Profits at the Periphery

What if AI doesn’t need to replace an industry’s core offering to undermine its economics?

There’s a loud conversation online about an impending ‘SaaS-pocalypse’, in which AI makes software so inexpensive to develop that businesses simply stop buying it. Why continue paying substantial subscription fees for something you can increasingly make yourself with a few prompts and the right harness?

I’m skeptical because I think this confuses the cost of producing software with the value of purchasing it. Open source has made extraordinary amounts of software available at little or no licensing cost, and in some parts of the industry it’s displaced proprietary alternatives. Yet businesses continue purchasing commercial software because they’re paying for everything that comes after the code: hosting, security, upgrades, scaling, and the engineers it takes to keep a system running, with those costs shared across many customers. An individual business can benefit from that collective investment without assuming the full responsibility itself. A business that builds its own tool takes all of it on, along with the job of hiring people to maintain something disconnected from what it sells. AI makes getting to code cheaper. It does much less for the five years after.

That burden grows with what you build. A customer who relies on most of a service would have to rebuild most of it. A customer who uses one corner only has to replace that corner. AI also makes it easier for new providers to bring narrower, cheaper offerings to market, without asking customers to own software themselves.

So I see a less dramatic, more consequential threat inside the SaaS-pocalypse argument. Software businesses stand to lose the customers who were never particularly great fits, and my hypothesis is that those customers are disproportionately profitable. Switching costs may remain stubbornly high, and if they do, I’m wrong. The pattern also reaches beyond software, to any business whose customers buy a shared capability because getting it themselves was too much trouble. That includes my own.

The customers at the edge

Consider a SaaS company generating $100 million in annual revenue. Some customers are exceptionally well matched to its product. They use a breadth of its capabilities, have integrated it into their operations, and value what makes it distinctive. They’re often the reason the service was developed or extended.

Others have a less complete relationship with the offering. One uses an elaborate analytics platform for a handful of reports; another, an enterprise content management system for a straightforward publishing workflow. They aren’t dissatisfied. Buying an existing, feature-filled product has simply been less expensive, risky, or distracting than commissioning something closer to what they need. Let’s call them peripheral customers.

Both groups have made rational decisions, benefiting from the economics of the shared software proposition. But their contributions to those economics are often quite different. The substantial costs of developing and operating a platform are spread across its customers, while serving an additional customer can be comparatively inexpensive. A subscriber who needs one capability, little support, and few new features contributes generously toward costs the provider would incur anyway.

This is the SaaS dream, but not a given. Some peripheral customers demand costly integrations or constant support. Let’s play with the possibility that a meaningful number are inexpensive to serve, and I think many are: a customer running a handful of reports files few tickets and asks for nothing on the roadmap.

The economics of leaving

AI lowers the cost of the alternative. Customers who need only a narrow slice of a platform can increasingly build or commission something closer to their requirements, or buy it from one of those newer, narrower providers.

There’s an echo of Clayton Christensen’s The Innovator’s Dilemma here: simpler alternatives finding customers whose needs are overshot by an incumbent offering. In Christensen’s account, incumbents give those customers up gladly, because the low end is the least profitable part of their business. My version runs the other way: the overshot customers are among the most profitable. I’m less interested in whether these alternatives eventually displace the incumbent than in what it loses when those customers go.¹Clay fans please don’t come after me!

Not every customer will consider leaving. Some depend on what is distinctive about the provider: its capabilities, reliability, integrations, or expertise. Call that earned dependence. Others depend on the economics of the arrangement. They could get a comparable outcome elsewhere, but the shared solution remains less expensive, or the costs and responsibilities of switching make it irrational. Call that default dependence. AI leaves earned dependence largely intact and progressively weakens default dependence.

The service hasn’t deteriorated, and neither group has suddenly become dissatisfied. But the alternatives have become more achievable. Some peripheral customers, the ones held by default, begin to leave, and the resulting loss of profit is substantially greater than the decline in revenue suggests.

Let’s run the numbers

Toy model time! Give our $100 million SaaS business 1,000 customers, 30% of them peripheral. Everyone pays $100,000 annually. Peripheral customers cost $5,000 each to serve, compared with $25,000 for the others; the company also carries $51 million in fixed operating costs. That leaves $30 million in annual operating profit.²Not a perfectly SaaSy business, although I’ve admittedly invented the economics to make my math easy.

Suppose AI-enabled alternatives persuade an additional 10% of peripheral customers to leave over a year, beyond those who would otherwise depart. What happens?

Thirty customers leave, taking $3 million in annual revenue with them while saving just $150,000 in servicing costs. Annualized operating profit falls by $2.85 million, or 9.5%, despite revenue declining only 3%.

