Keeping a complete case record across calls, police reports, and medical notes is the easy half of the relay. The hard part is the handoff into the demand letter: a sub-agent that finally reaches a witness on the fifth call can write that into the shared case file before anyone has reviewed whether the statement belongs in the chronology. I'd want the draft-for-review step to version those writes, so a later correction can show what the attorney actually saw versus what the agent appended after the letter went out.
"vendors can participate in the upside they create by pricing based on work products or outcomes." that only sticks in the plaintiff-law walkthrough if the agent also kept "a complete record of the case, from witness phone call transcripts to police reports and medical records." otherwise you're just arguing later about who closed the leak when the invoice shows up, which is the attribution fight Haber flags under outcome pricing.
Great piece. We see the same relay race in operations that never look like a "customer journey": onboarding a supplier, a client or a contractor. A bank letter arrives in the wrong name, a certificate has expired, legal hasn't signed, and someone has to chase it again. Then again.
One thing we've learned running these: the long-running agent isn't the hard part. The case is. Over days or weeks you need the case held open, fixed rules that decide what counts as done, and a record of every step and the rule behind it. That record is also what settles the attribution debate you mention, and that's what makes outcome pricing workable.
That's what we build at Floowed: the layer that runs the relay race, with our agents and anyone else's as the runners.
David, the post stopped me in my tracks. Especially #3 (Three Ways To Market). Felt like it was meant to be. We already built the agentic infrastructure at Relief. Our next step is going full-stack and buying the debt ourselves. Great post.
The interesting bit is buried near the end, not in the "press 3 for technical support" opening: intake specialists who call back fastest win the case, and that's a race decided in the first ten minutes. A long-running agent that babysits the journey for six weeks doesn't fix the part where the client already signed with somebody else on day one.
Still, the payment model is the real claim here. Getting paid on revenue expansion means the vendor suddenly has a financial interest in nobody falling off at handoff four — which is either the whole point or a lovely incentive to build software that harasses a plaintiff about medical records until they crack. Probably both. Did you want a take, or were you just leaving this on my desk?
The enterprise version has an extra leg at the very start, and the seller doesn't run it. The buyer's own assistant does the early research, same as you describe with Muse on the consumer side. So by the first demo the lead already has a picture of you, and sometimes it's wrong. Your agent then starts the call without knowing what they were told.
The "long running agent" concept has to sort out the same the same problem the automated driving cars did. Who is responsible for the errors and problems agents will inevitably do.
The AI usage policies say the user is responsible for any actions an agent takes. The less an agent is monitored, the more it can cause harm because of either LLM training problems, hallucinations or injected prompts.
Your article reads like the early days of driverless cars and Musk's statements. Long running agents are a long ways off because of the legal issues of the harms they can cause. Maybe an article on the realities given the history of driverless cars is relevant.
The same with the concerns over China leading in humanoid robotics. China is great at creating technology, China is very poor at making technology safe. Humanoid robots are a long way off because of the legal implications of any harm they cause. Chinese companies will not even consider this given they are mostly minimally funded startups.
AI is a great new technology. Use of AI without a human involved in any action the AI takes is fraught with legal issues.
The full-stack approach is interesting here. If an AI company can actually own the outcome, not just the software, it seems like the biggest opportunity might be building the business around the agent rather than selling the agent to an existing business.
Keeping a complete case record across calls, police reports, and medical notes is the easy half of the relay. The hard part is the handoff into the demand letter: a sub-agent that finally reaches a witness on the fifth call can write that into the shared case file before anyone has reviewed whether the statement belongs in the chronology. I'd want the draft-for-review step to version those writes, so a later correction can show what the attorney actually saw versus what the agent appended after the letter went out.
"vendors can participate in the upside they create by pricing based on work products or outcomes." that only sticks in the plaintiff-law walkthrough if the agent also kept "a complete record of the case, from witness phone call transcripts to police reports and medical records." otherwise you're just arguing later about who closed the leak when the invoice shows up, which is the attribution fight Haber flags under outcome pricing.
Great piece. We see the same relay race in operations that never look like a "customer journey": onboarding a supplier, a client or a contractor. A bank letter arrives in the wrong name, a certificate has expired, legal hasn't signed, and someone has to chase it again. Then again.
One thing we've learned running these: the long-running agent isn't the hard part. The case is. Over days or weeks you need the case held open, fixed rules that decide what counts as done, and a record of every step and the rule behind it. That record is also what settles the attribution debate you mention, and that's what makes outcome pricing workable.
That's what we build at Floowed: the layer that runs the relay race, with our agents and anyone else's as the runners.
Michel, co-founder, Floowed
www.floowed.com
David, the post stopped me in my tracks. Especially #3 (Three Ways To Market). Felt like it was meant to be. We already built the agentic infrastructure at Relief. Our next step is going full-stack and buying the debt ourselves. Great post.
The interesting bit is buried near the end, not in the "press 3 for technical support" opening: intake specialists who call back fastest win the case, and that's a race decided in the first ten minutes. A long-running agent that babysits the journey for six weeks doesn't fix the part where the client already signed with somebody else on day one.
Still, the payment model is the real claim here. Getting paid on revenue expansion means the vendor suddenly has a financial interest in nobody falling off at handoff four — which is either the whole point or a lovely incentive to build software that harasses a plaintiff about medical records until they crack. Probably both. Did you want a take, or were you just leaving this on my desk?
The enterprise version has an extra leg at the very start, and the seller doesn't run it. The buyer's own assistant does the early research, same as you describe with Muse on the consumer side. So by the first demo the lead already has a picture of you, and sometimes it's wrong. Your agent then starts the call without knowing what they were told.
The "long running agent" concept has to sort out the same the same problem the automated driving cars did. Who is responsible for the errors and problems agents will inevitably do.
The AI usage policies say the user is responsible for any actions an agent takes. The less an agent is monitored, the more it can cause harm because of either LLM training problems, hallucinations or injected prompts.
Your article reads like the early days of driverless cars and Musk's statements. Long running agents are a long ways off because of the legal issues of the harms they can cause. Maybe an article on the realities given the history of driverless cars is relevant.
The same with the concerns over China leading in humanoid robotics. China is great at creating technology, China is very poor at making technology safe. Humanoid robots are a long way off because of the legal implications of any harm they cause. Chinese companies will not even consider this given they are mostly minimally funded startups.
AI is a great new technology. Use of AI without a human involved in any action the AI takes is fraught with legal issues.
The full-stack approach is interesting here. If an AI company can actually own the outcome, not just the software, it seems like the biggest opportunity might be building the business around the agent rather than selling the agent to an existing business.