Back Ops bets end-to-end workflows beat partial automation in logistics

1. Whole workflows, not slices — ▶ 3:44

   A broken bench touches three functions: customer service takes the complaint, operations sources and reships a replacement, and finance files a claim and tracks recovery, since the company is now out two benches. Back Ops argues that automating only reshipment or only claim filing helps little. If a human does the first 25%, the agent takes 50 and the human then does another 20, the process is faster but still not off the team's plate. The company therefore targets complete workflows, starting in the back office and taking over more of it over time.

2. Browser operator by necessity — ▶ 5:04

   Most systems in this industry have no APIs. Carriers such as UPS and FedEx expose tracking but not claim filing, and third-party vendor portals expose even less. Where APIs exist, large enterprises resist granting unlimited read-write access to a vendor that has not earned trust. Back Ops therefore starts with a browser-based operator 60-70% of the time. Seeing the same workflows repeatedly also means it already knows which off-the-shelf tools get caught by bot detection, so it routes accordingly instead of redoing R&D for each task.

3. 60-80% reusable infrastructure — ▶ 6:41

   Sean says 60-80% of the core infrastructure is reused across customers, with 60-70% as the usual benchmark. The example is a temperature breach: a pallet of food or a failed refrigeration unit on a truck is handled by essentially the same machinery as breached cold-chain pharmaceuticals, which allows cross-vertical expansion. He is explicit that customization across stacks is still required. An auto manufacturer with hundreds of thousands of parts, pairing rules and minimum quantities shows how complexity grows. The reuse figure is a company benchmark, not an independently reported measure.

4. Process mining with outcome data — ▶ 7:41

   Onboarding began as a crude video recorder watched by a human who trained the agent. It has evolved into agentic onboarding, in which customers record a process and the system mines it. Sean distinguishes this from generic process mining, which offers basic suggestions such as dropping an Excel step. Back Ops says it applies ground-truth outcome data, for example advising that filing a certain claim type a certain way with a given carrier yields 11% higher throughput or approval. The 11% is an illustration given in passing, with no baseline or sample stated.

5. Billing gap and human touchpoints — ▶ 8:58

   Billing is the biggest enterprise opportunity Sean sees, provided customers will redo processes that have run the same way for years. In one case a customer's own analysis of a billing use case came to 440K, while Back Ops's analysis came to $4.98 million. That gap made the customer willing to rethink the process. A second lever is cutting unnecessary human oversight by writing rule sets for what reaches a person, such as a dollar threshold, a specific customer or a specific issue type. Both figures are speaker-reported.

6. Agents on both sides — ▶ 11:25

   Agent-to-agent communication in logistics is only beginning, but it is arriving. Back Ops has just signed one of the biggest US carriers, and one of the first things that carrier asked for was an agent on the other side, so claims would be filed on both sides. Sean says this is mostly written agent-to-agent exchange rather than voice agents calling each other. Agents already place calls to vendors and carriers, sometimes to people who do not know they are speaking to an agent. A claim with agents on both ends has not happened yet, but Sean says it is weeks away.

7. Escalation by email — ▶ 13:36

   Back Ops first built a portal where users could fill in missing context, but people did not use it. Its buyers are non-technical staff in customer service, operations and finance, and they may have 15 agents in flight. The fix was email, which is where they already work. Escalations are never blank. Each states what came in, what the agent tried, where it failed and the proposed next step, so a reply of yes or no suffices. The replies feed back into autonomous onboarding as commands. Sean's target is a handful of escalations out of 100 daily tickets, ideally edge cases or new customers.

Back Ops rests its business on a claim about where automation value sits. Sean argues that automating one step of a logistics problem barely changes a team's workload. A broken delivery sets off a chain: customer service takes the complaint, operations finds and reships a replacement, and finance files and tracks a carrier claim to recover the loss. If an agent handles only one link, humans still carry the rest. Sean's arithmetic is that when a person does the first 25%, the agent does 50 and the person then does another 20, the work is faster but is not off anyone's plate. So Back Ops goes after complete workflows, starting in the back office and widening from there.

Owning a workflow end to end is harder in logistics than it sounds, because the systems involved were not built for software to use. Sean says most have no APIs. Carriers expose tracking but not claim filing, and third-party vendor portals expose even less. Even where an API exists, large enterprises are reluctant to grant unlimited read-write access to a vendor that has not yet earned trust. Back Ops therefore starts with a browser-based operator 60-70% of the time. Repetition gives it an edge: having seen the same workflows many times, it knows which off-the-shelf tools get caught by bot detection and routes around them.

That repetition also underpins the economics. Sean puts reusable core infrastructure at 60-80%, with 60-70% as the usual benchmark, and customization making up the rest. His example is a temperature breach. A pallet of food or a failed refrigeration unit is handled by essentially the same machinery as breached cold-chain pharmaceuticals. If that holds, each new vertical should cost less to enter than the last. The caveat comes from Sean himself: customization across stacks is mandatory, and an auto manufacturer with hundreds of thousands of parts, pairing rules and minimum quantities shows how quickly complexity grows. The reuse figure is a company benchmark, not an audited measure.

The same accumulated data becomes part of the sales pitch. Onboarding began as a crude screen recorder that a human rewatched to train the agent. It has become an agentic process in which a customer records a workflow and the system mines it. Sean distinguishes this from generic process mining, which tends to offer basic suggestions like dropping a spreadsheet step. Back Ops says it adds ground-truth outcome data, such as telling a customer that filing a claim a particular way with a given carrier would yield 11% higher throughput or approval. The most striking example is a billing use case where the customer's own analysis came to 440K and Back Ops's came to $4.98 million. Both are speaker-reported, with no stated method.

The role of the humans is the second design problem. Back Ops first built a portal where users could fill in missing context, and people did not use it. Its buyers are non-technical operators in customer service, operations and finance. They are not versed in how much detail an agent needs, and they do not want to log in and find one of perhaps 15 agents in flight. The fix was email. Escalations are never blank. They state what came in, what the agent tried, where it failed and the proposed next step, so a reply can be a yes or no. Those replies feed back into onboarding as commands, which Sean says shortens iteration cycles.

The intake design shows how Back Ops thinks about autonomy. The agent is the first line across email, phone, text, WhatsApp, Slack and Teams. It classifies each issue with a confidence score, such as 93% that something is a short shipment, then follows a route. Each category has need-to-have and nice-to-have information, and if a need-to-have item such as an invoice or proof of delivery is missing, the agent contacts the sender itself. Sean's target is that only a handful of 100 daily tickets reach a person, ideally edge cases or new customers. Routine where-is-my-order emails, he says, are ones nobody should be answering.

The counterparty is where this model may be tested. Sean says a newly signed carrier, among the largest in the US, asked as one of its first requests for an agent on the other side of claims, so claims would be filed on both sides. He describes this as mostly written agent-to-agent exchange. A claim with agents on both ends has not happened yet, though he says it is weeks away. What the transcript leaves open is who operates the second agent, how conflicts are handled, and whether a vendor can serve both claimant and carrier. It also offers no error rates beyond a classification confidence example. As more of this traffic becomes software talking to software, the strength of the evidence matters more.