The choices that kept an operating responsibility in the room

In February 2025 we tested a proposed claims-correspondence assistant against 1,800 closed property files. The commercial proposition assumed that routine letters could be released with little intervention. Reviewers classified 1,440 drafts as having the correct supporting references, 216 as using an outdated endorsement and 144 as applying a reference from the wrong jurisdiction. Those categories covered the test set; a correct reference was not itself a complete assessment of a letter's insurance judgement.

The sponsor wanted to retain the original automation ambition. Claims colleagues were concerned that fluent wording would make a wrong reference harder to notice. I recommended a narrower release: the assistant could prepare a draft, but an authorised colleague had to inspect its source material and approve the communication. We accepted a USD4.6 million programme with a smaller initial scope, retained the review effort and moved automated dispatch out of the approved release. That was a material reduction in the promised convenience.

My team separated source retrieval, proposed wording and the release action. Claims and legal colleagues defined what the reviewer needed to see. We also recorded the reason for a substantive rewrite, so that the pilot would not report every approved final letter as an untouched machine success. The distinction made review slower at first, and some users asked for the extra record to be removed. We kept it while the evidence was immature.

In the subsequent supervised pilot, eighty-three of nine hundred letters were rewritten before release. Median preparation time for that cohort fell from twenty-two minutes to fourteen, excluding the later claim decision and any further investigation. Every letter still had a human approver. We did not extrapolate the observed rewrite rate to all products or describe the tool as an independent decision-maker.

I would involve experienced reviewers in the initial product brief next time. We spent too long debating whether people would remain in the process before specifying what a useful review actually required.

Our spring 2025 renewal pilot invited 7,600 policyholders to use a redesigned online service. The initial business case anticipated that most would complete digitally and treated telephone and posted responses as residual demand. Operations colleagues warned that the remaining work would still need careful handling, particularly where someone changed channel after starting online. The risk was a budget that counted the digital improvement while quietly removing the capacity needed for everyone else.

I supported an explicit assisted-service provision of USD540,000, retaining the capacity of eight colleagues through the pilot period. The programme sponsor challenged the cost because it made the digital business case less impressive. My response was that the insurer still owed a coherent renewal service to those members. A lower cost per online completion would not answer whether the whole renewal population had been served.

Technology and operations designed a shared record of the member's request, details already provided and the next responsible colleague. We tested handovers from the web form into a telephone conversation and from a posted instruction into the service queue. Product and legal specialists retained responsibility for the renewal communication and its terms. My team concentrated on preventing repeated questions and contradictory status messages when the channel changed.

Of the invited population, 6,100 completed through the online route and 1,500 used telephone or post. Those figures described completed renewals in the pilot; they did not establish why each person preferred a channel or prove that digital access had improved for every group. The assisted provision remained a real cost. We used the split to plan staffing and examine avoidable handover work, rather than classifying the 1,500 members as a failed conversion opportunity.

I would collect better evidence about channel changes at the start of another pilot. We knew the final route more reliably than the reason somebody left the first one, and that limited which design changes we could justify.

A severe regional storm in August 2024 generated 9,600 new claim notifications in seventy-two hours. During triage, claims supervisors identified 390 cases considered urgent that could not move through the ordinary intake form because a property photograph was mandatory. Some members could not safely reach the property. Others had submitted those details by telephone, leaving a colleague to work around a form that assumed an upload would be available.

The original requirement had a sensible purpose: photographs often helped a claims specialist understand the loss. Removing it everywhere would have discarded useful evidence. Leaving it mandatory in every circumstance would have allowed the form to determine which case reached a qualified person. I proposed a restricted exception route, activated by an authorised claims supervisor and recorded with a reason. The claims leader agreed to own its use; technology did not decide whether a loss was covered.

We spent USD280,000 on the intake change, testing and temporary support. The release displaced two lower-priority service enhancements from that month's schedule. My team made the missing evidence visible in the case record and preserved a follow-up task, so that bypassing the upload could not be mistaken for having supplied a photograph. Claims supervisors reviewed the 390 cases through the authorised route over the following thirty-six hours. That was a triage result, not a count of paid or accepted claims.

We kept the exception available under a defined operating procedure after the immediate event. The process required a named reason and retained the subsequent evidence request where appropriate. We did not conclude that every compulsory field should become optional. The lesson was to distinguish what someone needs for a first human assessment from what may be required for a later insurance decision.

Next time I would rehearse the route with service colleagues before catastrophe season. The initial design had been tested with complete example files, which concealed the conditions under which the most important member could not finish it.

A proposed expansion of assisted drafting used a simple benefit assumption: less time composing a letter would create an equivalent amount of available claims capacity. The claims leader was unconvinced because colleagues would also have to inspect sources, correct wording and record why they had changed it. I agreed to fund a six-week work study with sixty adjusters before committing to the larger capacity claim. Participation covered an observed cohort, not a random sample of every claims role.

We compared the tasks those colleagues performed before and during supervised use, keeping drafting and verification separate. The study found an average forty-two minutes a day released from drafting and eighteen minutes added through checking. That left twenty-four minutes in the observed working day. The time was not automatically available for another file: interruptions, the timing of calls and differences in case complexity still mattered.

The programme team wanted to translate the total into a staff-equivalent saving. I declined to put that number in the approval paper before operations had changed the work. The claims leader assigned fifteen of the twenty-four minutes to proactive customer follow-up and left nine unallocated while examining how unevenly the benefit was distributed. We retained the existing staffing plan. Finance could see a more modest operational proposition, including a USD160,000 measurement and training cost, rather than an immediate payroll reduction.

The outcome was a better allocation conversation. Supervisors could discuss whether customers were receiving the intended follow-up, and the technology team could see which checking tasks remained cumbersome. We had not demonstrated the same result for complex commercial claims or for colleagues outside the sixty-person cohort. Nor had the study isolated every other change occurring during those six weeks.

I would start with a more explicit comparison design in a future study. The exercise was sufficient to challenge an unsupported staffing claim, but it was not a controlled experiment from which a universal productivity estimate could responsibly be drawn.

A review of external claims-support access in November 2025 found 480 supplier user accounts across twenty-five partner organisations. The contracts described services at an organisational level, while the access records described individual users. Neither record alone established whether a particular person still needed to see a particular set of files. The claims function retained responsibility for delegated work; my team needed to make system access reflect that current instruction.

Working with fifty-six internal relationship leads, we identified eighty dormant accounts that could be closed and sixty further accounts requiring restricted access while their assignments were clarified. The remaining 340 continued under their existing approved scope after review. Dormancy was a reason to examine an account, not an accusation that its holder had misused information. We made no claim that the exercise had discovered a data breach or resolved every supplier risk.

Several partners objected to the additional confirmation work and warned that an indiscriminate shutdown would interrupt live files. I supported a staged review costing USD180,000 rather than a single deadline for removing all uncertain users. Claims supervisors identified time-sensitive work, relationship leads confirmed the people assigned to it, and technology applied the corresponding access changes. Restrictions and reinstatements had named approvers and an expiry for further review.

The result was a reconciled account population with a clearer connection to the work being performed. Eighty accounts were closed, sixty were restricted and 340 retained approved access. Those numbers describe the decision at that review date. They are not evidence that future joiners, transfers or departures will be handled correctly without continuing ownership. We therefore assigned the ongoing confirmation task to the relationship process rather than treating it as a completed technology clean-up.

I would have involved partner supervisors earlier. Some of the initial disagreement came from asking a corporate contact to confirm details that only the person allocating daily claims work could reliably explain.