ones? And there's always a pause before the no, because nobody ever wrote down what the tool was supposed to fix, so there's nothing to check the result against. Nine times out of ten it wasn't the bottleneck in the first place.

If you want that pause as a number, L.E.K. put one on it this week. They surveyed 100 US buyout professionals, and only 21% rate their own AI expertise as advanced or leading edge. The same people think their investment teams and portfolio companies are further behind than they are. And yet the group reports an average 28% productivity improvement from AI already.

Honestly, I'm sceptical of even the 21%. Advanced is a relative term and hard to quantify, and from what we see day to day the real figure would be much, much lower.

So the gains are real and measurable, and the overwhelming majority of the people delivering them don't really know what they're doing yet. That gap between day-one output and competence is where almost everyone I speak to is standing right now.

Day One Is Free. That's The Trap

A chatbot is good the first time you touch it. You type a question, something competent comes back, and the feeling of mastery arrives immediately and at no cost.

Nothing else in software has ever worked like that. Excel was hostile for a month. A CRM rollout took a quarter and a consultant. You knew you were a beginner because the tool reminded you constantly.

Claude never reminds you. It stays helpful and articulate right up to the point where the work needs real craft, and then it produces something that still looks fine. The output degrades far more slowly than your competence, so there's no signal marking the edge of your ability, and most people sail past it without noticing. I said this to a client last week and it came out blunter than I intended: you have to crawl before you sprint, and most firms are trying to sprint before they can crawl.

The $600,000 CRM

Which brings me to the post everyone argued about this weekend. Harry Stebbings shared a clip from a 20VC interview on Saturday, and it's done north of 450,000 views since:

Is it an anomaly or the start of a trend, he asked. The company is multi-billion-dollar, the build took three weeks, and 80% of their internal SaaS is apparently next.

The timeline was sceptical, and rightly so. Jason Lemkin's reply, seventeen minutes later, became the consensus: great for the 0.1% who can maintain it, build the integrations and handle what breaks, right up until the person who built it quits.

All true, and I'd sign every word. But it stops one step short of the point that matters.

The Same Ceiling, In A Nicer Font

Most CRMs are basic software. Salesforce and HubSpot included, they're tables with permissions and a reporting layer bolted on, and the value-to-cost ratio has been indefensible for years. So when people work out they can replace one, plenty do. The instinct is completely right.

The problem comes when you rebuild the thing you already have. Your requirements end up as a description of the system you're replacing, because that's the only version of the job you've ever seen. You confine yourself to everything you previously knew.

So you arrive at exactly the same ceiling. Same fields, same reports, same blind spots, and what's changed is the font, where the buttons sit, and the fact that you own it now. It's pantomime. And it's the most convincing form of false progress there is, because you did build something, it does work, and it did take $600,000 off the bill. Every signal says you've moved. You're standing where you started.

The question nobody in that thread asked is what the system should do if it were designed today, with what we now know, by someone who understands both the technology and the work. A CRM in the right hands can go far beyond anything currently on the market, and that version doesn't get built in three weeks, because most of the work sits underneath the interface, in the structured record a system needs before it can do anything the old one couldn't. That layer is where we spend most of our time at DealSage, and it's slow precisely because it's the part that matters.

The bar for what software can be just rose, on capability, quality, personalisation, all of it. Most people haven't noticed it move. Rebuilding yesterday's software without yesterday's licence fee isn't clearing it.

The System Underneath You

The other failure mode gets less airtime. You try the tools, nothing miraculous happens, and you conclude the whole category is oversold. Fair enough, given what you were promised. But the promise was pointing at the wrong thing.

Everybody wants the analyst in a box, the associate that builds the flashy model and drafts the whole memo. Right now most of the value sits somewhere much less sexy: the systems that keep the work running. We massively underestimate how much time they eat and how much they're worth.

My contact lists are a good example. Keeping them current used to be half an hour I reliably postponed, and postponing it meant this newsletter grew more slowly than it should have. It runs by itself now. The same goes for tracking a referral network, and for the obvious one, keeping the CRM up to date. A solo advisor I spoke to this week put the problem better than I could: how often do I update those Excel trackers if it's a mad week? Never. The task takes thirty minutes, thirty minutes doesn't exist in a mad week, and the tracker stops being true.

Underneath all of that sits the information itself. We come across valuable pieces of it all day, on calls, in emails, in passing, and mostly we do nothing with them. This is what AI is so good at now: capturing it, organising it, storing it, linking it, and doing the backend work that makes it all start to compound. You're building a system underneath yourself, and that's where I think the real strength lies.

That, for me, is what becoming AI native means. A system built underneath you, one unsexy piece at a time. It takes a lot longer than a weekend, and it compounds for years.

What The Builders Are Saying

More of the interesting arguments land on X before they land anywhere else. Five from this week. Worth following all of them.

@HarryStebbings (Harry Stebbings, 20VC) on switching costs collapsing

The post: a migration that would once have taken a large engineering team the best part of a year, dashboard conversion and all, done by a small team using agents. His conclusion: as migrations get cheap, legacy data structures stop working as a moat.

