PowerPoint Is Going To Die

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.

Last week's issue was Karp calling token pricing a wealth tax. Seven days on, the sovereignty debate has carried straight on and picked up its biggest voice yet: Satya Nadella posted an essay this morning arguing that in the AI age you pay for intelligence twice, once with money and again with the proprietary knowledge you hand over to make it useful. He even reaches for the same Karp quote this newsletter led with last week: firms want to own their compute, their data stack, and their alpha.

Delphi's Tommy Shaughnessy published twelve-month predictions pointing the same way, token prices halving while enterprises take ownership of their own stacks. More of the sharpest thinking now lands on X first, so the Builders section below is becoming a permanent fixture.

In the news: JPMorgan built agents that beat a 60/40 portfolio across twenty years of backtests, and CVC ran a sale process with a chatbot analyst instead of bankers. Plus Mercor at $20bn, Grok 4.5's price war, Carlyle's fivefold data-centre exit, and Meta's four-day image feature.

But first, my take on a question I now get on almost every call: where should the work actually happen? My answer tends to annoy people, so this week I've written it down properly.

In today's Acquisition Intelligence:

From The Trenches

The Front Door Is Disposable

On a call on Thursday, someone asked me whether I still track my to-dos with pen and paper. He does, slightly sheepishly, and so do I. When something needs actual thought, I write it down, because writing it down is the thinking.

It's one version of a question we face on nearly every DealSage call: where should this work actually happen? Should the memo be drafted in the platform or in a chat window, should the model live in Excel or behind an agent, should meeting notes be queried where they were taken or somewhere central. My answer is always the same and always slightly unsatisfying: it depends, and I just look at what I'm actually doing.

So, honestly, here is where my work started this week. A notebook. Voice-to-text in the car, dictating ideas to an agent that files them where they belong. Email. Claude when I wanted to think something through, Claude Code when I was building, Codex once, for a second opinion. Agents inside Notion. Agents inside our own platform. And far more speaking than I'd have admitted to a year ago: a growing share of what I "write" now starts as me talking, through Wispr (not a paid promotion), because speech is faster than typing and the agents on the other end don't care which one you used. Some things just feel right in a particular tool, and the feeling shifts with the task: sometimes I query my meeting transcripts inside Granola, and sometimes I pull the same transcripts into Claude and interrogate them there. That flexibility is the good part, and I'd encourage anyone to lean into it rather than hunt for the one true app.

The reason the flexibility works is that every one of those doors pulls from the same underlying record. I never open a session by re-explaining who we are, what the deal is, or what was agreed last week, because the context is already there whichever door I walk through, and everything that matters lands back in the same place: tasks into the platform tagged to whoever owns them, ideas into a database, deal work onto the deal record. The front doors are disposable. I could swap any of them tomorrow and nothing downstream would notice.

This is exactly how we design and deploy DealSage too. Where the agent loop runs matters far less than what it can reach, so everything is built around structured access to the deal data underneath, and the surfaces on top, chat, email, scheduled agents, whichever model is doing the thinking, stay swappable by design.

Skills Are The New Consistency

There's a second thing that makes the flexibility work, and it took me longer to see. When agents produce most of the output, the thing you actually need to standardise is how you like things done: how a memo should read, what a model must always include, which numbers get checked and against what. These have a name now, skills, and it's fast becoming common vernacular across the AI tools: your way of doing something written down, with examples to pull from, in a form any agent can follow.

Once that exists, it stops mattering which surface does the producing, because the same skill gives you the same memo whether it ran from a chat window, an agent inside the platform, or a scheduled job overnight. Consistency used to live in a senior person's head and a house-style deck nobody opened. Now it travels with the work.

