Your People Were Never The Bottleneck

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.

Salesforce shares rose 20 per cent after Thursday's earnings, up around 65 per cent since June. The SaaS panic is easing, and the model replacing it is a strange one: Salesforce, Workday and SAP are stripping the interface out of their products so agents can work directly against the data.

At the model layer, Anthropic's most capable model has stalled at around 11 per cent of enterprise spending on its tools, while the cheaper model released after it has already moved ahead. OpenAI is also proposing to cut Cursor off on 12 November, three months after SpaceX completed its $60 billion acquisition of the company.

Elsewhere: Apollo disclosed a social-engineering breach, Andreessen Horowitz raised $1.1 billion for AI infrastructure, Anthropic connected agents to laboratory equipment, and Morgan Stanley found more AI spending hiding off balance sheet.

But first, my take on why handing everyone Claude has made plenty of employees faster without producing the company-wide uplift management expected.

In today's Acquisition Intelligence:

From The Trenches

The Productivity Gain Is Stuck in the Queue

I was talking to a friend recently who works at a law firm that has rolled out Claude to everyone on the team.

He thinks it is great. Of course he does. He is fairly senior, works in a niche area and has no dedicated junior support, so help with drafting, research and the lower-value parts of the job already feels like a considerable step up.

It was making him quicker. I asked whether the firm would see that in the amount of work he took on.

"Fifty per cent of the productivity gain I'm getting here, I keep for me," he said. Then he thought about it for a second. "Maybe even a bit more."

He had precisely zero appetite for turning every hour Claude saved into another hour of billable work, and who can blame him? Whilst it doesn't entirely explain the situation, I do think it's one factor in why rolling Claude out and expecting miracles is foolish.

The firm's output only changes when that employee's speed was the thing holding the work up. In most companies, it wasn't.

The Gain Already Belongs to Someone

Management sees spare capacity waiting to be collected. Give 1,000 people a tool that saves each of them five hours a week, and somewhere a spreadsheet says the business has created 5,000 productive hours.

Employees who have spent years asking for more support or fewer internal meetings see a small correction to a job that had become unnecessarily difficult. Some saved time improves the work, and some becomes a proper lunch or an earlier evening.

If every efficiency an employee reveals earns them more work, keeping part of the gain is entirely rational. Nobody at the firm ever discussed who the saved time belonged to, so my friend decided for himself.

He reached the same place from the other direction when I asked whether it meant taking on another client. Claude helps, but it doesn't help enough for that.

Another client means another set of facts to hold in his head, another relationship to manage and one more thing that might become urgent on Friday afternoon. Claude reduces the admin around a matter, while the matter still occupies a slot in his brain.

Saved minutes across drafting, document review and call preparation do not combine neatly into capacity for another relationship. Judgement, responsibility and constant context switching determine how many clients he can carry.

I've seen this play out at a number of companies now, and management's initial reaction is usually some version of if only I could fix my employees. I think that's an unfair assessment. You need to dig a bit deeper and really understand where the bottlenecks live and why.

The Company Runs at the Speed of the Queue

Most work inside a company involves waiting for someone else: getting them on a call, collecting their numbers or waiting for a partner to decide. Claude can prepare the meeting and draft the follow-up. It cannot make Thursday's call happen on Tuesday.

Take a deal team reviewing a new opportunity. Claude cuts the first read of the CIM from half a day to forty minutes and produces the screening memo before lunch, then the deal waits for a partner, management's answers and the next IC slot.

The analyst feels a large productivity gain. The deal reaches the same decision on the same Friday.

A company is a system of dependencies and decisions, while a seat-based rollout treats it as a collection of individual jobs. Everyone reaches the next dependency faster, then sits in the same queue as before.

Meta Generated More Work

Reuters published an investigation this week into Meta's attempt to rebuild itself around AI. Internally it was called Project OT, short for Organization Transformation.

The plan imagined virtual workers taking over much of the daily work, overseen by smaller groups of what Meta called "talent-dense" staff. Its original pilot used five small pods of two or three engineers and a designer, removed layers of middle management and gave each pod a single senior unit head.

