a16z Names 2 AI Sales Strategies. There's a Third.

What happens when the vendor, not the buyer, holds the risk.

Based on the a16z article by Joe Schmidt IV and Julian Marx, “Lighthouse or Landgrab?”

Source: https://a16z.com/lighthouse-or-landgrab-how-to-pick-your-ai-sales-strategy/ 

This article by Joe Schmidt IV and Julian Marx, “Lighthouse or Landgrab?”, is foundational reading on enterprise AI go-to-market. The premise underneath it is the one worth sitting with: the buyer who signs is pricing their own risk, not evaluating your product in the abstract. They are asking what evidence makes that exposure bearable.

From that premise, two motions. Lighthouse: win a few marquee logos so a nervous buyer feels safe. Harvey needed Allen & Overy and Paul Weiss before law firms believed AI-drafted work was real. Landgrab: when the buyer already knows the problem and a mistake will not cost them their job, skip the logos and win on math, the way Decagon and Stuut went wide before an incumbent could react.

It is a genuinely useful map, and I want to take it one layer down. Underneath the two motions sits a single assumption they share. The authors frame these as two strategies. I think they are two versions of one move, and that there is a genuine third.

Lighthouse and Landgrab both leave the buyer holding the risk. Lighthouse lowers the buyer's perceived risk with borrowed credibility, the marquee logo that says this is safe. Landgrab makes the risk look worth it with self-evident ROI, the number that says this is cheap. Two different instruments, one target: talk the buyer into accepting a risk that never leaves their desk.

AI makes a third strategy possible. Move the risk onto yourself. And someone is already running it.

Lighthouse and Landgrab are two versions of one move: leave the risk on the buyer's desk and argue about how to make them accept it. The second real move is to take the risk off their desk entirely.

What follows takes their diagnosis one layer down, to the variable underneath it.

The third door has a name, and a price list

Call the third move the Underwrite. You stop persuading the nervous buyer and absorb the exposure yourself: price on the outcome, guarantee the result, eat the failure. You are no longer selling software. You are selling a priced guarantee with your own balance sheet behind it. The buyer's question shifts from does the product work to can the vendor pay if it fails?

This is not a thought experiment. Two companies are already running it, and each shows both the power and the boundary of the model. 

Sierra

Bret Taylor's customer-service agent company does not sell seats. It charges per resolved conversation, reportedly around a dollar-fifty, and nothing when the agent fails. Taylor's own framing is that the atomic unit of AI is a process, not a person, and that the market moves to outcome pricing.

Upside: The vendor, not the buyer, loses money when the product underperforms. That is the Underwrite in its purest commercial form, and it is winning enterprise logos on that basis.

Limitation: It works because a resolved support ticket is a clean, countable event. The model does not transfer to a domain where the unit of success is fuzzy. Sierra can price a resolution. It could not price a quarter of good strategy.

Intercom Fin

Fin publishes a flat 99 cents per resolution. You pay when it works, not when it runs, and a resolution is defined as a conversation the agent closes without a human.

Upside: Pricing is legible enough to print on a public page, which is only possible because the outcome is machine-loggable. The buyer carries no risk on a conversation the agent fails to close.

Limitation: The definition of the outcome does the heavy lifting, and edge cases live in that definition. A model that only works where you can write down the outcome precisely is a narrow model, by design.

Decagon, the useful contrast

In the same support category, Decagon bills per conversation, not per resolution. That is one notch back toward Landgrab. The ROI is compelling, but the buyer still carries the risk that a given conversation does not land. Same market, two different answers to a single question: how much of the customer's downside is the vendor willing to hold.

Upside: Per-conversation pricing is easier to underwrite for the vendor and easier to forecast, so it scales fast without balance-sheet exposure.

Limitation: It leaves the residual risk with the buyer, so it is not a true Underwrite. It is proof that outcome-alignment is a spectrum, not a switch.

The Underwrite is not a category. It is the far end of a spectrum, and the spectrum runs on one question: how much of the buyer's risk the vendor is willing to hold.