Orchestrate Here, Output There

Each surface has its own feel, and I've stopped fighting that. Claude's consumer app is lovely for artifacts and quick visual things, and frustrating the moment you go deep. Claude Code is brilliant for running several agent threads at once and hopeless when you want to actually look at something. The adjustment I'd push on anyone is getting comfortable with the work product materialising somewhere other than where you're orchestrating from: you direct from one window, and the deliverable lands in the app, the store, the deal record, wherever it lives. I still love a good UI, and a well-designed app or store to see your work in matters as much as it ever did. The platforms that age well from here will be the open ones, built so you can orchestrate from wherever you like while the work stays visible, organised and shared in one place. Viewing and producing are simply becoming different activities.

PowerPoint Is Going To Die

One other thing I've noticed over the past few months: how little I use PowerPoint, which borders on sacrilege for a former investment banker. I produce more materials now than I ever did in banking, and I cannot remember the last time I opened it. Everything I make is HTML, because HTML is what agents write natively, it renders anywhere, and it doesn't fight the tool producing it.

So here's a prediction you can hold me to: PowerPoint dies out. It survives on the muscle memory of people assembling slides by hand, and the number of people assembling anything by hand falls every quarter. Hear it here first.

The wider point is that the nature of work keeps evolving, and specifically the means and channels through which it happens: typed to spoken, assembled to generated, opened-in-an-app to delivered-by-an-agent. Betting on any particular channel is a losing game. Betting on the record underneath them all is the same bet every time.

Starbucks Reached The Same Conclusion

Bloomberg reported on Thursday that Starbucks is building in-house replacements for its Microsoft inventory system and its IBM maintenance platform, written with AI-assisted coding, as part of a plan to cut $2bn in costs. The company spends around $400m a year on software. A viral X post declared this the death of enterprise software and took a swing at IBM and Salesforce's share prices on the way through; the internal deck is narrower than that, two systems, rolling out through next year if testing goes well.

The direction matters more than the scope, though. Starbucks has been paying $400m a year largely for other people's opinions about how its own operations should work. Once AI can generate the code, the thing worth owning is the data and the processes underneath it, and the application layer becomes a preference.

I still like the notebook, for what it's worth. It's the one front door that never gets a software update.

What The Builders Are Saying

More of the interesting arguments show up on X before they show up anywhere else, so this section is here to stay. Four posts from this week that circle the same layer, from people building on it. Worth following all four.

@satyanadella (Satya Nadella, CEO of Microsoft) on the Reverse Information Paradox

Posted this morning and already past two million views. His argument: in the AI age you pay for intelligence twice, once with money, and again with the proprietary knowledge you have to reveal to make the intelligence useful. Models learn from your "exhaust", the prompts, the tool calls, and especially the corrections, and that learning flows in one direction, so value converges to whoever owns the learning infrastructure. His prescription is a hard trust boundary where your data, traces, evals and memory compound, and an orchestration layer decoupled from any single model.

Why this matters: anyone who read last week's issue will recognise the Karp quote Nadella reaches for, about firms wanting to own their compute, their data stack, and their alpha. When the CEO of Microsoft and the CEO of Palantir are making the same argument a week apart, it has stopped being a hot take. His test is one every fund can run today: if the model you use most was taken away tomorrow, does your capability survive it?

@levie (Aaron Levie, CEO of Box) on the battle for context

His argument: agent effectiveness comes down to domain expertise, the right context and tools, and fitting into workflows people can actually review. The applied AI layer organises critical domain knowledge and governs who gets access to it, and it improves the agent's context over time.

Why this matters: for a fund, "context" has a specific meaning. It's your deal history, your diligence questions, your pricing decisions. Levie is describing where the moat moved.

@Shaughnessy119 (Tommy Shaughnessy, Delphi Ventures) on the next twelve months

His predictions: token prices roughly halve while usage more than doubles, most enterprise usage shifts to open-source models at a tenth of the cost, and the cultural shift is enterprises owning their AI stack outright: models, data, memory, proprietary context, evals.

Why this matters: if the model layer keeps deflating, the durable asset inside your firm is the context you feed it. Renting intelligence gets cheaper every quarter. Your proprietary record doesn't exist unless you build it.

