Building the digital operating layer of a small insurance broker
What runs today, what’s still being built, and how the pieces connect
At a glance
- Problem
- People were carrying work by hand between disconnected tools.
- My part
- Business discovery, product and technical decisions, agency selection, hiring and managing two contract developers, and later AI-assisted delivery. Hands-on integration and infrastructure work throughout.
- What changed
- The website, CRM, communication tools and management reporting are running. Product releases still have dependencies; their status is shown below.
- A different way to report
- Instead of requesting each answer from a separate function, management can ask questions across the connected systems.
The work between the tools
Since April 2025 I’ve been helping rebuild the digital layer of a small insurance broker. When I started, work was spread across a CRM, spreadsheets, messaging, email, telephony and an ageing website. Commissions were processed manually, and there was no coherent B2B/B2C/Admin product layer.
Read more
None of those tools was disastrous on its own. The problem was all the work happening between them.
- CRM
- Old website
- Messaging / Email
- Commission spreadsheets
- Telephony
- Manual reporting
People carried the work between tools: messages → spreadsheet rows → checks → reports.
No coherent B2B / B2C / Admin product layer.
My role
I started by learning how the business worked: who did what, where information was copied between tools, and which problems needed a product rather than another workaround.
Read more
I worked with sales, advisors, marketing and management to turn that into requirements, priorities and acceptance criteria. I defined the system boundaries and technical approach, evaluated agencies and implementation partners, and coordinated the design agency’s branding and website work.
I hired two contract software developers for the product work, organized their tasks and dependencies, reviewed their work and coordinated delivery. Later, I moved a substantial share of implementation into specialized AI-agent workflows. I still owned the product decisions, unresolved dependencies and release decisions, with independent review of the implementation.
Alongside that coordination, I worked hands-on on the platform, integrations and infrastructure. The delivery team changed as the work grew; learning what the business needed remained part of the job.
How the product work gets built
A substantial part of implementation now runs through specialized AI agents, coordinated in Paperclip. I turn requirements into scoped tasks and acceptance criteria, map dependencies and keep the work aligned with the business priorities.
Paperclip tracks the dependencies, ClickUp mirrors the operational work, code and reviews live in source control, and Outline holds documentation and decisions. A handoff needs to explain what changed, what the next piece depends on and how to check it.
Agents leave code, test results and review artifacts. Independent QA checks the exact changes against the acceptance criteria, including tests and browser checks in an isolated environment. A person reviews that evidence and authorizes a release.
Requirements and acceptance → dependency planning → specialized implementation → independent QA → reviewed evidence → authorized release.
I also maintain the agent runtimes, test environments and recovery paths. Sentry and Grafana are implemented and help make errors and runtime behaviour visible. These are working parts of the delivery process; a completed tracker task still needs evidence before it can be treated as a working change.
What changed
This is ongoing work. Some systems are live; the product layer still has release dependencies.
Read more
- Team communication
- CRM
- New website
- Telephony
Business workflows ↕ product and operations layer.
- B2B
- Admin
- B2C
- Commission tool
Cross-system questions → reporting · product work → reviewed software delivery.
- Dashboards
- Management reporting
- AI-assisted delivery
The rows group the operating layer; they do not imply that every tool has a direct integration with every product.
A concrete example: commissions
Read more
The manual process requires accounting to work through deals more than once:
- A closed deal is entered into Excel.
- Commission is calculated using product, team and partner rules.
- When the insurer’s payment arrives, the accountant reviews the rows again.
- Eligible commissions are paid.
The dedicated tool is being built to represent those rules and move processing out of Excel. It has not yet replaced manual reconciliation in production.
Three products, three audiences
- B2B — for business partners
The partner platform has substantial implementation in place and is being prepared for beta. Release dependencies remain open.
- Admin — for the internal team
The associated administration product has been implemented; its release is tied to B2B. Implementation readiness does not mean it is already in production.
- B2C — for customers
The customer product has a core in place while feature requirements continue to evolve. It is at a different stage from the partner and admin products.
From Slack questions to recurring operations
One part of the work began as a Slack assistant for marketing questions. It grew into recurring workflows that bring advertising, CRM, call and website signals together, produce reports and suggest warm-lead follow-ups for advisors.
APIs, MCP connections and custom scripts connect the sources. The systems of record still own the data. Dashboards cover familiar questions; the reporting workflows combine evidence for growth reviews, telephony analysis and questions that cross teams.
The harder work was making those runs repeatable: handling inconsistent source data, retaining evidence, recovering from failures and checking that the output reached the person who needed it. The runtime evolved as the original setup became too coupled to operate confidently.
Recommendations and external actions have different consequences. The workflow requires human authorization for CRM changes, customer communications and advertising spend; that responsibility remains with people.
Read the sources
Bring together CRM, advertising, call and website signals through APIs, MCP connections and custom scripts. The source systems keep their own records.
Aggregate, then analyse
Add up the source data first. Then review it from marketing, sales, PPC and SEO perspectives, with a second pass that tries to poke holes in the conclusions.
Report to management
Answers to specific questions, a daily growth review and monthly telephony reports. Dashboards cover the questions asked regularly.
Decide before acting
The workflow calls for human authorization before changing CRM records, sending communications or changing advertising spend. A recommendation stays separate from the action.