Certainty Is the Product

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.

Silver Lake spent the week circling a take-private of Workday, a company now worth roughly $51 billion after the shares jumped 18 per cent on the report. The potential deal is a very large bet that boring software will be the last to go: Workday is deeply embedded, and the upside from replacing a core employee record is rarely worth the pain.

OpenAI's own pre-IPO sprint looks less orderly. Nearly half a dozen reorganisations, a string of senior departures and the closure of its preparedness team are piling up as Sam Altman tries to prepare the company for a listing that could value it at as much as $1 trillion.

Also inside: a 600-fold corporate AI spending gap, an agent-only RuneScape economy, Anthropic's $6 billion Decart talks, Apple and Alibaba, and the FT on whether AI creates a permanent underclass.

But first, my take on the human quest for certainty, and why it creates such a problem when we use LLMs. Claude can be entirely right within the frame we give it while missing what actually happened.

In today's Acquisition Intelligence:

From The Trenches

The Countdown Clock Problem

When Transport for London put live arrival boards at bus stops, passengers started reporting shorter waits. Perceived waiting time fell from 11.9 minutes to 8.6, and 65 per cent said they had waited less than they used to.

The buses were running exactly as before. Over the same period reliability actually declined slightly, and 64 per cent of passengers said they thought the service had got better.

Nothing about the service had changed. People simply knew how long they were waiting, and knowing turned out to feel almost identical to being served better.

We will pay for that feeling in hard currency. A stated preference study on Dutch railways found travellers willing to give up more than seven minutes of journey time in exchange for certainty about their wait, which is to say they would rather wait longer and know than wait less and wonder.

There is a blunter version in a Nature Communications study from 2016, where researchers ran 45 people through a task with electric shocks at varying odds. A fifty per cent chance of a shock produced more stress, measured in skin conductance and pupil dilation, than a hundred per cent chance. Being sure you are about to be hurt is easier to sit with than not knowing.

Hold that next to how we now work. We have put a machine in front of every analyst that will close any question you bring it, instantly, in fluent prose.

I was reading Rory Sutherland's Alchemy this week and came across a chapter called "Context Is Everything". His argument is that human behaviour makes very little sense when you try to force it into one universal set of rational rules, because the same person will behave differently at work, at home, inside a club or when dealing with a stranger.

Economic decisions are just as context-dependent. Sutherland thinks organisations lose sight of this as they get larger, because conventional logic is self-explanatory and therefore safer for the person making the decision. It is easier to defend a failure produced by sound reasoning than a success case built around something counterintuitive.

A few lines felt very familiar in 2026, for a book published in 2019. Sutherland describes narrow, conventional logic as the natural mode of thinking for a risk-averse executive, for a simple reason: "you can never be fired for being logical". A decision built on familiar reasoning is easy to defend, even when the result is unimaginative.

He then makes the competitive point directly: "logic always gets you to exactly the same place as your competitors". We have explored that theme before in the context of AI, where access to the same models and the same generic prompts produces very similar work.

His other analogy is even more relevant to how firms are adopting AI. Organisations apply conventional logic in social and institutional settings where it has no place, so "we end up using inappropriate software for the operating system, neglecting the psycho-logical approach". A general chat box can be extremely capable and still be the wrong interface for how a firm actually operates, because the firm's real operating system includes its accumulated decisions, relationships and unwritten judgement.

One passage caught my eye because it gets at something deeper than either of those:

That is a very good description of what happens when somebody uploads a reporting pack to Claude and asks what is going on in the business.

Claude will give you an answer. It will probably be clear, well structured and conclusive enough to paste into an email. The answer can be perfectly logical within the four corners of the file and completely wrong about the company.

The Answer Inside The Frame

Sutherland uses a satnav as the analogy. The route can be mathematically correct based on the roads it knows about, while a driver looking out of the window can see that the road is flooded or that a train would get there faster. The satnav is solving the problem it was given inside the frame it was given.

Give Claude an ageing schedule and ask why cash collection slipped, and it might tell you debtor days stretched by eleven. The arithmetic could be right. What it cannot see is that one customer is withholding payment over an installation that went badly, the billing team moved systems in June and dated a batch of invoices wrong, and the sales director agreed longer terms on two contracts to get them signed.

The file creates a clean boundary around the problem. Claude fills that boundary with a clean answer. We mistake neatness for truth.

Why Certainty Wins

We create part of this problem ourselves. We like answers that close the question, particularly in finance, where the work is expected to reconcile and every conclusion eventually has to survive somebody else's scrutiny.

