Andreessen Horowitz (a16z) partner David Haber argues that the commercial model for AI is shifting away from selling “seats” and counting labor savings, and toward charging for outcomes and sharing in customers’ revenue growth.
In the analysis translated by TechFlow and republished by PANews, Haber says AI agents are no longer limited to short, clearly bounded interactions. New capabilities are allowing them to track leads and execute tasks over much longer periods, which opens the door to managing complex customer journeys that stretch across weeks or even months.
From short support interactions to long-running execution
Haber begins with a familiar example: calling a telecom carrier to report an internet outage. The traditional process often involves automated menus, long waits, multiple handoffs, and repeated explanations.
AI has already started to change that. Instead of moving through a robotic phone tree and talking to several representatives, a customer can be greeted by an AI agent that answers immediately, checks the account, determines whether others in the area are affected, resolves simple issues, or schedules a technician.
Until recently, Haber writes, that was close to the practical limit of the technology. The newer shift is that AI agents can now track work over longer time horizons and handle goals that unfold across many steps. In his view, that changes the role of AI from improving isolated interactions to redesigning entire customer journeys.
He points to insurance, mortgage, and enterprise sales as examples. AI could guide an insurance customer from the first call through payout, move a mortgage borrower from pre-approval to closing, or push an enterprise lead from an initial demo to a signed contract. That, he says, creates room for AI vendors to be paid on revenue growth, not only on cost reduction.
“Value appears only at the end of the relay race”
Haber describes these businesses as relay races: long, multi-step processes in which value is created only if the entire sequence is completed. Every handoff introduces the chance of customer loss or value leakage.
He uses the workflow of a traditional plaintiff law firm as a detailed example. A potential case often starts as a phone lead. Intake staff gather key facts, then pass the matter to an attorney to decide whether to sign the client. If the firm takes the case, the investigation begins. Paralegals collect police reports, try to reach witnesses, send preservation letters to defendants, and keep the client updated. Once treatment is complete, the firm prepares a medical timeline and the attorney finishes the demand package.
This process breaks down in several places. Prospective clients often contact multiple personal injury firms and sign with the first one that responds. If they get voicemail or if screening takes too long, the lead can disappear. After a case is signed, evidence gathering requires many parallel threads, and a firm may not make a fifth call after four attempts go unanswered. Limited staff can also mean evidence is reviewed only superficially, especially in cases that look routine or small. Haber notes that if a doctor’s records contain signs of brain injury and no one reads closely enough to catch them, the settlement value can be materially reduced. Cases can also drag on for months, with information scattered across documents, systems, and staff members, making it hard to maintain a full picture of what is happening.
What long-running agents are supposed to do
Most AI agents so far have worked on bounded assignments: send a follow-up email, summarize a call, draft a document, or handle a narrow interaction in customer support. Customer service was an early fit because the problem space was well defined, the output could be checked, and unresolved edge cases could still be escalated to a human team.
Haber says long-running agents change that equation. They can operate across dozens of steps, multiple modalities, and longer time spans in pursuit of a goal.
In the plaintiff law firm example, an agent could be assigned a case at the moment the intake call begins, with the goal of maximizing the client’s recovery. It would store and retrieve case context as needed, wake itself up around deadlines or key events, dispatch sub-agents to contact witnesses, update the client, and communicate with lawyers through phone, email, or text.
It could then combine what it learns, decide on the next action, draft documents for review, and keep pushing the matter toward the best possible result. Across multiple cases, the system would continue learning, using data from one matter to improve the handling of the next.
Haber argues that this directly targets the main points of value leakage in today’s relay races. At the beginning of the process, clients are less likely to be lost because the agent is available at any hour and can evaluate cases almost immediately. During evidence gathering, the system is not constrained by staffing and can spin up another sub-agent to make the fifth, tenth, or twentieth call. Routine-looking matters can still receive full attention, and every piece of evidence can be reviewed. The agent can also preserve a complete record of the case, from witness calls and police reports to medical records, reducing the odds that critical information gets dropped.
