Andreessen Horowitz partner David Haber says the business model around AI is starting to move from charging for labor saved toward charging for results delivered. In his essay, titled Winning the Relay Race: Long-Running Agents in High-Value Customer Journeys, Haber argues that the largest opportunity may sit in customer journeys that run for weeks or months, not in one-off automations or short support interactions.
The article, translated into Chinese by TechFlow, frames the shift as a move away from pricing AI like software seats or hourly efficiency tools and toward a model where vendors can claim part of the revenue lift they help create.
From bounded tasks to long customer journeys
Haber opens with a familiar telecom support example. A customer reporting an internet outage may first hit an automated voice system, get routed through a menu, sit through hold music, then repeat account details and the problem description to multiple support agents before finally reaching a technician booking.
AI has already changed parts of that experience, he writes. An agent can answer immediately, look up the account, check whether other users in the same area are affected, solve simple issues, or schedule a repair. But until recently, that was close to the practical limit: short interactions, tightly scoped tasks, and clearly defined workflows.
What is changing now, according to Haber, is the emergence of AI agents that can track leads and execute work over much longer periods. That opens the door to redesigning entire customer journeys rather than improving a single touchpoint. He points to insurance customers moving from first contact to final payout, mortgage borrowers moving from pre-approval to closing, and enterprise prospects moving from an initial demo to a signed contract.
Why value shows up only at the finish line
The essay describes these workflows as relay races. The value is not created at one isolated step but at the finish, after many handoffs, follow-ups, reviews, and decisions.
Haber uses plaintiff law firms as the main example. A potential case usually starts with an inbound call. Intake staff gather key facts, then hand the matter to a lawyer for a decision on whether to take the client. Once the case is signed, investigation begins: staff pull police reports, try to contact witnesses, send preservation letters to defendants, and keep the client updated. After treatment is complete, the firm builds a medical timeline and hands the file to a lawyer to prepare the demand letter.
Each handoff creates a chance for value to leak out. Prospective clients may call several personal injury firms and sign with the one that responds first. If they hit voicemail or wait too long for case evaluation, the lead can disappear. After signing, evidence collection often requires several parallel threads, and a firm may not have the patience or staffing to keep calling after four unanswered attempts. Evidence review can also be uneven, especially in cases that look routine or small at first glance. Haber writes that a careful reading of a doctor’s notes might reveal signs of brain injury and materially increase the settlement value, but a firm may instead accept the first offer. Over time, information gets scattered across people, documents, and systems. Bills may never make it into the demand package, and gaps in treatment may go unexplained, pulling down the final outcome.
What long-running agents are supposed to do
Haber contrasts that with the current generation of AI agents, which have mostly handled bounded tasks such as sending a follow-up email, summarizing a call, drafting a document, or managing a narrowly defined interaction. Customer support worked well as an early use case because the problem space was clear, reliability could be kept high, and edge cases could still be escalated to a human team.
Long-running agents, he argues, can pursue a goal across dozens of steps, multiple modalities, and longer spans of time. In the law firm example, an agent could be assigned at intake with a clear goal: maximize the client’s recovery. It would store and retrieve case context as needed, wake up automatically around events or deadlines, send sub-agents to contact witnesses, update clients, and communicate with lawyers by phone, email, or text. It could synthesize findings, plan the next action, draft documents for review, and keep the case moving toward the best possible outcome.
At a portfolio level, Haber says, these agents can also learn across cases. Data generated in one matter can improve how the next one is handled.
Fixing the main sources of value loss
The essay says long-running agents can address the main places where relay-race workflows lose value today.
- At the start of the journey, prospects are less likely to drop because an agent is always available and can evaluate a case almost instantly.
- During evidence gathering, the workflow is no longer capped by headcount because the system can launch another sub-agent to make a fifth, tenth, or twentieth call.
- Routine-looking cases can still receive full evidence review, which may surface clues that would otherwise be missed, including signs of brain injury in medical records.
- The system keeps a complete record of the matter, from witness call recordings to police reports and medical files, reducing the odds that important information disappears over a multi-month process.
For a16z, the central point is simple: the highest-value opportunity appears when the agent can run the whole race, not just one leg of it.
