How The Middle Market Wins

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.

Google spent the week losing the people who built its AI. Demis Hassabis stepped back from running Google DeepMind to become Alphabet's chief scientist, and Jeff Dean left alongside three other senior figures to start a company called Discovery Loop. Alphabet fell 4 per cent, and the flagship Gemini still has not shipped after a June launch date came and went.

Earnings went the other way. Palantir raised guidance again on 93 per cent revenue growth, and Airbnb hit a four-year high after telling the market its AI assistant had cut support cost per booking by about 16 per cent.

Also inside: ByteDance training a 10 trillion parameter model, Apple renting Alibaba's Qwen to run Siri on Chinese Macs, LinkedIn shipping a button that lets you report AI slop, and Wall Street bonuses climbing while the hiring does not.

But first, my take on the middle market, because the numbers everyone quotes about AI failing came out of companies that look nothing like the ones I spend my week inside.

In today's Acquisition Intelligence:

From The Trenches

How The Middle Market Wins

Someone asked me this week what he was actually supposed to believe. One thread had open source models closing on the frontier and token costs running away with themselves. The next had a founder swearing he had automated an entire marketing department at the click of a button. Then a damning report on deployments that went nowhere, then a run of posts calling the whole category a bubble.

He runs a good business and he is nobody's fool. He simply could not tell which parts of any of it were true, and while he waited to find out, another quarter went past.

The first half of my answer was blunt. Most of what he is reading on LinkedIn and X is rubbish, in both directions. Nobody automated their marketing department by lunchtime. There are no shortcuts here, the same as anywhere else in business, and anybody promising him one is selling something.

On his second question, whether to hold off a little longer, I urged him heavily not to. The experimenting phase of AI is behind us. We are well and truly into a known, quantifiable playbook, and it is available to him right now.

The shovel sellers on LinkedIn and X were nothing new. What struck me was everything underneath them. So much of the prevailing narrative, the serious arguments and the reporting included, is skewed towards enormous enterprises and the frontier providers implementing for them. Very little of it is relevant or appropriate to a middle market business.

Why The Enterprise Is Distorting The Picture

OpenAI and Anthropic are embedded with the Fortune 100 and the largest allocators. Rational, given that is where the budgets are. The press follows the labs and the labs' biggest customers, so the running commentary on AI describes the problems of the very largest organisations on earth.

Those problems are real. Token costs bite, and I have spent issues of this newsletter arguing they bite everyone rather than only the giants. Data sovereignty is just as serious, more so the moment you handle anything a customer would mind seeing leaked.

Those pressures reach a middle market business too, and they are worth managing properly. What they should not do is hold up the work. Caution calibrated to forty thousand seats has been handed down to companies that could have the thing running before a large enterprise finishes its procurement review.

What It Takes Inside A Giant

I remember this from my J.P. Morgan days, and Ben has the same scars from his time at Palantir. Walk into an organisation of that size, ask for the five datasets you need to rebuild one workflow, and months disappear collecting sign-off from every owner of every field. Nobody has written a line of code yet.

It can still be worth doing. Two of them proved it this week. Palantir raised guidance again on 93 per cent revenue growth. Airbnb put a 16 per cent cut in support cost per booking on the record. Both are written up below, though look at the price of admission: years of work, enormous sums, and in Airbnb's case a rebuild of how the company operates.

Why The Middle Market Is Different

A middle market business carries two decades of accumulated workaround, plus a handful of systems that have never spoken to each other. Plenty to go at, in other words.

It also has a surface area small enough for one competent team to hold the whole thing in view. The person who can authorise the data is usually already on the call. When a decision is needed, a room settles it that afternoon. Put those together and you can run a genuinely aggressive timeline, which is the thing no budget buys you inside a Fortune 100.

The Reports Are Catching Up

RSM published a survey on 21 July that almost nobody picked up: 1,030 companies, 827 American and 203 Canadian, fielded in March, margin of error 3.1 points.

Eighty-six per cent have partially or fully integrated AI into their operations. Ninety-seven per cent report moderate or high success from their pilots. Fifty-four per cent say the return has already beaten what they expected. The whole thing is worth reading.

So the evidence is starting to point where I would expect. There is obviously still a long way to go.

The Hard Part

That same survey asked the firms reporting moderate or limited pilot success what was stopping them scaling:

Not one of those gets fixed by a better Claude subscription. It is slow, human-led work: sitting with the people who do the job, learning what the fields actually mean, reconciling systems that have spent twenty years ignoring each other. Which is also why the cheap-provider-and-see-what-happens plan returns nothing. Nobody sells you the hard part by the month.

The Ambition Gap

According to the RSM survey, 45 per cent of middle market firms are focused on implementing AI where it delivers clear value today. Only 17 per cent are pursuing transformation across the enterprise.

Going after the whole business is what actually pays, because a company is one connected thing. Say you want to automate the month-end close. Doing it properly means sitting with whoever reconciles by hand, then the billing team feeding them numbers, then the operations lead who knows why one region always books late. Three functions mapped. Data quality fixed in all of them.