Lose 30 core customers instead and profit falls $2.25 million, or 7.5%. Most of the damage is ordinary operating leverage, and it applies to any customer. The difference is selection. AI exposes the customers whose subscriptions make unusually large contributions to the business, while their usage and satisfaction look healthy. Change the assumptions and the outcome changes too: peripheral customers who are expensive to serve do less damage when they leave; those who cost more than they pay can actually improve profitability by departing. What matters is which customers go, and how much of the cost base stays behind.

Now stretch the model

The argument rests on two conditions: customers bought a shared capability because getting it any other way cost too much, and the provider’s costs stay put when some of them leave. Software meets both. Professional services meet the first and only partly the second, which changes where the periphery sits and what it costs to lose it.

A mid-sized law firm doing high-volume commercial work complicates the hypothesis. Clients engaging it for routine contracts or legal research might look peripheral, particularly if AI can produce much of that work cheaply. But law itself is interpretive, shaped by jurisdiction, precedent, and regulators, even on a routine contract. That’s earned dependence. The peripheral part of a law firm’s practice is smaller than the volume of standardized work suggests.³I’d expect a white shoe law firm has even less peripheral work, if only because the value of even a small edge in creative interpretation for its clients is massive.

Management consulting looks rather different. A consultancy can be indispensable as an advisor and replaceable as a source of analytical capacity. Firms sell at least three kinds of work: professional capacity embedded in client teams, statement-of-work projects with defined deliverables, and advice grounded in expertise, relationships, or judgment. Clients will become less dependent on borrowed staff as AI allows their own teams to do more. Some defined deliverables will become easier to reproduce. Specialized advisory work, especially where the client values credibility or help navigating a consequential decision, is harder to substitute. And consulting’s costs complicate the result. A firm can lose an engagement before it can shed the salaries of the consultants who delivered it. Those costs eventually adjust. The harder problem is where the margin lives. In a leveraged firm, the capacity and deliverable work staffed from the base of the pyramid often pays for everything else, and it’s the work AI currently makes easiest to replace. A consultancy can keep its most valued relationships and still lose the work that made them profitable.

Advertising agencies make the periphery harder to locate still. A single client often buys brand strategy, creative development, campaign execution, and content production from the same agency. AI makes production easier to obtain elsewhere without diminishing the client’s dependence on the agency’s creative judgment. But production often pays for the strategy, which gets bundled in or given away to win the account. When production leaves, the agency keeps the expensive part of the relationship and loses the part that funded it. Whether the lost work damages profit also depends on whether it was performed by permanent staff, freelancers, or outside partners.

Across consulting and advertising, two businesses I know well, the periphery is modular. It lives within a service, an engagement, or even a single client relationship. So the model has to count the parts of a relationship that become easier to obtain elsewhere, along with the customers who leave.

Finding the periphery

The arithmetic is straightforward. Deciding who belongs in the model is much harder. Is Salesforce’s periphery made up of customers using only a few features? A law firm’s, of clients bringing only routine matters? A consultancy’s, of engagements that mostly supply people rather than distinctive expertise?

None of these holds up. I defined the periphery by how much of a product a customer uses, and that’s the wrong measure. Someone using a narrow slice of the service can depend heavily on its reliability and integrations. A routine legal client can depend on the institution behind the advice. A consulting client buying a great deal of professional labor can depend very little on the firm supplying it.

The better measure is dependence: how much of what a customer buys relies on something distinctive about the provider and its service. That’s earned dependence. Whatever is held in place by the cost and inconvenience of going elsewhere is default dependence, and AI is changing it even when earned dependence remains intact.

As alternatives become cheaper, relationships with weaker dependence on a provider’s distinctive capabilities will be more susceptible to AI-driven substitution, even when satisfaction and usage remain healthy. Where those relationships also contribute heavily toward shared costs, profit will suffer disproportionately. That gives a business two dimensions⁴ to map its customers against: how much each contributes toward shared costs, and which kind of dependence holds it.My first two-by-two on this site! The customers to worry about sit where high contribution meets default dependence. The time to measure dependence is before customers leave, and to test whether it predicts departures better than conventional customer-health measures.

So the useful question is what a business knows about its customers today. Most companies can tell us what customers buy, how much they use it, what they pay, and how satisfied they are. How many know what those customers actually depend on them for? And how much of that dependence exists because getting the same outcome elsewhere has historically been too expensive or inconvenient?

The businesses least prepared for AI-driven erosion are the ones least able to distinguish which of their apparently healthy customers depend on them, and which simply haven’t had a good enough reason to look elsewhere.