Earlier this month, some of the most powerful clients in corporate law sent their lawyers a blunt message. Per the Financial Times, Goldman Sachs, Morgan Stanley and Citi have been pressing their outside law firms to cut fees, arguing that AI has made routine legal work faster and the bills should reflect it.
Citi alone has reportedly started asking firms bidding for its business to disclose how much they’re saving with AI, while Morgan Stanley plans to put most of its outside legal work out to competitive bid, with fixed fees on the table.
They’re not the first to try this. In February, KPMG International pressed its own auditor, Grant Thornton, to pass on its AI savings, and the fee fell 14%. Grant Thornton, however, said its fees reflect the cost of both its people and the technology that supports them.
Most founders don’t have Goldman’s negotiating leverage, but they’re paying for the same kind of work. Ask a founder where AI will save them money and you’ll usually hear about software: a cheaper CRM, a coding assistant, a tool that writes the first draft of the newsletter.
And while those savings are real, they’re small, precisely because software was never the biggest expense for most companies. The bigger number sits a few rows down the P&L, in the services column.
That’s where you find the agency on retainer for sales decks and pitch materials, and the answering service that covers the phones after 5 p.m. You’ll also find the staffing firm that sends a team of contract reviewers every time a legal matter lands, billing by the hour until the job is done.
These costs are easy to accept because they feel like the price of doing business. They’re also hard to negotiate, since you’re paying for people’s time and there’s rarely a clean way to measure what that time produced.
Venture firm Foundation Capital thinks this is where the real AI story is, estimating that businesses spend $4.6 trillion USD a year on salaries and services that AI could reshape. What’s more: roughly half of it goes to outsourced IT and business process work, based on Gartner figures.
For comparison, the firm puts the entire SaaS market at around $200 billion. Foundation Capital’s shorthand for the shift is “service as software,” meaning the AI does the work rather than helping a person do it, and the seller takes responsibility for the outcome.
Most of the debate about AI and employment frames this as a question of which jobs disappear. For anyone running a startup, a more useful version is which invoices shrink first, and what you should still be willing to pay full price for.
The companies building in this space offer a clue; most aren’t even trying to remove people entirely. They’re moving the human to a different point in the process, usually the point where someone has to make a call and stand behind it.
Why the middle layer goes first
Outsourced services have a cost structure that makes them especially exposed, with them being built almost entirely on labor, and much of that labor being repetitive.
Contact centers show how quickly that’s being tested. A Gartner survey of service leaders this spring found AI spending up 38% while overall service budgets grew just 2%, with leaders pulling money away from labor and overhead to pay for it.
And the outsourcers feel it, too. One executive at TaskUs, in fact, told CX Dive on September 28 that clients are driving roughly a 7% to 10% reduction in call volumes each year through automation, nearly all of it now powered by AI.
When most of a bill is handling high-volume, rules-based tasks, even partial automation changes the math quickly. In 2022, Gartner predicted conversational AI would cut contact center labor costs by $80 billion in 2026, while expecting only about one in ten interactions to be fully automated.
In other words, the savings were never expected to come mainly from replacing agents, but from stripping out the routine volume surrounding the conversations people actually need to have.
Small businesses feel a version of this every day, usually through the phone. For a plumber, a veterinary clinic or a small law firm, it’s still where new business arrives, and the people running those businesses are usually busy doing the actual work when it rings; a dentist in the middle of a procedure can’t pick up, and neither can a contractor on a roof.
The cost of that delay is higher than most stakeholders expect. When researchers audited 2,241 U.S. companies for Harvard Business Review, they found that firms contacting a lead within an hour were nearly seven times more likely to have a meaningful conversation with a decision maker than firms that waited even an hour longer. Almost a quarter of the companies in the audit never responded at all.
Traditionally, small businesses had two options: send callers to voicemail, which many people simply hang up on, or pay for a human answering service, which can cost hundreds or thousands of dollars a month and often knows little about the business.
That second option is now competing with a wave of AI receptionists, from startups like Upfirst to features built into larger phone platforms, that answer, book appointments and hand off the calls that need a person for a small fraction of the price.
The more interesting shift is in what the owner is actually paying for. The routine calls get handled, and the owner’s attention goes to the handful of callers who genuinely need them. What a small business is really buying is responsiveness, which used to be something only larger firms with a front desk could afford.
