Score AI Agent Claims: Workflow Lock-In Beats Cost Cuts

BrainPath says its AI workforce cuts a workload that previously cost €7,500 a month in salaries to €40 a month. The number sells attention, not retention. The operator's question is harder: what makes the customer stay when the model improves, the price drops, or a rival ships a similar agent? A defensible product needs a durable role in the customer's process, not just a cost line.

Cost cuts do not create switching costs

If the value is only a lower invoice, the customer can leave when another vendor offers a better price. Roadmaps add urgency, but they do not add friction. BrainPath says its CFO and HR assistant agents are scheduled to launch in Q2 2026. Launch dates do not create switching costs. They can only show that the vendor has more tasks to sell. The operator should ask what happens after the first invoice.

Embedded agents pass the same tests

Three agent archetypes show the difference. Cost-cutting products sell cheaper output. They win on price, then lose when the task is simple enough for a general model. Workflow-embedded products sit inside a high-friction process, where the user must trust the output, review it, and feed the result back into the next step. Compliance-locked products accumulate process data, enforce guardrails, and become part of the audit trail. Demos that save money are easy to replicate. Workflows that require review, exceptions, and downstream steps are harder to replicate.

Evenup, an NFX-backed legal firm, integrated demand-letter drafting into the workflows of personal injury attorneys. The agent does not merely write a letter. It enters a case file, follows a legal process, and produces a document the attorney must review. The attorney remains the reviewer, so the agent is part of the case workflow, not a standalone writer. The more steps it touches, the harder it is to remove. The cost is narrowness. Workflows built for one profession may not scale to adjacent markets without rebuilding the process.

Larry Page and Sergey Brin developed PageRank in 1996. Google's daily search volume rose from 10,000 searches per day in 1998 to more than 42 million searches per day by 2000. The algorithm was the technical edge. The default path to information was the moat.

Groupon reached a $1 billion valuation in about 16 months and turned down a $6 billion buyout offer from Google. That speed did not guarantee a durable moat. It showed how quickly an offer can scale, and how quickly the market can move on. Fast offers can outrun weak workflows. Speed is not the moat.

Score the five moats before you buy or build

Score each claim from one to five. Products should not need all five to be defensible, but they should not rely on one. The scorecard is the operator's filter: it separates a cheaper task from a process that is hard to remove.

• Workflow lock-in. Does the agent sit inside a required step, not beside it? If the user can export the output and continue without the product, the lock is weak.
• Proprietary process data. Does the product learn from how the customer actually handles exceptions, approvals, and rework? Raw usage logs are not enough. Process data is the moat.
• Integrations. Does it connect to the systems where the work already lives? Each real integration raises the cost of leaving, but it also raises the cost of building.
• Compliance. Does it create audit trails, access controls, or policy checks that make the product easier to defend in a regulated environment? Compliance can be a barrier, but it can also slow adoption.
• Switching costs. Would the customer lose time, training, data, or accountability by moving? If the answer is no, the product is a tool, not a system.