Author: David Haber, a16z partner
Compiled by: Shenchao TechFlow
Shenchao Introduction: This a16z analysis points directly at the most critical turning point in AI commercialization: from saving costs by the hour to sharing revenue by outcomes. If you are a SaaS practitioner or investor, this article explains why the next batch of AI unicorns will no longer sell "seat fees," but instead take a share of the revenue growth they generate for customers.
Traditionally, calling your internet provider to report an outage went something like this: a robotic voice first calls you a "valued customer," then makes you fumble through a phone menu for several minutes: "Press 3 for technical support, press 2 for broadband, then press 4 for other issues." Then you are placed on hold with waiting music, feeling forgotten, until a customer service representative finally picks up and asks for your name, account number, and description of the problem. If restarting the router does not fix it, you have to keep waiting, then explain the whole thing from the beginning to another representative, and only then can you schedule an on-site repair visit.
AI is already changing this kind of customer service scenario. Instead of dealing with a robotic voice and several customer service representatives, an AI agent can now pick up the phone immediately. It can look at your account, check whether other users in your area are also affected, and either directly resolve simple issues or help you schedule a repair technician. There is no more "all customer service representatives are currently busy," and no more waiting music. Until recently, this was basically the upper limit of the technology.
But now, new capabilities are enabling AI agents to track threads and execute tasks over longer time horizons. We believe this will allow AI companies to move beyond single customer interactions and rethink how to handle complex customer journeys that last weeks or even months. AI will guide insurance customers from the first call to final payout, guide mortgage borrowers from pre-approval to closing, and guide enterprise sales leads from the first demo to signing. This will not only deliver better outcomes for customers, but also allow AI vendors to get paid based on revenue growth, rather than charging only on the basis of cost savings.
Value only appears at the finish line of the relay race
There are many types of businesses where creating value for customers requires many steps, spanning weeks or even months. Take the traditional plaintiff law firm case preparation process as an example. Potential cases usually come in as phone leads. An intake specialist collects key information and then hands it to a lawyer to assess whether to take the client. Once the firm signs the case, investigation and evidence gathering begin: paralegals obtain police reports, try to contact witnesses, send preservation letters to defendants to preserve evidence, and continuously update the client on progress. Then the firm waits for the client to finish treatment, drafts a medical timeline, and gives the materials to the lawyer to complete the demand letter.
This process is what we used to call a "relay race": a long journey where every handoff carries the risk of leakage or value loss. Potential clients usually call several personal injury law firms and often sign with the one that responds fastest. So if the client only reaches voicemail, or if qualification is too slow, the case falls out of the funnel. After the client signs, evidence gathering involves multiple parallel threads, and the firm may not have the patience to make a fifth call after four calls go unanswered. In addition, staffing limits may prevent the firm from reviewing every piece of evidence in detail, especially cases that appear routine and low-value. If carefully reading the doctor's records could have revealed signs of brain injury, greatly increasing the settlement amount, the firm may simply accept the other side's first offer. Finally, a case may drag on for months, involve many participants, and have information scattered across documents, people, and systems. This makes it nearly impossible for the firm to maintain a global view of every client. As a result, important bills may not be included in the demand letter, gaps in treatment may go unexplained, and the final settlement amount is pushed down.
The opportunity for long-horizon agents
Until recently, AI agents mostly handled bounded tasks: "send this follow-up email," "summarize that call," "create this document," and very narrow interactions. For example, customer service became an excellent early AI deployment scenario because the problem space is clearly defined, high reliability can be achieved, and it can still be customized to a company's specific policies. Interactions are short, outcomes are verifiable, and if the AI agent encounters an edge case it cannot handle, the existing customer service team is usually still there to escalate it.
Now, long-horizon agents can pursue goals across dozens of steps, multiple modalities, and longer time spans, enabling them to run the entire relay race from start to finish. In the plaintiff law firm example, starting from the initial intake call, a long-horizon agent can be assigned the case with the goal of maximizing the client's recovery. The agent saves and retrieves case context as needed. It wakes up automatically based on events or deadlines, proactively dispatches subagents to contact witnesses, update the client, and communicate with lawyers via phone, email, or text. It then integrates findings, plans the next steps, drafts documents for review, and drives the case toward the best possible outcome. Across clients, it continuously learns, and the data generated by each case helps it handle the next case better than it does now.
