What Consumer Agents Could Do to Business Models

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.

Agents are starting to compare, negotiate and buy on our behalf. Quite a few businesses earn their fees doing that work for us, or benefit when we leave it undone.

Reuters got a look inside Anthropic's confidential IPO prospectus this week. Alongside the growth story sits a harder financing question: how do you support enormous infrastructure commitments when major customers can reduce their spending? Broadcom's agreement to provide up to $42 billion of financing adds another layer to the relationship between the labs and their suppliers.

Accenture, meanwhile, offered some encouragement for anyone buying a services business. Demand for help implementing AI is supporting growth, although clients are pushing for lower prices. Bloomberg also completed its acquisition of Canoe, bringing more of the work behind private-markets data into its business.

Elsewhere: Amex puts acquired software to work, AMD agrees an $8.2 billion deal for World Labs, and the FT looks at whether computing capacity can become a futures market.

But first, it's been the week of the personal agent. I'm still unconvinced by the experience, but following what these tools could do to the businesses underneath it gets much more interesting.

In today's Acquisition Intelligence:

From The Trenches

It's felt like the week of the personal AI agent. With Muse, Instinct and now OpenAI's Dots, the industry is making a concerted push towards software that can get on with things on our behalf.

I'm still fairly bearish on consumer AI in its current form. Another demo of an agent booking a restaurant doesn't do much for me, and I'm not convinced we've found the right way to interact with these products.

That said, the paid market is still small, which helps explain why there's so much attention on winning it. PNC Research data shared by a16z on October 1 put the share of US households paying for an AI subscription at 2.2 per cent in April, more than double a year earlier.

We're still in the early innings, and I doubt this is going to be the final form. I don't think these launches represent the iPhone moment for personal AI agents, largely because of the interface.

For consumers and enterprises alike, we want to see what we're discussing and have something in front of us to question. I suspect that goes some way towards explaining PowerPoint's extraordinary staying power.

For a restaurant booking, a short instruction and a confirmation might be enough. When changing my insurance or moving money, I want to see the options, understand the trade-offs and intervene before something happens.

I suspect voice will help bridge that gap. Talking through a complicated decision while looking at the same information still feels natural to me, and the current apps haven't quite found that combination.

Looking past the restaurant bookings, though, some of the more tangible uses point towards a bigger commercial question.

Cancelling unwanted subscriptions is one of them. It's useful, it's been possible through other services for years, and I doubt it will send personal-agent adoption through the roof on its own.

What it does illustrate is how many businesses operate in small pockets of friction. Some earn money because we leave an arrangement alone; others get paid to find the alternatives and organise the change.

Agents, whether personal or enterprise, could steadily reduce the work involved in both. Thinking through what happens to the businesses underneath those everyday tasks is where this starts to interest me.

The Price Of Leaving Things Alone

Insurance is an obvious place to start. Insurers have to assess and price risk, but customers also face the separate job of finding comparable cover and working out whether they're paying too much.

Brokers and comparison sites have built businesses around helping with that search. Even with their help, the process can be annoying enough that a customer accepts another renewal.

One such example came up on Twitter, where Joe Devoy described uploading his existing car insurance policy to Muse and asking it to find equivalent cover at a better price. He reported a replacement saving $3,500 a year, with the agent buying the new policy and cancelling the old one.

Extrapolate that further and an even bigger example is how banks attract and retain customer deposits. Those deposits are an important source of funding for lending, and keeping that funding cheap is valuable.

Apollo's Torsten Slok raised this in his September 27 post, “Is an Agentic Bank Run Coming?” He asked what would happen if agents routinely moved household cash into higher-paying accounts, threatening the cheap deposits banks rely on.

A bank having to compete harder for those balances could face higher funding costs even without a crisis. The customer gets a better return and the bank has less room to earn its margin.

The important change is how consistently an agent could pursue these opportunities. Given the right permissions, it could keep checking rates and following up on requests long after a person would have decided they had better things to do.

