The Incumbent's Dilemma: Cannibalize Your Own Seats, or Watch Startups Do It For You

Table of Contents

• What the Incumbent's Dilemma Actually Means in the Agent Era
• Why Seats Are the Wrong Thing to Defend
• The Cannibalization Math No CFO Wants to Run
• The Startup's Asymmetric Advantage
• What Incumbents Actually Have That Startups Don't
• Four Strategic Postures, Ranked
• How to Cannibalize on Purpose
• Insights Most People Overlook
• Frequently Asked Questions
• Conclusion
• References

What the Incumbent's Dilemma Actually Means in the Agent Era

Clayton Christensen described the original innovator's dilemma decades ago: well-run companies fail precisely because they listen to their best customers and protect their most profitable lines, leaving the low-margin, "good enough" entrant room to climb the value chain. The agent era didn't invent this problem. It just compressed the timeline from a decade to a couple of fiscal years and aimed it directly at the seat-based subscription, the single most lucrative business model software has ever produced.

Here is the specific version of the trap. An incumbent sells a SaaS product priced per user per month. A support tool at $80 a seat, a sales engagement platform at $150 a seat, a legal research suite at $300 a seat. The value of that pricing rests on a quiet assumption: a human sits in the chair and does the work. Agentic AI breaks the assumption. When one autonomous agent can resolve the tickets that previously needed eight support reps, the customer no longer needs eight seats. They need a result, and they would very much like to pay for the result instead of the eight logins they no longer use.

The dilemma is that the incumbent is the best-positioned company in the world to build that agent and the worst-positioned company to want to. This connects directly to the broader debate over whether SaaS pricing breaks when one agent replaces a ten-seat team, and it is the load-bearing tension under the entire GaaS thesis.

Why Seats Are the Wrong Thing to Defend

Seats were never a measure of value. They were a convenient proxy. Counting logins is easy, predictable, and it scales with customer headcount, which historically grew. Investors loved it because net revenue retention above 120 percent was practically automatic: existing customers added people, added people meant added seats, and the curve went up and to the right without anyone having to sell anything new.

Agents sever the link between value delivered and humans logged in. That is the whole disruption in one sentence. Once value is decoupled from headcount, defending seat count means defending a number that no longer tracks the thing customers actually buy. You can hold the line on price per seat all you want; if the customer needs a third as many seats, you have defended the unit and lost the account.

The categories most exposed share a profile worth naming plainly: high-volume, repetitive, well-structured knowledge work where the output is verifiable. Tier-one support, sales development, data entry, claims processing, routine document review, bookkeeping. If you sell seats into those workflows, the agent is already underwriting your competitor's pitch deck. The categories that hold up longest involve judgment that's hard to verify, relationships, regulatory accountability, or genuinely creative work, which is why the question of which seat-based models are most exposed to agents is really a question about how verifiable your customer's output is.

The Cannibalization Math No CFO Wants to Run

Let's make it concrete, because the abstraction hides how uncomfortable the decision is.

Suppose you run a customer-support SaaS doing $100M in ARR across roughly 12,500 seats at $80 a month. Your gross margin is 80 percent. Beautiful business. Now an agent, yours or a competitor's, can deflect 60 percent of ticket volume autonomously. A typical customer who bought 10 seats genuinely needs 4 going forward.

If you do nothing and the market reprices around outcomes, you don't lose 60 percent of revenue gracefully over five years. You lose customers at renewal to a startup charging, say, $0.50 per resolved ticket, and that startup's bill to the customer comes in 50 percent below the old seat total while still carrying a fat margin, because the agent's marginal cost is inference, not payroll. Your $100M doesn't erode. It walks out the door at contract boundaries.

If instead you launch your own outcome-priced agent, you cannibalize. The same customer moves from $800/month in seats to maybe $450/month in resolved-ticket fees. Revenue per account drops. That is the cannibalization, and it is real, and your board will see it in the next quarter. The difference is that you keep the customer, the data relationship, and the renewal, and you can expand the agent into adjacent workflows you never sold seats into at all. This is precisely the seat-count shrinkage metric SaaS investors have started to fear, and the only thing worse than reporting it yourself is having a competitor cause it for you.

The honest framing for the CFO is not "will revenue per customer fall." It will. The framing is "do we want to capture the smaller, faster-growing outcome-based pool, or surrender the whole pool to keep a vanity seat count for two more quarters." Andreessen Horowitz has written at length about how agent pricing is shifting software from seats toward work delivered, and the through-line is consistent: the pool of money is moving from software budgets to labor budgets, and labor budgets are an order of magnitude larger.

