Insurers, reinsurers and brokers have bought Business Process Outsourcing for decades on a single premise: the same task costs less when it runs on someone else's lower-wage bill.
Agentic AI removes the premise, because it changes what the work is.
BPO is language work
Strip away the delivery centers and the org charts and BPO is language work. Submissions arrive as emails and PDF attachments. Claims arrive as narratives, photographs and adjuster notes. Policy administration means moving information between formats, systems and counterparties. The industry called it labor because for 50 years only a person could read a broker's slip, interpret a loss description or key an endorsement.
Agentic systems can now do language work. Intake, indexing, extraction and data entry, the tasks that fill BPO delivery centers, sit squarely inside what these systems handle today.
One condition applies. The system must be grounded in insurance. Domain-specific language models paired with an insurance knowledge graph, the approach we take at mea, read the market's documents the way an experienced technician does, at machine speed and machine volume.
Exceptions used to break automation. Now they route.
The industry has automated before. Robotic Process Automation moved structured data between systems by imitating keystrokes. Optical Character Recognition turned scanned documents into searchable text. Both broke the moment reality deviated from the template, and those breakages are precisely where BPO labor took over.
Agentic systems treat the exception as part of the workflow. The system executes the repeatable work, flags the consequential decisions and routes them to a person through an operations dashboard. The human in the loop is designed into the process. People supervise the flow, resolve the flagged items and own the outcomes. Work moves continuously, and human attention goes only where judgment is required.
We call this model supervised autonomy. The system runs the transactions; people run the system. It preserves the cost advantage that drew firms to BPO and restores the control they surrendered to get it.
The economics invert
Labor-based BPO starts expensive and grows more expensive.
Wages rise, attrition forces continuous retraining and every increase in volume demands more headcount. Compute starts cheap and gets cheaper every year. When agentic systems execute the transactions, human effort concentrates where it earns the most: supervision, exception handling, judgment and ownership of the process. The cost curve bends down while capacity scales up.
Regulators get a better audit trail
Supervised autonomy suits regulated industries because the audit trail is a byproduct of doing the work. Properly implemented, an agentic system logs every action from inception: each extraction, each decision, each exception and its resolution, recorded and explainable to regulators and reinsurers.
A labor-based operation can be made auditable, at the cost of documenting and verifying the actions of hundreds of people across multiple systems and shifts. One model produces evidence as it runs; the other produces evidence through effort.
Choosing the right foundations
General-purpose language models made all of this possible, and they belong in every insurer's architecture.
Operational progress comes from the layer above them: models trained on insurance transactions, paired with an insurance knowledge graph, that read a slip, a bordereau or a claim file the way an experienced technician does. Domain grounding is what converts capable AI into accurate operations, and it is the standard to hold any system to before it touches live workflows.
The remaining question is pace. Supervised autonomy delivers lower cost, faster cycle times, full control and an audit trail built in. Traditional BPO delivered one of the four. For insurers, reinsurers and brokers, the decision now concerns pace: how quickly transaction execution can move onto the new model.