Why now: AI made outcomes attributable, not pricing new

Outcome-based pricing is not new. Contingency fees, revenue-cycle collections that take a percentage of what they recover, performance contracts: risk transfer is decades old. Healthcare has paid collectors a cut of collected revenue since long before software touched it. So the Underwrite is not a new idea. It is newly possible.

Outcome pricing is old. What is new is attribution.

You could never underwrite a result you could not isolate. Contingency stayed confined to the few domains where causation was legible, such as collections, where the collector either recovered the dollar or did not. Everywhere else, you could not separate the vendor's contribution from the buyer's own team, so you sold access and hoped. Agentic AI changes exactly that. When the system executes the whole workflow instead of assisting a human through it, the outcome becomes attributable to the product itself.

Attribution is not automatic: it needs three things

Agentic AI does not create attribution by default. Attribution requires three conditions to line up at once, and most deployments today miss at least one.

1.  The agent completes the whole task, not part of it. Attribution works when the agent owns the workflow end to end. Sierra resolves the ticket. Fin closes the conversation. The moment a human is in the loop finishing the job, causation smears and you can no longer say the product produced the result. Most enterprise AI today is copilot-shaped: it drafts, suggests, and assists. Harvey is the clearest case. It runs more than 25,000 custom agents and still sells seats, because a lawyer reviews and owns the final output, so the result cannot be attributed to the tool. Copilots cannot underwrite. Only full-workflow agents can.

2.  The outcome is a discrete, countable event. A resolved ticket, a recovered invoice, a qualified lead: binary, loggable, and hard to dispute. This is why support and AR moved first, because the unit of success is obvious. It breaks the instant the outcome is continuous or subjective. There is no clean definition of a resolved quarter of marketing or a successful strategy memo. If you cannot draw a hard line around the outcome, you cannot price a guarantee on it, no matter how agentic the system is.

3.  The counterfactual is clean. Even with a completed, countable task, you have to separate your contribution from everything else moving the number. Sierra can, because the ticket was either resolved by the agent or it was not. A fraud-detection agent that claims to have prevented losses runs straight into the counterfactual problem: you are claiming credit for events that did not happen, and the buyer can dispute every one. Attribution is not just measuring the outcome. It is tracing it back to your product, not the buyer's team.

Put the three together and the honest scope is narrow, and that narrowness is the point. Today the Underwrite works in a specific shape: high-volume, transactional, single-turn, machine-loggable outcomes. Support, AR and collections, chargebacks, parts of sales qualification, appointment setting. That is not most companies. It is a beachhead.

And a beachhead is exactly what makes the framework dynamic rather than static. Attribution is not a state AI has reached. It is a border AI keeps pushing outward. Every time agents get good enough to own one more complete workflow with a countable outcome, one more market crosses from un-underwritable to underwritable. The claim is not that the door is open everywhere. The claim is that the door keeps opening.

Bounded and unbounded: the line that decides everything

Attribution is one of the two conditions that make an Underwrite possible. The other is whether the downside is bounded. The distinction is simple and it decides which markets are reachable.

A downside is bounded when the worst case is a knowable dollar figure you can name in advance. If your support agent gives a wrong answer, the cost is a re-handled ticket and an annoyed customer. Bad, but finite, and you can price it. A downside is unbounded when the worst case has no ceiling. If an AI drafts a merger document and gets a figure wrong, the loss is a mispriced deal or a malpractice claim with no fixed size. You cannot name the number in advance, so you cannot price a guarantee against it.

This is why the third door stays shut in law and trading. Ask whether a vendor could underwrite a law firm's merger documents: I drafted them, and if I am wrong I will cover the claim. No one writes that guarantee, because the tail is unbounded. It is exactly why Harvey, valued at 11 billion dollars, still sells seats to Am Law 100 firms like Allen & Overy and Paul Weiss rather than outcomes. It cannot bound the downside, so it runs the Lighthouse. The authors' own flagship example proves where the boundary sits.