A data-justified answer is safe. You can point to the debtor days, show the movement and explain that collection slipped because terms drifted. Saying the numbers do not explain enough yet, and that you need to speak to the credit controller, leaves the work visibly unfinished.

So we narrow the frame until certainty becomes possible. We choose the data that can be reconciled, turn it into a narrative and call the result correct. The answer may be correct relative to the inputs while missing what actually happened.

AI matches this instinct almost perfectly. Large language models are very good at turning incomplete evidence into coherent prose, often sounding more certain than the evidence deserves, and we generally ask them for a conclusion rather than a list of reasons the question may be badly framed.

Once the collection explanation appears in a portfolio review, it starts to acquire authority. It gets repeated in the board pack and lender update, then becomes the accepted account of the quarter because the original calculation was real and the story around it sounded complete.

The model has helped us do something humans already do all the time. It has taken the evidence we chose, supplied the certainty we wanted and made the narrative easier to defend.

A Different Kind Of Wrong

Hallucination has become a catch-all term for any wrong AI output, though it describes a narrower problem: the model invents a fact, a number or a source. Everything in the receivables example could be real. The debtor days tie to the ledger and the calculation is flawless.

The model is doing what it is designed to do, taking the material provided and producing the most useful answer it can from it. It has no awareness of the disputed installation or the conversation with the credit controller, so it has no reason to tell you either may change the conclusion. You can ask it to challenge its assumptions and it will surface predictable gaps, though it cannot ask about a dispute if nothing in its view suggests one exists.

An experienced finance director at least has a chance to recognise that the ledger cannot explain why the number moved. They know a business exists beyond the file and can decide where to dig next.

This kind of wrong is unusually persuasive because every individual part looks defensible. A fabricated number often gives you something concrete to catch. A real number inside an incomplete explanation can survive several rounds of review.

Bigger context windows do not solve this. Every new model swallows more tokens, which fixes a memory constraint, while the hard questions inside a company stay larger than the prompt and the documents somebody thought to attach: CRM records, customer conversations, pricing decisions, delivery problems, board discussions and the judgement of people who know which exceptions matter.

And there may be no single correct answer. A collection problem can have an accounting explanation and an operational one, with a commercial decision sitting behind both. The useful answer may be a set of causes with different levels of confidence, rather than the clean narrative our board-pack instinct wants to produce.

For people looking for the end state of AI inside a company, I think it is the ability to encode as much of that accumulated context as possible and make it reachable at the moment a decision is being made. That means formal data alongside previous decisions, rejected options, conversations and the unwritten knowledge that currently disappears when somebody leaves.

Once AI can work across those sources, it can surface conclusions that no individual was in a position to reach because the evidence sat across different teams, systems and years. A pricing change and a delivery bottleneck might interact in a way that neither finance nor operations could see from its own view, and the result may be several plausible explanations rather than one definitive cause.

This is one of the design choices we make at DealSage. Sources stay attached, missing or conflicting inputs remain visible, and the system can keep digging across the firm before a plausible explanation hardens into the official one. The aim is to give more of the company's context a voice in the answer, including evidence that changes the question entirely.

Until that is solved, the person in the loop matters more than ever. Their job is to take the answer apart, push into the parts that look thin and feed back the context the model had no way of seeing. Checking the arithmetic is the least of it.

The uncomfortable part is that a good system will sometimes hand you less certainty than the bad one did, and it will feel like a downgrade.

Sutherland calls the alternative a "not-perfect-but-rarely-stupid conclusion", reached by considering a much wider range of factors. That is less satisfying than the unambiguously right answer when somebody wants the board pack finished tonight, but it is closer to how difficult company decisions actually work.

What The Builders Are Saying

Two posts this week, one showing how far apart corporate AI adoption has become and the other showing what happens to an economy once agent labour is effectively free. Follow both accounts if you want to see these arguments before they make it into a conference deck.

@omooretweets (Olivia Moore, partner at a16z)

The post: the median company spends about $12 per employee each month on AI, according to Ramp's AI Index shared by a16z. The top one per cent spend roughly $7,500, more than 600 times as much.

Why this matters: A $12 budget buys somebody a chat subscription. Spending $7,500 per employee means AI has moved into production workflows, API usage and the variable cost base of the business. Those companies are operating the technology in a fundamentally different way.

My take: Spend does not prove return, and I would be very interested to see the P&L attached to the top one per cent. The gap still shows why licence adoption is such a poor measure of progress. For a portfolio company, the useful question is how much real work runs through the technology and which operating line moves as a result.