Agents solve the main value leakage points observed in today's relay race. At the start of the journey, clients no longer drop off because case evaluation is slow, because the agent is online at any hour and on any day of the week, and can complete case evaluation almost instantly. Evidence gathering is no longer constrained by staffing, because the case agent can spin up an additional subagent at any time to make the fifth, tenth, or even twentieth call. Because agents can scale almost infinitely, even cases that appear routine are handled thoroughly, and every piece of evidence is carefully reviewed, exposing brain injury clues that might otherwise have been missed. Finally, the agent keeps a complete record of the case, from recorded witness calls to police reports to medical records, ensuring that no information is lost.
Charging by outcomes
Beyond plugging these leaks, owning the complete customer journey can also change the agent vendor's business model. When an agent can deliver outcomes with higher quality and greater speed, it can generate revenue growth for customers, not just cost reduction. More importantly, because the impact is observable and attributable to the agent, vendors can share in the incremental value they create by pricing per work product or outcome. In the plaintiff law firm example, a single case is a natural unit of work, and the value of the final outcome, the settlement amount, is clearly identifiable. In addition, because the agent accounts for a growing share of the final work product, the outcome is easier to attribute to the agent. Vendors can therefore share in revenue growth, reaching contract values that were previously unattainable through cost reduction alone.
Admittedly, shifting to outcome-based pricing is not frictionless. In the plaintiff lawyer industry, most states prohibit non-lawyers from sharing legal fees, so pricing needs to align with settlement amounts without directly taking a cut of them. In other verticals, vendors may encounter resistance from procurement teams because they have no precedent for signing outcome-based contracts. And even if agents come closer to fully owning outcomes, attribution disputes may still arise initially. Therefore, successfully deploying long-running agents is as much a business challenge as a technical one.
Still, charging by outcomes in high-value customer journeys is not a new idea.
The case for an AI-native Red Ventures
Red Ventures (RV) is a good historical example. Founded by Ric Elias and Dan Feldstein, RV pioneered digital end-to-end outsourced sales operations for long-cycle consumer contracts. In insurance, for example, RV would collect customer information when customers searched for these products (through sites such as homeinsurance.com), proactively contact customers, walk them through coverage options, and ultimately close the policy. Because it fully owned this journey, RV received a payment from the insurer for every policy sold. This model was once extremely effective: according to S&P Global data, RV generated about $370 million in EBITDA on about $1.4 billion in sales in 2022.
Today, this model appears to face a structural threat. At the top of the funnel, the way consumers discover products is changing. Instead of relying on Google search or comparison sites to explore and compare products, they can directly ask personal agents such as Meta's Muse or Instinct for product recommendations, which reduces the value of owning discovery websites in the past. Further downstream, instead of hundreds or thousands of people operating a linear funnel, long-running agents can handle far more complex journeys than before. They can work across products, across communication methods, and in parallel, support any language and operate at scale, with unlimited persistence. RV's EBITDA is expected to fall to about $160 million in 2025 due to customer churn and weakening advertising spend, which already offers an early sign of the model's fragility, and AI will further transform the entire industry.
Three paths to market entry
There are different ways to bring long-running agents to market:(1) Selling the platform: In this model, vendors sell long-running agents as software that enterprises can use within their existing operations. The plaintiff law firm case, where the agent was integrated with the existing team and its workflows, is a good example. Selling the platform offers revenue upside because there is usually no comparable software or outsourcing services contract to anchor pricing. But it also lengthens the sales cycle, since a new budget line must be created, and when an agent interacts with multiple departments across an enterprise, it is difficult to attribute outcomes clearly to it.
(2) Selling the service: Here the vendor takes over a service that has already been outsourced and delivers the desired outcome more efficiently, just as Red Ventures owns the entire customer acquisition journey for insurance companies. This approach shortens the sales cycle because the buyer's ROI calculation is simple and the BPO budget line already exists. However, existing pre-AI pricing models may not reflect the value that long-running agents can truly deliver, and many high-value journeys today are run in-house by enterprises, limiting the reach of this approach.
(3) Full-stack entry: Full-stack means vertical integration, becoming an AI-native player. For example, rather than selling long-running agents to a debt collection company, you could buy the debt portfolios yourself and collect on them yourself, so you capture the full upside from AI and solve the attribution problem. But it also means taking on licensing or compliance and other service layers that sit outside the actual customer journey, adding distraction and operational complexity.
It is too early to tell which approach will win, but we are confident that long-running agents will win in a large-scale relay race and expand the AI-addressable market from a slice of the labor budget to a slice of revenue. Plaintiff case preparation is just one example, and we expect agents to drive outcomes in other high-value journeys, such as mortgage origination, insurance claims processing, complex enterprise sales, and countless other scenarios. If you are capturing this opportunity, we would love to hear from you!