That was the sensible part: give a handful of capable people the context, tools and authority to own an outcome from beginning to end.

At scale, the numbers moved in very different directions. AI helped employees produce 220 per cent more changes to Meta's internal software platforms and infrastructure year on year. New or upgraded features reaching users rose only 36 per cent.

Meanwhile, major technical and security incidents increased 40 per cent, according to the internal posts reviewed by Reuters. The time employees spent firefighting them rose 70 per cent.

Meta generated vastly more code, fed it into the same complicated machine and created extra work downstream. The constraint had moved from writing code to deciding what should ship, integrating it safely and dealing with what broke.

This is why founders and small teams can report something much more dramatic. Some days AI lets me do ten or twenty times what I could before because I can research an idea, build it and ship it without booking six calls or waiting for a committee. I own the queue.

Finding Where The Work Actually Stops

Most of the value here comes from correctly identifying where work stops, which sounds obvious and turns out to be the hard part.

Since we started doing AI implementation at the operating company level, this has become a core part of what we do at DealSage. It is far less technical than people expect. Mostly it is sitting with people at every level of the business and asking what slows them down.

You get very different answers depending on who you ask and where they sit. Management usually says visibility, because visibility is what management lacks. Ask the engineer on site and he tells you the job form takes twenty minutes, won't submit from a phone in a basement with no signal, and so nothing gets invoiced until he is back at the depot on Friday.

Sales leadership says the problem is lead volume. The rep says every quote above a certain size waits on one person who is in meetings all day. The CFO says month-end close is too slow, and the controller says two subsidiaries send their numbers in a different format and somebody rekeys them by hand.

None of those are fixed by a Claude license.

What you are hunting for is the chains of translation: every point where information changes hands, or format, or system, and loses a day and some of its meaning on the way. Shorten those chains and the business speeds up, because the time comes out of the gaps rather than out of the typing.

Giving everyone Claude is still sensible. It makes tedious work less tedious, gives unsupported people something resembling junior help and raises the baseline quality of everyday output.

What it will not do is deliver a 20 per cent efficiency gain to the firm because everyone now has a licence. A value creation plan built on that has underwritten the gain at the employee level, and that gain leaves the building at six o'clock, which is roughly what my friend told me he does with his. The gains that reach the P&L come from removing an approval, collapsing a handoff between two teams, or making one team's data reachable by another without a request and a two-week wait.

The individual gains are already here. My friend is quicker, happier and finishing earlier, while his firm's output looks much as it did in March.

Everyone involved would probably describe the rollout as a success. The value creation plan still has to explain how time saved by an employee becomes output for the firm.

What The Builders Are Saying

Two posts worth reading this week.

@firesidealpha (Bill Gates clip, August 27)

The post: Gates says historical analogies are liable to mislead us on AI. The microprocessor, PC and internet created jobs on balance, but each remained a tool inside a field of human capability. He is staking his reputation on general machine cognition, followed by cheap humanoid robotics, producing a different result.

Why this matters: this is the strongest version of the argument against treating AI as one more familiar technology cycle, made by someone who lived through the three cycles everyone cites.

My take: Gates is talking his book, and the robotics timeline is doing quite a lot of work. Still, his direction of error feels right. Previous technology cycles rewarded the continuity case, which is precisely why so many firms are assuming this one will as well.

@samzliu (Sam Z Liu, August 25)

The post: Liu argues that agent traces sit somewhere between logs and business records. Few people will read them, but they contain the agent's reasoning and work product. Standard monitoring software treats them as exhaust when they need the persistence and ownership of a CRM record.

Why this matters: once an agent screens a CIM or runs a sourcing pass, the firm needs an answer to where that work lives and who can use it next.

My take: this is our thesis at DealSage. Agent work belongs on the deal from the moment it is created, alongside the sources and decisions that shaped it. Parking it in a log only recreates the coordination problem from the FTT in a newer system.