Why this matters: he posted this the same day as the $600k CRM clip, and together they're one position: the cost of moving off software is collapsing faster than the cost of building it. If you hold software assets, the moat you underwrote two years ago needs repricing. If you're the buyer, the lock-in discount runs the other way.

@nicbstme (Nicolas Bustamante, founder of Fintool, acquired by Microsoft) on what survives a model upgrade

The post: "Most system prompts and skills have a very short life expectancy. They are usually patches for the current model's shortcomings. But those shortcomings are exactly what the next model gets trained and post-trained to fix." A skill that teaches a model how to build a DCF or run a standard workflow eventually disappears, because the model learns it. What lasts is "context that cannot live in the model weights: your preferences, your company's private knowledge, your tools, your data, and the specific way you want work done."

Why this matters: I agree with this. A lot of the harness and guardrails you're putting in place right now won't really be needed in a year, maybe a handful of months. The organised, clear, structured data will be.

@jawwwn_ on Alex Karp's warning

The post: a clip of the Palantir CEO with Mathias Döpfner. Karp argues Silicon Valley is overselling AI while telling the public "don't believe your lying eyes". His worry is distributional: AI raises the average standard of living while the people building it get 10 to 100 times wealthier, and dismissing people's fears just drives protest voting: "screw these people, I'm going to burn down the house."

Why this matters: the gap between the story being sold and the thing being lived is the theme of this whole issue, and Karp is describing it at country scale. Inside a firm it produces disillusioned teams and quiet non-adoption. At his scale it produces politics no deal model prices in.

@demishassabis (Demis Hassabis, CEO of Google DeepMind) on a FINRA for frontier AI

The post: a long essay arguing AGI is a few short years away, "a way to make sand think", and proposing a US-initiated Frontier AI Standards Body modelled on FINRA: federally overseen, setting capability benchmarks, testing frontier models before release, able to coordinate a slowdown if it ever came to that.

My take: sure. But he and Karp both run enormous companies that are perpetuating the very risks they warn about, prioritising commercial success at the expense of everything else. A standards body is easier to propose when you already own the frontier it would police.

@saneord on how fast this all moves

The post: a quote-tweet of Arena's announcement that Kimi-K3 had taken the number one spot on its Frontend Code leaderboard, and the entire reply reads "29 days ago...". The joke lands because the result was hours old and already felt like history.

Why this matters: the model race now moves faster than anyone's evaluation cycle, including yours. Which is a decent way into the story below.

News Digest

Kimi K3 Tops the Agentic Banking Benchmarks, Then Runs Out of GPUs

Moonshot released Kimi K3 on 16 July, 2.8 trillion parameters, billed as the world's biggest open-source model. The weights haven't shipped yet. Within 48 hours it had stopped selling subscriptions, because demand had run to the limits of its GPU capacity.

The details:

Why it matters: The model that just took the top spot on an agentic banking benchmark is Chinese, costs half what Opus does per task, and can't currently sell you a subscription.

My take: Sticker prices at the frontier are rising while cost per task keeps falling, so the deflation story survives once you adjust for capability. The benchmark mix matters more: agentic banking, legal and SaaS workflows are exactly the work a fund would hand an agent, and a model almost nobody in this industry is evaluating leads on most of them. And the subscription pause tells you compute, not capability, is the binding constraint on Chinese labs.

Read more

Other Interesting Things I've Read or Seen This Week

Stripe and Advent offer $53bn for PayPal (July 16) - $60.50 a share all cash, a premium of roughly 28% to 30%, the two owning it equally with no plan to break it up. The board reportedly thinks it undervalues the company. Enrique Lores, four months into the job, has said PayPal underinvested in its technology and wants to strip out layers to spend more on AI, which is a diagnosis a sponsor can underwrite.

BofA names an AI transformation head (July 17) - Kevin Milsom gets the role inside global markets, reporting to a platforms head. Morgan Stanley created its firmwide equivalent in March 2024, reporting to two co-presidents, with a governance mandate and a steering group. Twenty-eight months and several rungs of seniority apart, which tells you most of what you need to know about how seriously each bank takes adoption.

Goldman and JPMorgan both post record quarters (July 14) - Goldman revenue up 39% to $20.3bn, JPMorgan up 27% to $58bn, equities trading $4.4bn above expectations between them. Goldman's CFO called it "an AI capex super cycle where there are demands on financing in every single financing instrument, in every region of the world and across every single industry", a sentence that would have got you laughed out of a credit committee three years ago.

US private equity software deal value plunges (July 10) - the largest quarterly decline of any sector on PitchBook data, as AI revaluation squeezes buyers and sellers at once. The SaaSpocalypse showed up in the deal data before it showed up in anyone's model.

Wall Street wants to trade AI compute like oil (July 6) - a16z-backed Ornn is building a GPU pricing index, and exchanges are looking at compute futures. Given Moonshot just ran out of GPUs mid-launch, the demand side of that market is not hypothetical.

Netflix paid $587m for Ben Affleck's AI firm (July 18) - the price only surfaced in an SEC filing, months after the March deal closed.

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. We build the structured record and the repeatable processes underneath the deal work, so the grunt work goes away and the judgement stays with the people who have it. Reply here or have a look at dealsage.io.