Why the pricing model changes
For Haber, plugging operational gaps is only part of the story. The bigger shift is that when an AI agent owns more of the customer journey, the vendor’s pricing model can change as well.
If an agent delivers better outcomes faster, the benefit is no longer limited to lower labor costs. It can produce observable revenue gains. Because that impact is easier to see and easier to attribute to the agent, the vendor can price on work product or on outcomes and capture part of the incremental value created.
In a plaintiff law firm, a single case is the natural unit of work, and the end result — the settlement amount — is a clear economic outcome. As the agent takes on a larger share of the work that produces that result, the link between the agent and the final value becomes easier to establish. In that setup, Haber says, the vendor can share in the revenue uplift and reach contract values that pure cost-savings software could not support.
He also notes the friction. In the plaintiff bar, most U.S. states prohibit non-lawyers from sharing legal fees, so pricing would have to align with settlement outcomes without becoming a direct fee split. In other verticals, procurement teams may resist because they are not used to signing outcome-based contracts. Attribution disputes may still surface in the early stages even if agents take on more of the process. That means deploying long-running agents is as much a commercial challenge as a technical one.
Red Ventures as a historical reference point
Haber points to Red Ventures as a useful precedent. Founded by Ric Elias and Dan Feldstein, the company built end-to-end outsourced digital sales operations for long-duration consumer contracts.
In insurance, for example, Red Ventures would collect customer information when people searched for products through sites it owned, including homeinsurance.com, contact those customers, guide them through coverage options, and close the policy. Because it controlled the full journey, it could be paid by insurers for each completed sale.
That model was highly effective. Citing S&P Global data, Haber says Red Ventures generated about $1.4 billion in revenue and roughly $370 million in EBITDA in 2022.
He argues that the model now faces structural pressure. At the top of the funnel, consumers are changing how they discover products. Instead of relying only on Google search or comparison sites, they can ask personal agents such as Meta’s Muse or Instinct for recommendations, which reduces the value of owned discovery sites. Deeper in the funnel, long-running agents can handle customer journeys that are far more complex than the linear processes traditionally run by large human teams. They can work across products, communication channels, and languages at scale, with no practical limit on persistence. Haber says Red Ventures’ EBITDA is projected to fall to about $160 million in 2025 because of customer losses and softer ad spending, an early sign of how vulnerable the model may be as AI reshapes the sector.
Three go-to-market paths
Haber outlines three ways long-running agents could be brought to market.
Sell the platform
Under this model, the vendor sells the agent as software and supports enterprises using it inside their existing operations. The plaintiff law firm example, where the agent is integrated with current teams and workflows, fits here.
The upside is pricing power, because there may be no obvious software or outsourcing benchmark to anchor the contract. The downside is a longer sales cycle. Buyers may need a new budget line, and if the agent interacts across several departments, attributing outcomes can become harder.
Sell the service
In this approach, the vendor takes over an already outsourced function and delivers the desired result more effectively. Haber points to Red Ventures’ ownership of the insurance acquisition journey as an example.
This can shorten the sales cycle because the buyer’s ROI math is simpler and a BPO budget line already exists. But older pricing models, built before AI, may fail to capture the true value of long-running agents. Haber also notes that many high-value customer journeys are still run in-house, which limits the reach of this model.
Go full stack
The third path is vertical integration: become an AI-native operator rather than a vendor. Haber gives the example of buying debt portfolios and collecting on them directly instead of selling long-running agents to a debt collection company.
That approach lets the operator capture all of the upside from AI and avoids attribution problems. The tradeoff is that the company must take on licensing, compliance, and other service-layer obligations that sit outside the customer journey itself and add operational complexity.
From labor budget to revenue budget
Haber says it is still too early to know which path will win. His conviction, though, is that long-running agents will succeed in these large-scale relay races and expand AI’s addressable market from a portion of labor budgets to a portion of revenue.
He closes by saying plaintiff case preparation is only one example. Similar outcome-driven models could emerge in mortgage origination, insurance claims handling, complex enterprise sales, and many other high-value journeys.