Insurance

Core systems have had thirty years of investment. Distribution has had a portal. Recruitment sits in spreadsheets, licensing in a regulator’s website, training in an LMS bought by a different department, sales in a CRM the agents avoid, and commission in a monthly PDF. Nobody is being careless — each of those was a sensible local decision. The cost only appears at the seams, and it appears three times: head office spends its capacity coordinating between systems, agency leaders manage a downline they cannot see, and agents spend selling hours on data entry.

A portal shows an agent their policies. Digitising distribution means the lifecycle itself advances — a candidate progresses through documents, licence verification, training and assessment without somebody chasing each step; an expiring licence raises escalating notifications before it stops someone selling; a neglected lead becomes an exception rather than a statistic discovered at quarter end. The difference is whether the software is a window onto work or the thing doing the work.

Growth in agency distribution is bounded by how quickly a recruited person becomes a producing one. That interval is consumed almost entirely by waiting: for a document, for a licence to verify, for a mandatory course, for a contract. Each wait is individually reasonable and collectively expensive. Compressing it is a systems question — configurable gates that genuinely block, evaluated at the moment activation is attempted rather than the last time anyone looked.

An agent’s authority to sell is a live state that changes with a date. Treating it as a periodic audit means discovering lapses after the fact, which is the expensive order to discover them in. Anticipation — expiry surfaced ahead of the event, with escalation, and enforcement that holds server-side rather than by hiding a button — turns a compliance exposure into a routine piece of scheduling.

An insurer spends real money approving product material — benefits, exclusions, waiting periods, comparisons — and then leaves an agent to remember it in front of a customer. The valuable move is to make approval a single act with two effects: the material becomes learnable by a person, and it becomes the only thing an AI assistant is permitted to answer from. That closes the loop between what compliance approved and what the customer is told, which is the loop that actually carries regulatory risk.

The useful question is not whether an insurer has AI. It is what the AI is allowed to do, what it is allowed to know, and who is accountable when it acts. In distribution the honest shape is narrow: monitor the conditions that matter, identify exceptions, assemble context, prepare the next step, act on low-risk work within an explicit ceiling, and escalate everything else to a person. A regulated decision — approve, decline, underwrite, adjudicate — is not on that list, and a platform that leaves the boundary vague is one somebody will eventually have to defend.

An insurer already owns a policy administration system, a rating engine, a claims system, a commission engine and a data warehouse. A distribution platform that positions itself as a replacement for any of those is not a serious proposal. The realistic model is a system of engagement over systems of record: canonical contracts on the platform side, adapters on the client side, and visible failure when a source is stale or unreachable rather than an empty result that looks like an answer.

For many insurers, and in many jurisdictions, regulated operational data cannot leave an approved environment — and that constraint has to survive contact with AI, which is where most platforms quietly fail it. The workable answer is that the platform runs where the insurer needs it to, and that the model provider is a configuration choice with a residency gate behind it, including running with no external model at all.