In high-stakes work, the human moves to the end of the line
The picture gets more interesting in fields where a mistake carries legal or regulatory consequences. Here, AI takes on the first pass through the work, and the experts who remain become more concentrated and, arguably, more valuable.
Legal document review is a good example. When a company gets sued or investigated, someone has to read through thousands or even millions of documents to decide what’s relevant, what’s privileged and what contains sensitive personal data.
A RAND study of large corporate cases found that review made up 73% of eDiscovery production costs, and most of that work went to outside counsel. The same research found that human reviewers are highly inconsistent, and that computer-assisted review found at least as many relevant documents as traditional eyes-on review.
An entire industry still grew up around supplying reviewers. Alternative legal service providers, which handle work like review outside traditional firms, reached an estimated $28.5 billion USD market by 2023, though Thomson Reuters found that some law firms expect their own AI capabilities to eventually reduce how much they rely on those providers.
That market is now splitting in two directions at once, sometimes inside the same company. Legal tech firm Altorney, for instance, runs an AI platform that categorizes documents for relevance, privilege and sensitive data before a human reviewer opens a file, alongside a marketplace where firms hire those reviewers directly rather than through a staffing agency.
It’s a useful snapshot to where the industry is heading: the machine takes the first pass across the entire pile, and the lawyers spend their time on the documents that matter to the case, which is the part of the job that needed a lawyer in the first place.
It’s also the kind of breakdown Citi is now asking its firms to put on paper.
The agency gets smaller… and smarter
Biopharma follows the same logic as legal review, with even less room for error. A presentation for physicians or a national sales meeting has to match the clinical data exactly and meet regulatory rules on what can be claimed.
Much of that production work has historically gone to specialist agencies. McKinsey has noted that pharma’s creative and production process is almost completely outsourced, and it expects agencies to concentrate on what they do best, such as creative strategy and media planning.
The risks of rushing are just as real. The consulting firm describes one life sciences company that spent months building an external-facing generative AI tool, only to pull the launch because its medical and legal teams raised risk concerns after the fact, for example.
That tension is on the agenda this week in Philadelphia, where Prezent is hosting Articulate 2026, a summit for medical affairs and communications leaders from companies including Pfizer, Novartis and Sanofi.
Panels cover how AI strengthens or weakens scientific communication and where biopharma organizations are resisting it, and the day closes with a conversation about what the company calls “the rise of the Neo Agency.”
The term is the company’s own, describing a model that brings together AI, scientific expertise and human expertise to produce life sciences communications, but the idea behind it is one all entrepreneurs will recognize.
A traditional agency grows by adding people: account managers, strategists and junior staff who produce the drafts, with clients paying through hourly rates or retainers. In the newer model, software handles the first drafts at volume, and a smaller group of experts checks the science, the compliance and the story.
The pricing shifts with it, toward fixed-price projects where the client pays for a finished deliverable rather than the hours behind it, which is exactly what Goldman, Morgan Stanley and Citi are now asking of their law firms.
Paying for judgment, not volume
Founders don’t need to work in pharma or law to act on this. Start with the vendor list and go through it line by line, asking one question about each: am I paying for judgment, or am I paying for volume?
Volume is the answered call, the first read through a stack of contracts, the tenth version of a slide template. That work is getting cheaper quickly, and it’s fair to expect vendors to reflect that in their pricing, or to replace them with tools that do.
Simply put, the banks are asking their law firms to show the math, and there’s no reason a startup can’t ask its agency or answering service the same thing at renewal time.
But judgment is harder to automate and easier to underprice; it’s the lawyer deciding what a document means for the case, the medical director signing off on a claim, or the owner calling back the customer who’s genuinely upset.
As a matter of fact, Grant Thornton’s pushback to KPMG makes the same point from the vendor’s side; the expertise still costs money, even when the grunt work doesn’t.
The trap is treating AI as a way to cut the whole services column at once. Accountability doesn’t transfer to software just because the work does, and the companies that learned this the hard way usually did so in public.
A better approach is to strip out the volume first, then spend some of the savings on the people who make the calls. Over the next few years, the companies that get ahead will probably be the ones that stop paying premium rates for routine work and start paying properly for the moments that need a human.
Featured image: Resource Database via Unsplash+