The Startup's Asymmetric Advantage

The startup attacking you has one enormous structural gift: nothing to protect. It has no installed base of seat revenue to cannibalize, no board conditioned on net revenue retention, no sales comp plan built around expanding seat counts, no nine-figure number that goes down if it prices honestly. It can walk into a prospect and say "pay us per resolved outcome, $0 if it doesn't work," and there's no internal civil war about it.

That asymmetry is the entire reason the dilemma is a dilemma. The startup isn't smarter or better-funded. It is simply unconflicted. It can lead with the pricing model that customers prefer because it has no legacy revenue that model would undercut. Christensen's original insight, restated cleanly in the Harvard Business Review's treatment of disruptive innovation theory, was never that incumbents lacked the technology. It's that their existing customers and existing economics made the right move look irrational right up until it was too late.

There's a second-order advantage worth flagging: startups can run at zero or negative margin on the outcome to win logos, funded by venture capital, in a way a public incumbent answerable to quarterly earnings simply cannot. The incumbent has to make the agent pay. The startup has to make the agent spread. Different games.

What Incumbents Actually Have That Startups Don't

This is the part the "SaaS is dead" crowd consistently underweights. The dilemma is real, but incumbents are not defenseless, and the smart ones are leaning hard on three durable assets.

Proprietary data and workflow context. An agent is only as good as what it can see and act on. The incumbent already holds years of the customer's structured data, the labeled outcomes, the edge cases, the integrations. A startup agent has to either rebuild that context or beg for API access, which is exactly why the data moat agents can't easily cross is the incumbent's strongest card, and why some vendors are quietly tightening integration terms.

Distribution and trust. Selling into enterprises is slow, relationship-heavy, and procurement-gated. An incumbent with a signed MSA, a security review already passed, and a champion inside the account can ship an agent to thousands of customers in a release note. The startup has to win each of those battles cold. As the broader cluster argues, distribution often beats raw capability when the capability gap is closing anyway.

The system of record. Whoever owns where the data lives has leverage over whoever wants to act on it. The unfolding system-of-record versus system-of-action battle is the incumbent's home turf, agents need a place to read and write state, and that place is usually the incumbent's database.

None of this makes incumbents safe. It makes them capable of fighting back, if they're willing to use the assets to undercut their own pricing rather than to lock the doors and stall.

Four Strategic Postures, Ranked

Incumbents are visibly sorting into four responses. They are not equally good.

1. Cannibalize on purpose (best)

Build the outcome-priced agent yourself, route your own customers to it, and accept lower revenue per account in exchange for keeping the account, the data, and the expansion surface. Salesforce's move to per-conversation Agentforce pricing and the wider repricing for the agent era is the clearest large-cap example of an incumbent choosing to eat its own seats before someone else does.

2. Bundle the agent into the suite (good, defensible)

Give the agent away or near-away as part of the existing subscription so the customer never has a reason to evaluate a startup. This protects the relationship and uses scale as a weapon, the bundled-agent strategy of the mega-vendors. The risk: you train customers to expect agents for free and erode pricing power across the suite.

3. Bolt an agent on as a premium upsell (risky)

Keep seat pricing intact, sell the agent as an add-on. This is the path of least board resistance and the most common, and it's frequently lipstick rather than transformation. It defends the quarter and concedes the model. Customers still get the outcome-pricing pitch from startups; you've just made yours more expensive.

4. Deny and defend (worst)

Insist the workflow needs a human, slow-roll the agent, and tighten API access to keep startups out. This buys time and forfeits the future. The data-access gatekeeping can work for a while, but it signals to the market exactly which revenue you're afraid of.

How to Cannibalize on Purpose

If posture one is correct, and for exposed categories it usually is, the execution problem is organizational, not technical.

Run the agent business as a separate P&L with its own targets, the way Christensen prescribed: a unit that's allowed to undercut the mothership without having its budget held hostage to the seat-revenue line. Comp the sales team on outcomes booked, not seats retained, or you've asked them to sell the thing that shrinks their commission. Price the agent against the customer's labor budget, not your old software price, McKinsey's research on the economic potential of generative AI consistently sizes the prize in displaced labor cost, which dwarfs the software line you're protecting.

And move before the renewal cycle does the deciding for you. The window where you can reprice your own base on your own terms is open right up until a competitor reprices it for you. After that, you're negotiating from the back foot, matching someone else's outcome price while explaining to your board why retention slipped. The five-year SaaS-to-GaaS transition rewards the incumbents who start the painful repricing while they still have the leverage to set the terms.

Insights Most People Overlook

The cannibalization is not 1-to-1, and that's the whole opportunity. Everyone models the agent eating existing seat revenue. Far fewer model that the agent reaches workflows you were never invited to sell seats into. When you stop charging for logins and start charging for resolved work, your addressable surface inside the account expands into adjacent processes the customer would never have bought a separate seat-based tool for. The revenue per account drops; the revenue per account over five years can rise. Most CFO models miss this entirely because they only run the subtraction.