Take it one use case at a time and you pay that price on every project, each stopping dead at its own edge. Go at the business whole and the second workflow costs a fraction of the first. By the fourth it is close to free. Everything you wire in raises the value of what is already wired.

My honest view is that this conversation has been aimed at the wrong companies for two years. The firms best placed to do it well are hearing about it last, second hand, from people whose constraints they do not share. If you run a middle market business or a portfolio of them, you be considerably more ambitious than the coverage suggests.

What The Builders Are Saying

Three posts this week. The first one is the best context you will get on the Google story below, and the second is a decent signal about where the general audience has got to. Worth following all three accounts.

@firstadopter (Tae Kim, tech writer) on Sergey Brin

Why this matters: Google had the research, the talent and the compute, and lost three years to caution. That is a different failure mode to being out-innovated, and it is the one large organisations actually suffer from.

@emollick (Ethan Mollick, Wharton)

Why this matters: agent-to-agent coordination has left the architecture diagrams. Worth the eighteen minutes.

@elvin_not_11 ("Me when I use Opus 5")

If you know, you know.

News Digest

Google's AI Leadership Comes Apart

Alphabet announced on 5 August that Demis Hassabis is stepping back from running Google DeepMind to become Alphabet's chief scientist and the unit's chairman. The same day, Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le left to found Discovery Loop. Koray Kavukcuoglu takes over day to day with the title of senior vice president rather than CEO, which tells you which direction the unit is being pulled.

The details:

Why it matters: the company with the deepest research bench in the industry has spent three years unable to convert it, and the people who built the technology have now left to do it somewhere smaller.

My take: the more interesting read is Richard Waters in the FT, who argues the competitive frontier is moving off the model itself and onto the software harnesses that control how agents call on models, plus the domain data used to sharpen them. That is the argument I made back in March in "The Model Doesn't Matter", and it is now the FT's house view. It is also why we keep DealSage model-agnostic and put the work into the layer above: the harness and the data are the durable part, and this week is a reminder that even the people who invented the model layer cannot hold a lead in it.

Read more

The Earnings Catch Up With The Pilots

Two sets of results this week did more for the AI investment case than any survey. Palantir raised full-year guidance again on Monday, and Airbnb hit a four-year high on Friday after putting a number on what its AI assistant has done to the cost base.

The details:

Why it matters: these are audited cost lines and guidance revisions rather than survey responses, and they are the first clean examples of AI showing up somewhere a CFO has to sign it.

My take: Emarketer's Jacob Bourne called Palantir the clearest counterexample to the idea that enterprise AI never scales past pilots. Fair, though Airbnb is the more useful one to study. A 16 per cent cut in support cost per booking is one number attached to one process, which is exactly the shape of result we build towards with clients. Ask what your own portfolio companies could put on a board slide this month.

Read more

Other Interesting Things I've Read or Seen This Week

Bankers are making it rain, with a catch (Axios, August 5) - Investment banking and trading bonuses are projected to rise sharply on the deal boom and the AI investment surge, while automation holds down hiring. (Record bonuses and fewer people to hand them to, which is lovely news if you are already in the building.)

Inside Google's $200bn Wall Street finance machine for Anthropic (FT, August 5) - Private credit, chip leases and data centre guarantees underpinning the capital flowing to Anthropic, with banks separately arranging to offload about $15bn of the debt. (Structured finance has found its new favourite collateral, and it plugs into the wall.)

Thoma Bravo to sell Command Alkon to Francisco Partners (PE Hub, August 7) - The construction software platform trades from one software specialist to another. (Sponsor to sponsor, as is tradition.)

ByteDance is training a model of up to 10 trillion parameters (FT, August 7) - Three times the size of Moonshot's Kimi K3 and above industry estimates for Mythos 5, from a team of about 2,000 that has refused to distil anybody else's models for over a year. (Doing it the hard way, on purpose, at a scale nobody else is attempting.)

Apple opens Mac Siri to Alibaba's Qwen in China (Reuters, August 8) - Chinese Mac users can now route Siri and Writing Tools through Alibaba's models, as Apple defends a 9 per cent share of a PC market where Lenovo has 31 per cent. (The most valuable company in the world is renting its assistant's brain from an ecommerce firm.)

LinkedIn wants users to lean less on AI (WSJ) - The platform has added a "seems like AI slop" report button, is killing its own "enhance your post" feature, and blocks more than 200,000 AI-generated spam comments a day. Pangram Labs puts 41 per cent of long-form LinkedIn posts as entirely AI-generated. (Their editor's list of tells includes the "not this, but that" construction, which I have been deleting from my own drafts for a year.)

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. The whole-business approach in this issue is the work we do: we go through a company unit by unit, find where the value and the bottlenecks actually sit, and build the system that fixes them. If you run a middle market business or a portfolio of them, reply here and I'll show you what the first pass looks like, or have a look at dealsage.ai.