Spread that behaviour across millions of customers and seemingly small inefficiencies become material. Any business that lets customers negotiate a bill through a chatbot should be preparing for a lot more requests to lower it.

Suppliers will adapt, perhaps by restricting discounts or changing contract terms. But the ability to charge more simply because most people never ask becomes harder to rely on.

The cost of switching, particularly the time and effort involved, could fall dramatically. We've been thinking about that largely in terms of software recently, but consumer agents could extend it into the rest of the economy, including our insurance policies and banking relationships.

The Businesses That Do The Shopping For Us

That puts pressure on businesses that benefit from customers staying put. It also brings us to the businesses that have made a good living doing the shopping around on our behalf.

Comparison sites and travel platforms aggregate a fragmented market, make the options easier to assess and help arrange the purchase. Much of that is exactly the work we're now asking agents to do.

Skyscanner earns referral fees and commissions from travel providers, alongside advertising. Booking.com earns commissions on accommodation reservations, with payment handling and other services adding to what it provides.

Booking Holdings, across its brands, generated $26.9 billion of revenue in 2025.

If my agent can compare suppliers and complete the booking, I have less reason to visit a comparison site. The supplier also has an incentive to transact directly if it can offer me a better deal and retain more of the proceeds.

That is where the flight-booking demo starts to become economically interesting. Saving me twenty minutes could change who gets paid for arranging the holiday.

There is still valuable infrastructure here. Skyscanner already supplies travel data through APIs, and an agent might use that service without ever sending me to its website.

Booking's supply relationships and the work of servicing reservations also matter. But a business that loses the customer relationship may have to accept a different role and a different fee.

The risk extends to any intermediary whose main contribution is collecting information and arranging a straightforward transaction. As that work gets cheaper, some of those businesses will need a much better explanation of what customers are paying for.

And If Both Sides Have An Agent?

The same thought takes us a step further into marketplaces. My agent could know I'd sell an unused camera for $400, while yours is looking for that model below $450, and they could work through the terms without either of us spending an evening on messages.

Anthropic's Project Deal experiment already tested this with agents representing 69 employees, negotiating 186 real deals worth just over $4,000. It was a controlled workplace setting, but it makes the idea considerably less abstract.

An open marketplace still needs trust, payments and dispute resolution. Those remain valuable services for eBay to provide as agents take over more of the transaction.

But how much of the commercial internet would we still visit ourselves? If agents can find each other and transact, many of the pages we browse could become unnecessary.

The internet would still connect them, with existing platforms or new providers organising access and establishing trust. Whoever does that could end up controlling a powerful marketplace of their own.

When people have talked about agents becoming the main users of the internet, I've struggled to picture it. Following the progression from checking a bill to comparing suppliers and then negotiating with another agent makes that direction of travel easier to understand.

The potential reach of consumer AI gives those small changes much bigger consequences. That's what I'd be thinking about when assessing businesses built around aggregation and friction.

I still think we're waiting for the personal agent's iPhone moment, and I suspect it will arrive in a form beyond another app on a phone. In the meantime, I'd be looking closely at businesses whose fees depend on work those agents are learning to do.

News Digest

Anthropic's Growth Comes With Commitments

Reuters' September 28 report on Anthropic's confidential IPO prospectus gives investors more detail on the business behind the proposed listing. Revenue reached nearly $4.6 billion in 2025, while the operating loss widened to $8.06 billion.

The reported $42 billion net loss needs care: roughly $34 billion came from revaluing financing that could convert into shares. That is an accounting charge, not cash spent running the business.

The details:

Why it matters: Public investors will have to assess flexible customer spending against much longer infrastructure commitments.

My take: I'd spend more time on that mismatch than the proposed valuation. Broadcom's position as supplier and financier also means the relationships need to be assessed together. A business can have an excellent product and still leave its investors with a difficult financing problem.