"Protecting margin" can be the most expensive decision a SaaS company ever makes. The instinct to keep the agent as a high-margin upsell feels prudent and is often fatal. It optimizes the income statement for the exact quarters in which the market is deciding the new pricing norm. By the time the margin-protectors notice the norm has set without them, the outcome price is anchored at a competitor's number and they're a follower in a category they used to own.

The startup's lack of a data moat is temporary; the incumbent's lack of nerve is not. Everyone frets about the startup's missing context. But context can be acquired, through usage, through partnerships, through the customer simply handing it over because the outcome pricing is too good to refuse. The thing that's genuinely hard to acquire is an incumbent board's willingness to vote for lower revenue per customer. The scarce asset in this fight is courage, not data.

Outcome pricing quietly transfers risk, and that's why customers love it. Seat pricing puts all the utilization risk on the buyer, you pay whether or not anyone uses the login. Outcome pricing shifts it to the vendor. That shift is a feature customers will pay a premium for and abandon you to get. Incumbents keep analyzing this as a pricing-mechanics question when it's really a risk-allocation question, and risk allocation is stickier than features.

Internal sales compensation will kill more agent strategies than competitors will. The most underrated obstacle isn't the startup or the technology. It's the account executive whose commission is built on seat expansion being asked to sell the product that shrinks seats. Until comp is rebuilt, the field organization will quietly bury the agent. The disruption that takes incumbents down usually walks in through their own incentive plan.

Frequently Asked Questions

Does cannibalizing seats always mean lower total revenue? In the short term for a given account, usually yes, revenue per customer drops as seats convert to outcomes. Over a multi-year horizon it can reverse, because outcome pricing lets you expand into adjacent workflows and price against the customer's much larger labor budget. The mistake is modeling only the subtraction and ignoring the expansion surface that opens up.

Why can't incumbents just match the startup's outcome pricing later, once it's proven? They can, but from a weaker position. By the time outcome pricing is "proven," a competitor has anchored the price in the customer's mind, won the early logos, and started accumulating the workflow data that makes the agent better. Matching late means competing on someone else's terms with a worse product and a skeptical board.

Is the data moat a real defense or just incumbent cope? It's real but not permanent. Proprietary data and workflow context genuinely make an incumbent's agent better and raise the startup's cost of entry. But customers can hand data to a challenger if the outcome pricing is compelling enough, and gatekeeping API access to block startups tends to signal exactly which revenue you're protecting. Treat the moat as runway, not a wall.

Which SaaS categories are relatively safe from this dilemma? Workflows where output is hard to verify, where regulatory accountability requires a named human, where relationships drive the value, or where the work is genuinely creative and non-repetitive. The more verifiable and repetitive the work, the faster the agent erodes the seat model.

Should a startup founder attack incumbents on price or on workflow ownership? On the model, not just the number. Leading with "pay per outcome, $0 if it fails" is powerful precisely because the incumbent can't easily match it without cannibalizing. But durable wins come from owning the workflow end to end so you accumulate the data and become the system of action, not just the cheaper line item.

How does this change what enterprise buyers should ask vendors? Buyers should ask how a vendor prices when seat counts fall, whether the agent is a genuine outcome-priced product or a seat-preserving upsell, and who owns the workflow data the agent generates. The answers reveal whether the vendor has actually chosen to cannibalize or is quietly defending a number.

Conclusion

The incumbent's dilemma in the agent era reduces to a single uncomfortable sentence: you must undercut your most profitable product before a startup does it for you. Seat-based pricing was a brilliant proxy for value right up until agents decoupled value from headcount, and now defending the seat count means defending a number that no longer tracks what customers buy. The cannibalization math is real and the revenue per account will fall, but the alternative isn't keeping that revenue, it's handing the whole pool to an unconflicted challenger pricing by outcome.

The incumbents who win won't be the ones with the best agents. They'll be the ones willing to point their data moat, their distribution, and their system-of-record advantage at their own pricing, running the agent as a separate P&L, comping the field on outcomes, and pricing against labor budgets while the window to set the terms is still open. This is one node in a larger argument running through the GaaS cluster: that the shift from software budgets to labor budgets is rewriting who captures the value, and the companies that hesitate to cannibalize are simply choosing a slower way to be cannibalized.

References

• The Pricing of AI Agents, Andreessen Horowitz (a16z)
• What Is Disruptive Innovation?, Harvard Business Review
• The Economic Potential of Generative AI: The Next Productivity Frontier, McKinsey & Company