@maxbittker (Max Bittker, creator of RuneBench and the RS-SDK agent environment)

The post: an agent-only RuneScape server has developed some strange and recognisable economics. Cheap labour makes most commodities abundant, currency inflation pushes the bots towards exchanging goods directly, and resources with a limited respawn rate attract huge swarms because they remain genuinely scarce.

Why this matters: It is a neat simulation of what happens when the marginal cost of execution collapses. Output becomes abundant very quickly, then value moves to the fixed inputs the agents cannot reproduce.

My take: People are over-indexing on agent-led simulations because they give us a clean miniature of the future. RuneScape has fixed rules, enforced scarcity and known respawn rates, so the outcome happens to fit a lot of the current narrative about cheap intelligence and scarce resources. It is a fun experiment, not an investing strategy.

News Digest

Silver Lake Bets On Boring Software

Silver Lake and Workday are discussing a take-private of the enterprise HR and finance software company. Workday's shares rose 18 per cent after Reuters reported the talks, lifting its market value from roughly $43 billion to $51 billion, and Silver Lake may bring in other investors to finance what would rank among the largest software buyouts ever.

The details:

Why it matters: Silver Lake would be betting that the least glamorous enterprise software is also the hardest to displace, because switching risk outweighs most of the upside from a newer interface.

My take: I remain very bearish on mass-market software, but Workday is close to the last category I would expect to disappear. Core HR and ERP are dull, deeply embedded and mostly there to maintain an authoritative record of employees, approvals, payroll inputs and the rest of the machinery. Nobody wants a cleverer employee database enough to volunteer for a painful migration, which makes boring SaaS a reasonable place for Silver Lake to hide from the disruption hitting more replaceable applications.

Read more

OpenAI's Pre-IPO Sprint Gets Messy

OpenAI has reorganised nearly half a dozen times this year as Sam Altman prepares the company for one of the largest IPOs ever attempted. Chief revenue officer Denise Dresser announced her departure this week after less than a year, joining a widening group of senior executives leaving or stepping back while co-founder Greg Brockman takes on more power.

The details:

Why it matters: public investors may soon be asked to underwrite a trillion-dollar AI company while its commercial leadership and internal structure are still being rebuilt.

My take: Nearly half a dozen reorganisations in eight months is the pattern worth underwriting. OpenAI is trying to professionalise ahead of a listing while power concentrates around Altman and Brockman, and its enterprise chief is leaving just as Anthropic pulls ahead with business customers. Revenue growth can carry a lot through an IPO, though public investors at this valuation will eventually want a stable organisation behind it.

Read more

Other Interesting Things I've Read or Seen This Week

Anthropic is in talks to buy Decart for roughly $6 billion (Reuters, August 13) - The chip-optimisation and world-model company was valued at about $4 billion in May, and a deal would be Anthropic's largest known acquisition. (Three months is apparently enough time for another $2 billion of context.)

Apple trains a China-specific AI model with Alibaba (Reuters, August 14) - Apple is reportedly using Alibaba's Qwen to help build a model for the Chinese market as regulation forces a separate AI stack. (Designed in California, retrained in Hangzhou.)

Microsoft prepares its Maia 300 AI chip (Reuters, August 10) - Microsoft plans to unveil the next generation of its in-house chip this autumn after delays left the previous effort trailing Nvidia on efficiency. (The Nvidia independence project remains heavily dependent on buying Nvidia chips in the meantime.)

Could AI create a permanent underclass? (FT, August 9) - The useful version of the labour argument is about expertise, with automation raising the value of workers who know how to direct it while deskilling the people whose judgement it replaces. (A little less cinematic than the San Francisco version, which is probably why it is more convincing.)

Zuckerberg publishes a 6,500-word manifesto for universal superintelligence (WSJ, August 10) - Meta's chief executive argues that widely distributed and open models are safer than concentrating advanced AI inside a handful of institutions. (The metaverse has been excused from the vision this time.)

DeepSeek formally releases V4 Pro (DeepSeek, August 13) - The production release brings a one-million-token context window, agentic coding and API pricing well below the leading closed models. (Somebody had to make a million tokens feel inexpensive just as I finished an essay arguing that more context is not the whole answer.)

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 DealSage around the context a firm has accumulated across its deals, systems and decisions, with every answer traceable to source and every missing input made visible. If you want to see what that looks like on your own data, reply here or have a look at dealsage.ai.