News Digest

Enterprise Software Loses Its Interface

Salesforce's 20 per cent rise after Thursday's results took its recovery since June to roughly 65 per cent, and Richard Waters in the FT thinks the worst of the software sell-off has passed. The feared outcome, customers replacing enterprise applications with things they build themselves, looks less likely. A stranger version is arriving instead: Salesforce, Workday and SAP are opening their products so agents can use the underlying data without touching the interface.

The details:

Why it matters: enterprise applications may survive as systems of record while losing the interfaces and seat economics that supported their valuations.

My take: going headless gives agents direct access to whatever is already in the system. In many CRMs, that means stale stages, thin contact records and meeting notes abandoned months ago. The diligence question is whether the target holds context a competitor would have to reconstruct, because an interface-free database full of neglected fields is still a database full of neglected fields.

Read more

Anthropic's Frontier Model Loses the Buyer

Anthropic may attempt a $2 trillion-plus IPO as soon as September, but the FT reports that enterprise spending on Fable 5 has flattened at about 11 per cent of total Anthropic usage. Ramp's data covers 70,000 companies and shows the cheaper Opus 5, released later in July, already moving ahead. Corporate buyers that once defaulted to maximum capability have started buying enough instead.

The details:

Why it matters: the labs are funding frontier development on the assumption that superior capability wins enterprise spend, and current buying behaviour points elsewhere.

My take: I wrote an issue in March called The Model Doesn't Matter, and I would write it again: capability stopped being the main constraint for most corporate work some time ago. What matters now is the system around the model, including the context it can reach, the actions it can take and the review it receives. Vendors whose advantage rests on access to the cleverest model need an answer for a market where the cheaper one clears nearly every task.

Read more

Cursor Learns About Supplier Concentration

OpenAI told Cursor on August 28 that it plans to end their model agreement on 12 November. Cursor became part of SpaceX this month after a $60 billion all-stock acquisition of parent company Anysphere. OpenAI says it cannot rely on a Musk-owned company respecting its terms, pointing to X's conduct after Musk bought Twitter.

The details:

Why it matters: an AI company's change of control can cost it access to the models on which its product was built.

My take: model access belongs in the supplier-risk section of diligence. Ask what the agreement says about a change of control, how much notice the provider owes and what moving to another model requires. Cursor can route around OpenAI; a target built around one provider may discover that its most important supplier also wants its market.

Read more

Other Interesting Things I've Read or Seen This Week

Apollo reports a breach caused by social engineering (Bloomberg, August 21) - Attackers impersonating IT staff reached Apollo's cloud systems and took personal information. (A trillion dollars under management, defeated by somebody sounding helpful on the phone.)

Andreessen Horowitz raises $1.1 billion for AI infrastructure (Bloomberg, August 28) - A dedicated fund for the tools and systems beneath the models. (The picks-and-shovels trade now has its own limited-partnership agreement.)

OpenAI becomes the sole investor in its latest venture fund (WSJ Pro) - The new vehicle has no outside limited partners. (This should shorten the annual meeting.)

The model harness gets a name (WSJ CIO Journal) - Tools, instructions, context and subagents around the model increasingly determine whether an agent works and what it costs. (Readers of The Model Doesn't Matter may recognise the plot.)

Morgan Stanley finds more AI spending off balance sheet (Axios, August 27) - Data-centre leases and purchase commitments make the build-out larger and more leveraged than headline capex implies. (The footnotes have started buying GPUs.)

Anthropic gives agents control of laboratory equipment (Reuters, August 27) - Its Model Hardware Standard has already driven microscopes, robotic arms and plate readers in a Genentech protein assay. (USB-C for laboratories, with slightly higher stakes.)

OpenAI took a week to notice its models had breached Hugging Face (FT) - Autonomous models exploited vulnerabilities and reached another company's systems before OpenAI detected it. (A strong quarter for observability vendors.)

Meta criticises Anthropic while spending heavily with it (NYT, August 27) - Internal projections had Meta paying Anthropic as much as $10 billion a year while executives attacked it in public. (Procurement and communications remain in separate queues.)

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 start with where the work stops, then structure the deal history, contacts, operating data and investment decisions so agents can carry context across the firm. If you want to see what that looks like on one of your workflows, reply here or have a look at dealsage.ai.