// Operations leader
Capacity without a hiring cycle.
Whatever volume growth you have been handed has to come out of throughput, because it is not coming out of headcount. The sequence matters more than the tooling: instrument first, automate the intake second, route the exceptions to someone who owns them, and staff the review layer before you need it.
// The situation
Two numbers describing the same year.
89% of carriers say they will increase or maintain staff, against projected industry employment growth of 0.78%. Technology, underwriting and claims are named as the greatest hiring needs.
Jacobson / Aon · Q3 2026 Insurance Labor Market Study.
Intent is not supply. The gap closes through throughput or it does not close.
So week one is counting: touches per file, cycle time by stage, rework rate and exception volume on the process you already run. That count is the deliverable before anything is automated.
// The stack you actually run
The automation baseline is lower than most people assume.
Only 38% of carriers auto-assign claims to adjusters, and only 16% have automated coverage opening and reserve setting.
Five Sigma · n=100 senior claims executives · 2022 data · dated.
Four years old. The count we take in week one replaces it with yours.
The reason the baseline stays low is not tooling. 72% of AI budget goes to technology and 28% to change management, and 47% of employees with AI access report no change to their workday after 18 months.
Capgemini World Property & Casualty Insurance Report · May 2026 · vendor research.
// How we sequence it
In this order, and the order is the argument.
Instrumentation before automation
Most AI programs cannot say what they saved, because nothing was measured before the tool went in. We instrument the current process first — touch counts, cycle time by stage, rework rate, exception volume — because you cannot defend a number you did not have before. Sometimes the finding is that the process is faster than you thought and the money belongs elsewhere.
Surface: your work queues, system timestamps and audit logs. Week-one deliverable, every engagement.
Intake automation
Document analysis and summarization, case analysis and submission ingestion are the three use cases with a real production track record. What we do not lead with: subrogation, fraud and FNOL scoring are heavily promoted, and we found no verifiable named-carrier production result for any of them.
Surface: the shared intake mailbox and the submission or claim record in your core system.
Exception routing that a person owns
Straight-through where the file supports it, routed to a named queue where it does not, with the reason attached. The value sits in the routing rules being explicit and editable by your team rather than held inside a model.
Surface: assignment rules in the claims or policy system, versioned.
The QA layer, staffed
Somebody reviews the output, owns the correction loop, and feeds errors back into the rules. We define the role, write the runbook, and train the people who will hold it — a priced line in the proposal rather than an assumption buried in a blended rate.
Surface: a named review queue, a runbook and a training pack that live in your documentation.
Two lines get funded in a softening market.
Cost per file and headcount avoidance. Both need a defensible baseline, which is why the counting comes first. The definitions: percentage of files touched, and by whom, before and after; cycle time by stage on one cohort definition both sides; rework rate; exception volume and the share resolved inside the named queue; and correction rate on AI output, reported rather than hidden.
// Our own operating record
We owned the expense ratio we were changing.
That is Clearcover’s operating record — ours, as the people accountable for it. Not a client’s, and not a forecast of yours. It sits on this page because the servicing expense line is the one an operations leader answers for.
Clearcover · verified July 2026.
Servicing is measured on headcount, throughput and an expense line, in front of a regulator and a live P&L. That is the frame these builds are scoped in.
// What's hard about this
Two things we tell you in week two.
The 72/28 split is the root cause and the hardest part to defend internally.
Change management, training, role redesign and the QA function read as soft-cost line items to a CFO in a cost-cutting year, and they are the difference between a deployed system and the 47% who report no change. So they appear in the proposal as explicit, named, priced lines with the role definition and the runbook as deliverables.
Your project team dissolves at launch.
Launch budgets fund development and implementation but not monitoring, retraining, documentation, user support or performance review. So the transfer plan sits in the first scope: exit criteria written at kickoff, an owner named on your side before go-live, and the runbook in your repository on the day the system turns on. How the transfer works →
// What ships with it
The operating model is a deliverable.
The QA runbook and the correction-loop definition. The exception taxonomy, with the routing rules and their change history. The role definition and the training material for the review seat. A named human decision-maker specification for every automated step that touches a regulated decision.
Roughly half the states have adopted the NAIC AI Model Bulletin. Connecticut requires an annual AI compliance certification attested by a named officer.
NAIC and state bulletins · current at September 2026.
What the governance file contains →Instrument it before you automate it.
The counting is week one on every engagement, and it occasionally ends the project early by proving the money belongs somewhere else. That result is in the report either way.