Read more

Accenture Has Work. Clients Want A Discount.

Accenture's October 1 results offered some relief for the IT services sector. The company forecast annual growth ahead of expectations as businesses continued to buy help with AI implementation, sending its shares up 22 per cent during trading.

There was a useful qualification: pricing was lower in many areas during the quarter, as clients sought a share of the savings from AI.

The details:

Why it matters: Growing demand for implementation can coexist with pressure on what providers charge for it.

My take: This is a useful check on any services acquisition model that gives the seller all the productivity gains. Clients get a vote in where those gains go. I'd want a clear view of which work earns a fee for the outcome, and which work has historically been priced by how long it takes.

Read more

Bloomberg Buys More Of The Work Behind The Data

Bloomberg completed its acquisition of Canoe Intelligence on October 1. The deal was announced in July, so this week's news is the completion.

Canoe handles the collection and processing of alternative-investment documents, turning the information inside them into structured data. For anyone dealing with private-fund reporting, that is a familiar source of work before the analysis can begin.

The details:

Why it matters: Bloomberg is extending its role in how private-market information becomes usable inside an investment workflow.

My take: This is close to our view at DealSage: the usefulness of the AI depends heavily on the information and processes underneath it. A model can answer a question quickly; the harder work is making sure the relevant information has been captured correctly and stays current. Owning that work gives a business a role that survives a change in the model.

Read more

Other Interesting Things I've Read or Seen This Week

The cheaper AI model can produce the bigger bill (WSJ, October 4). In one example, a model with lower token prices ran up a $14 bill and failed, while a pricier model completed the task for $1. A useful complication for anyone budgeting on the assumption that cheaper tokens mean cheaper work.

California sets rules for lawyers using AI (Reuters, October 1). Newly signed legislation requires lawyers submitting court filings to personally verify citations and disclose AI use. The drafting may get cheaper; someone still has to make sure the cases exist.

Amex puts its software acquisitions to work (WSJ, September 30). Its new corporate card integrates expense-management software following the Center and Hyper acquisitions, with more AI agents and accounts-payable features planned. Buying the capability was the easy part; now finance teams get to use it.

Synopsys finds a way to charge OpenAI (Reuters, September 30). OpenAI will pay a training subscription for a specialist chip-design model, then share customer revenue with Synopsys while conventional tools check the output. The software incumbent has found several places to send an invoice.

AMD agrees to buy World Labs for $8.2 billion (WSJ, September 28). The all-stock deal brings Fei-Fei Li's spatial-AI research company into the chipmaker, with closing expected by year-end. Knowing what the next models need seems a reasonable thing for a chipmaker to want to own.

Meta starts an enterprise AI business (WSJ, September 28). Former MongoDB CEO Chirantan Desai will lead a business bringing together Meta's agents, coding tools and model access. Another supplier for the procurement team to assess, once someone explains the pricing.

OpenAI shelves GPT-6.1 Astra (Reuters, September 28). The planned October release failed internal safety and alignment checks, including concerns about staying within authorisation and accurately reporting its actions. A useful reminder that doing more work and doing the work you asked for are separate product requirements.

Amazon explores an $8 billion chip financing (Reuters, October 2, reporting the FT). The proposed sale-and-leaseback would bring outside investors into financing Nvidia hardware. Last week's infrastructure bill is still looking for somewhere to sit.

The FT examines futures for AI compute (October 2). Marc Rubinstein looks at CME's proposed contracts and the difficulties of standardising computing capacity across chips and configurations. Hedging your AI bill would be useful, provided the hedge still resembles what you need to buy.

Visa opens up part of its AI cyber defence (Reuters, September 29). The payments company has released defensive software after AI exposed vulnerabilities, with its technology president describing the experience as humbling. A more useful word than the ones usually found in technology announcements.

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 help firms put AI to work with their own information and workflows. If you're assessing where AI changes the economics of a business you're buying or already own, reply and tell me what you're looking at.