It is 4pm on a Thursday. You have spent most of the day on things that needed to happen but did not actually need you. Chasing a document trail for a renewal. Rewriting an internal checklist that already exists somewhere on a shared drive. Drafting an email almost identical to one you wrote last month. The work that requires your expertise, the conversations and the judgement calls, got squeezed into the margins.
This is the part of the insurance day that does not appear in any productivity report, and is the single biggest opportunity AI offers anyone in the industry right now. McKinsey’s analysis of generative AI in financial services flags claims, underwriting and policy operations as among the highest-impact applications across banking and insurance, with potential annual value over USD 1.1 trillion across the sector. That number is for the industry. The version of it that matters to your career is the four-to-six hours a week you can move from admin into the work that actually earns and renews business.
The mistake most people make on day one with AI
People open ChatGPT for the first time, type something vague, get something vague back, and conclude AI is not for them. The shift happens when you treat the tool the way you would treat a sharp new analyst on their first day. Brilliant, but reliant on you for context, scope, and the specific outcome you want. Once you give it that, a workflow that has been living scattered across your inbox, your head and a half-finished spreadsheet becomes a clean, shareable, team-ready document in five minutes.
The other tool worth knowing about for this kind of work is Claude, which tends to be stronger at long structured documents and at holding context across a multi-step conversation. Use either. They produce broadly similar quality on this category of task.
For sector-specific automation that handles policy data, claims documentation, and underwriting workflows directly inside your operating systems, Guidewire and Shift Technology ship AI features built on insurance-specific models. Those tools are the next step. The first step is what you can do this afternoon with a chat window.
The workflow-building prompt that actually works
Here is the prompt that turns a scattered process into a usable document. Open ChatGPT or Claude and try it on something you actually do every week:
I am an insurance [BROKER / ACCOUNT MANAGER / UNDERWRITER] and I need to create a clear internal workflow for [SPECIFIC PROCESS, e.g. managing personal-lines policy renewals]. Here is roughly how it currently works: [DESCRIBE IN 6 TO 10 BULLET POINTS, INCLUDING WHO DOES WHAT, WHAT TIMINGS APPLY, AND WHERE THINGS COMMONLY GO WRONG]. Please write this up as a step-by-step workflow document suitable for sharing with a small team. For each step include: who owns the step, what trigger or input starts it, what information or documents are needed, what timing or service-level applies, and what the success criterion looks like. Flag any steps where things commonly go wrong in practice and suggest specific safeguards. End with: a checklist version of the same workflow on one page, and a list of three places this process is likely to break under load.
What comes back is not a rough sketch you have to fix. It is a polished document with checklist, escalation logic, and a list of failure points, written in your professional voice if you described the process in your own words.
That same document, run back through the model with a follow-up prompt, becomes a staff training guide, a client-facing summary, or an email template set:
Take the workflow document above and produce three derivative versions: a one-page training brief I can hand to a new hire on their first morning, a short client-facing summary explaining how we manage their renewal so they know what to expect, and a set of five email templates with bracketed placeholders for the routine touchpoints in this process.
One process. Four artefacts. About ten minutes total.
A counterintuitive insight: the document is the deliverable
Most insurance professionals think of process documents as overhead. The version that actually changes your career is the one where the document becomes the artefact your team standardises on, with your name on it. Two consequences. The first is that the process now runs the same way whether or not you are in the room, which is a meaningful step away from being the bottleneck. The second is that you have a written, reviewable artefact that demonstrates a specific kind of senior judgement: the ability to translate tacit knowledge into something a team of three can execute consistently. That is exactly the capability principals are looking for when they think about who to put on more strategic work.
A worked example. A broker working in commercial lines used the prompt above on their cyber-renewal process, which had previously lived as “Sarah’s brain plus a Trello board”. The output, lightly edited, became a 4-page workflow document, a one-page checklist, and a five-template email pack. Total time: 35 minutes, including a 15-minute revision pass. Six weeks later, that document was the agency’s standard for cyber renewals, and Sarah had been moved onto the agency’s broader cyber strategy work because the renewals queue was no longer entirely dependent on her.
The failure mode to watch for
Asking AI to write a process you have not yet thought clearly about. The model will produce something polished and confident, and you will paste it into your shared drive, and three weeks later you will discover it is wrong in two places that matter, because the failure happened in your description, not in the tool. The fix is to start with bullets, not prose, and to be specific about where things break. If your description of the process does not include “this is the step that goes wrong when [SPECIFIC THING]”, the document AI generates will not anticipate it.
A second failure mode worth naming, particularly in regulated work: do not paste real client data, real policy numbers, or anything personally identifiable into a free-tier general-purpose model. Use redacted examples. Bracketed placeholders. Synthetic data that follows the same structure. If you need to work with real records, the right path is the version of the tool your firm has approved for that purpose, not the public chat.
Why this matters now, specifically
Regulation is increasing. Client expectations are rising. The professionals who keep more of their week available for the work that requires actual judgement are the ones who will have headspace to grow, take on more strategic accounts, and move up the value chain. Every hour reclaimed from process drift is an hour you can spend on the renewal that needed more than a phone call, or the new business conversation you keep meaning to have.
The smartest move this year is not to learn every AI tool. It is to take three of the workflows that currently live in your head and your inbox, run them through this prompt, and put the artefacts somewhere your team can see them. The compounding starts there. The career consequences follow.