@PatrickOjo_ on why enterprise AI pilots fail to scale

His point: the most common killer has nothing to do with the agent. The data it needs sits in systems owned by different departments, with different governance policies and no shared definition of a clean record. Pilots succeed in a sandbox and die in the org chart.

My take: all four are describing the landing zone problem from a different seat. The model gets cheaper, context gets more valuable, governance decides who's even allowed to play, and now Microsoft says the learning loop itself is the thing to own. Deal teams that treat their data layer as an afterthought will find the agents they buy in 2027 have nothing worth reading.

News Digest

JPMorgan Puts Agents In Charge Of The Money

Banks have used models to advise on allocation for decades. What JPMorgan's researchers just tested is a different category: AI agents that make the allocation themselves. A team led by Thomas Salopek built an array of AI-powered investing agents that shift between stocks and bonds as market conditions change, and Bloomberg reported the results on Thursday. The headline number: the best system beat the classic 60/40 portfolio, with less volatility, over two decades of backtests.

The details:

Why it matters: this is a major bank publishing evidence that agents beat its own house methodology at the job the industry considers sacred. Allocation was supposed to be the last thing to automate.

My take: backtests flatter everything; twenty years of hindsight has never lost anyone money. The real story is that a bank is now comfortable saying agents made the decisions end to end, which makes this a governance milestone more than a performance one. Expect every large allocator to publish a version of this within a year, and expect the differentiator to be the data each firm's agents are allowed to see. As I argued above, the agent is just another front door. The record it runs on is the edge.

Read more

CVC Ran A Sale Process With A Chatbot Analyst

When CVC put Skroutz, the Greek e-commerce business, up for sale earlier this year, it skipped the investment bankers. Private Equity News reported on Wednesday that prospective buyers were sent a link to a data portal that stood in for the investment memo, with a chatbot "analyst" answering their questions on financials and due diligence, and prompting serious parties to book time with management directly.

The details:

Why it matters: one of the largest buyout firms in the world just automated the Q&A layer of a live sale process, and the process still worked.

My take: most of the sell-side question queue was always just structured data access, and CVC has now proved it on a live deal. The judgment parts of banking, the positioning, the price tension, the negotiation, stayed human here and will for a while. The implication runs the other way too: if the sell side's data layer answers in seconds, a buy-side team still compiling question lists in Word is setting the pace of someone else's process. Whoever holds the cleaner structured record sets the tempo.

Read more

Other Interesting Things I've Read or Seen This Week:

Carlyle sells Copia to EQT for $2.6bn (July 8) - The data-centre power platform returns roughly five times Carlyle's money. (The picks-and-shovels trade is alive and well, and the shovels are substations.)

Mercor in talks at a $20bn valuation (July 9) - The training-data startup has told investors it holds at least one term sheet at that number. (Selling expert judgment to the labs by the hour remains this cycle's least glamorous goldmine.)

SpaceXAI ships Grok 4.5 (July 8) - "Opus-class, but faster and cheaper" per Musk, at $2 per million input tokens against Opus 4.8's $5. (Shaughnessy's price-halving prediction, above, is running ahead of schedule.)

OpenAI launches GPT-Live (July 8) - Voice models that listen and speak at the same time. (At last, an AI that interrupts like a real managing director.)

Meta pulls Muse Image after four days (July 10) - The auto-opted-in image generator built on public Instagram accounts lasted Tuesday to Friday before SAG-AFTRA saw it off. (Move fast and break things, then apologise to the actors' union.)

The AI CEO who kept dealmaking after a secret fraud plea (July 8) - Prosecutors unsealed papers showing a founder pleaded guilty in 2025 and carried on raising capital and announcing partnerships into 2026. (Diligence: still undefeated, on the rare occasions anyone does it.)

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're building the place deal work lands: whichever front door it arrives through, chat, email, an agent or a spreadsheet, it ends up as structured, traceable deal intelligence your whole team can use, whichever model happens to be in fashion that week. Reply here or have a look at dealsage.io.