The rule with a door in it

Schmidt and Marx stake out one hard claim: when exposure and social proof point in opposite directions, exposure wins every time. A scared buyer cannot be reasoned out of fear with math. That rule is right, with one caveat that reopens it. Exposure only wins if you treat it as fixed, if the risk has to stay on the buyer. The frontier move is to buy the risk down with capital. When the vendor eats the exposure, the buyer has nothing left to fear. The rule does not break. It has a door in it, and bounded plus attributable is the key that opens it.

The deeper axis: who holds the risk

So what actually separates the markets where the Underwrite works from the ones where it cannot? Not the size of the buyer's downside, but the transferability of it: whether the vendor can take the customer's decision risk onto its own books at a price it can afford to carry. Insurance is one instrument for that. Guarantees, outcome pricing, deferred payment, and shared upside are others. The unifying question is not whether a risk is insurable. It is whether the vendor can credibly absorb the buyer's risk.

Transferability has exactly two conditions, and they are the two axes of the map below. The downside must be bounded, and the outcome must be attributable. Run the authors' own cases through it. Support and AR are bounded, measurable, and recoverable, so they are transferable, and those are their Landgrab examples, now migrating toward the Underwrite as attribution improves. Law and trading are unbounded, where one wrong figure misprices a position, so they are untransferable, and that is Lighthouse, exactly where they placed it. The framework does not contradict their map. It just shows the axis underneath it.

What the Underwrite requires

A guarantee on your own balance sheet needs an architecture, not a slogan. Four conditions, in order.

Attributable measurement. You can isolate that your product, not the customer's team, moved the number. Sierra can point to a resolved ticket. A consultancy usually cannot point to a resolved quarter.

Bounded downside. The worst case is a figure you can name in advance. This is the line that rules law and trading out and rules support and collections in.

A selection model. You guarantee only where you predict the outcome better than the buyer does. This is the whole game. The Underwrite is not a pricing trick, it is an edge in forecasting who will succeed.

Capital to hold the risk. Someone carries the loss until the outcome lands. This is why the Underwrite is a weapon only the well-funded can wield, and why it is a strategy at all rather than a billing choice.

Two traps, because the map deserves them

The insurer with no actuarial table. Guarantee outcomes indiscriminately and you adverse-select the buyers most likely to fail, because the ones who cannot get the result alone come running to the vendor who eats the downside. The disciplined Underwrite underwrites selection, not product.

And selection is only half of it. The other half is definition: the guarantee has to state precisely what success means and which steps it covers, because the edge cases and the muddy, high-complexity work are where a vague promise turns into an open-ended liability. An outcome you have not defined is an outcome you cannot price.

The cautionary history is already written in cyber, where vendors announced breach warranties, found the claims math brutal, and quietly withdrew them. CrowdStrike famously paid out on none for years. A guarantee you cannot price is a marketing splash, not a moat. 

The multiple you just torched. Absorbing customer risk turns a capital-light software company into something that looks like a specialty insurer, and the market pays very different multiples for the two. The reframe is the moat: on top of a strong product, the balance sheet becomes a go-to-market weapon. Product quality is what makes the outcome attributable in the first place. Capital is what lets you stand behind it. You need both, and only firms already sitting on venture capital can do the second. In a category where everyone rents the same models, who can afford to hold the risk may be the last durable advantage left.

The close

Schmidt and Marx say a marquee logo is proof that travels, and that in fragmented markets, where the controller in Des Moines will never hear of your other customers, proof does not travel, so you grind out every deal on the math.

A guarantee travels where a logo can't.

It needs no peer network, no status hierarchy, and no one watching who signed first. It works in exactly the fragmented markets the authors say proof cannot reach. Which raises the question I cannot put down. Once a vendor can credibly underwrite the outcome, does “proof travels” stop being an axis at all? A priced guarantee may be the one form of proof that travels to any buyer, in any market, with no one else's permission.