From cost savings to outcome pricing
Haber ties that technical capability to a commercial shift. If an agent can deliver faster and better results across the full journey, the value it creates may show up as revenue growth, not just lower operating cost. In that setting, a vendor may be able to price by work product or outcomes and share in the incremental value the system creates.
In plaintiff law, a single case is a natural unit of work, and the end result — the settlement amount — is easy to identify. As agents take on a larger share of the final work product, Haber says, the connection between the tool and the outcome becomes easier to see. That gives vendors a path to contract values that pure cost-cutting software often cannot reach.
Friction remains
The essay does not present the shift as frictionless. Haber notes that in most U.S. states, non-lawyers cannot share legal fees, so pricing in plaintiff law would need to align with settlement value without directly taking a cut of the legal fee. In other verticals, procurement teams may resist because outcome-based contracts are not yet standard. Attribution disputes may also appear in the early stages even if the agent is doing more of the work tied to the result.
That is why Haber describes long-running agents as both a technical challenge and a business challenge. Building the system is only part of the problem. Vendors also have to structure contracts, align incentives, and defend attribution.
Why Red Ventures matters in the argument
To show that outcome-linked pricing is not a new idea, Haber points to Red Ventures, or RV, as a historical example. Founded by Ric Elias and Dan Feldstein, RV built end-to-end outsourced digital sales operations around long-duration consumer contracts.
In insurance, for example, RV collected consumer information when people searched for products through sites it owned, including homeinsurance.com. It then contacted those users, walked them through coverage options, and ultimately closed policies. Because RV controlled the journey from lead generation to sale, it could get paid by insurers for each policy sold.
According to S&P Global data cited in the essay, RV generated about $1.4 billion in revenue and about $370 million in EBITDA in 2022. Haber presents that as evidence that controlling a high-value customer journey from end to end can support a powerful outcome-based model.
But he also says the structure is under pressure. At the top of the funnel, product discovery is changing. Consumers may no longer rely mainly on Google search or comparison sites and may instead ask personal agents such as Meta’s Muse or Instinct for recommendations. That reduces the value of owning discovery websites. Farther down the funnel, long-running agents may outperform large teams managing a linear process. Haber says those agents can work across products and communication methods, support any language, scale broadly, and keep going with effectively unlimited persistence.
The essay says RV’s EBITDA is expected to fall to about $160 million in 2025 because of customer churn and softer ad spending. Haber treats that as an early sign of how exposed the old model may be as AI changes the market.
Three paths to market
Haber outlines three go-to-market approaches for long-running agents.
Sell the platform
Under this model, the vendor sells long-running agents as software and plugs them into an enterprise’s existing operation. The plaintiff law example, where an agent works alongside the current team and workflow, fits this category.
The upside is pricing room. There may be no close software or outsourcing contract to anchor the price. The downside is a longer sales cycle, since buyers may need a new budget line, and attribution becomes harder if the agent touches several departments.
Sell the service
Here, the vendor takes over a service that has already been outsourced and delivers the desired result more efficiently with long-running agents, much as Red Ventures controlled the insurance acquisition journey.
This can shorten the sales cycle because the buyer’s ROI math is simpler and BPO budgets already exist. The catch is that pre-AI pricing models may not reflect the real value these agents can create. Haber also notes that many high-value journeys are still handled in-house, which limits the reach of this path.
Go full stack
The third route is vertical integration. Instead of selling a long-running agent to a debt collection company, for example, a business could buy debt portfolios itself and run collections directly. That would allow it to capture all of the upside from AI and ease attribution problems.
It also adds operational weight. A full-stack operator has to take on licensing, compliance, and other service-layer responsibilities that sit outside the customer journey itself.
a16z’s bottom line
Haber says it is still too early to know which model will win. Even so, he argues that long-running agents are likely to outperform existing approaches in large relay-race workflows and expand AI’s addressable market from a share of labor budgets to a share of revenue.
He closes by saying plaintiff case preparation is only one example. Similar opportunities, in his view, exist in mortgage origination, insurance claims, complex enterprise sales, and many other high-value journeys.