Automated Report Pipelines¶
Gen H has an untapped pipeline sitting in its existing tooling: Lightdash CLI for data, Claude Code for intelligence, and frontend-slides for presentation. Chain them together and you get a fully autonomous path from data to publishable report. Three report categories are ripe for this. Funder reports are now proven - Nottingham Building Society is fully automated.
Progress: Funder Reports¶
What's Done¶
Nottingham Building Society - Gen H's biggest funder - is fully automated as of 12 Feb 2026. Run the script, get the report. The output is slightly more accurate, better presented, and completely removes the manual workflow. This is the proof of concept for the entire pipeline.
What's Next¶
Additional data points. Now that the pipeline is automated, the capital markets team want to add metrics they'd always wanted but couldn't justify when each addition meant more manual steps. Automation has unlocked scope that was previously too expensive to include. This is a nice example of the broader pattern: automation doesn't just replicate the manual process, it enables things that were impractical before.
Remaining funders. Five more to go. These should be easier for two reasons: the pipeline pattern is built, and the remaining funders are simpler than Nottingham. The goal is for each new funder to be a config exercise, not a development exercise.
Handover to capital markets. This is the critical unsolved problem. Hal built this, but the capital markets team need to own it going forward. "Run the script" isn't enough - they need the ability to:
- Add or tweak custom content (commentary, context, narrative sections)
- Adjust look and feel (formatting, branding, layout)
- Add new data points without waiting for Hal
- Maintain per-funder customisation as funder requirements evolve
The current thinking is packaging this as a Claude Code skill, which would give the team a natural-language interface over the pipeline rather than requiring them to work with scripts directly. This is the right instinct - the gap between "Hal automated it" and "the team owns it" is where most internal tooling dies. A skill wrapper turns a personal automation into a team capability.
Questions worth answering on the handover:
- How technically comfortable is the capital markets team with Claude Code? Have they used it, or is this introducing a new tool alongside a new process?
- What's the failure mode if the skill breaks or produces wrong numbers? Is there a human review step, or does this need to be bulletproof before handover?
- Should the skill be opinionated (guide the user through choices) or flexible (let them do whatever they want)? Opinionated is faster to build and harder to misuse.
The Three Report Types¶
1. Funder Reports - IN PROGRESS¶
Reports for external mortgage funders (financial institutions). Structured, predictable, but with per-funder customisation:
- Each funder cares about slightly different metrics or defines them differently
- Time frames vary by funder
- Presentation needs to be polished - these go to institutional partners
- Lowest complexity, highest immediate credibility impact
- Nottingham Building Society: complete. Five remaining funders: queued.
2. Network and Corporate Account Reports - NOT STARTED¶
For the broker sales team to share with networks and corporate accounts. Similar shape to funder reports but more opinionated:
- "What are all the brokerages under this network doing with Gen H?"
- Highlights what's different, what's interesting, what stands out
- Based on standardisable metrics across networks, but surfaced with editorial judgement
- Medium complexity - the data is structured but the insight layer needs curation
3. ExCo Management Information - NOT STARTED¶
The big one. Currently an enormous spreadsheet with thousands of data points across multiple tabs, maintained by many people monthly, read by almost nobody in a meaningful way.
The goal: ingest the full surface area of MI data and condense it into a digestible report that highlights what actually matters this month. Specifically:
- Commercial funnel changes (always important to management)
- Key metrics moving significantly in either direction
- Anything that warrants management attention rather than routine monitoring
This is less about reporting and more about automated editorial judgement - taking a wide, noisy dataset and producing a focused narrative. Hardest of the three, most valuable if it works.
The Pipeline¶
Lightdash CLI (data/queries) → Claude Code (analysis/opinion) → frontend-slides (HTML presentation)
↓
Hosted link + PDF export
Each report type is fundamentally: a set of Lightdash queries or charts, processed through Claude for structure and (where needed) editorial judgement, then rendered as a frontend-slides presentation. HTML works for hosting; PDF for document distribution. Both are needed.
Rollout Phases¶
Phase 1 - Supervised generation: Claude produces reports autonomously, a human reviews and tweaks, feeds corrections back as instruction changes. This is where the pipeline gets calibrated. Funder reports are through Phase 1 for Nottingham.
Phase 2 - Automated distribution: Once report quality is predictable, automate the sending on a defined cadence to defined recipients. Funder and network reports will reach this phase before ExCo MI, which needs more iteration on the editorial layer.
Where This Gets Interesting¶
The real value isn't any single report - it's the general-purpose pipeline. Once the pattern is proven (data -> analysis -> presentation -> distribution), it becomes reusable for any recurring report across the business. The time currently spent on manual report production, formatting, and distribution is substantial and spread across many people.
The Nottingham result already demonstrates this: automation unlocked additional data points that were too expensive to include manually. Every report type will likely have its own version of "things we'd love to include but can't justify the effort". The pipeline makes them all free.
What Needs Sharpening¶
The handover problem is now the bottleneck for funder reports. The technical pipeline works. The open question is how to transfer ownership to the capital markets team in a way that sticks. A Claude Code skill is the leading approach - it needs to be scoped and built.
The ExCo MI "opinion layer" is the hard problem. Pulling data from Lightdash is solved. Rendering HTML is solved. The unsolved piece is: how do you reliably teach Claude what "interesting" means to Gen H's management team? This likely needs:
- A set of heuristics or rules about what always matters (funnel metrics, lending volumes, pipeline changes)
- Threshold-based flagging (anything moving more than X% month-on-month)
- Context from previous reports (what was flagged last month, has it improved or worsened?)
- Iterative calibration from ExCo feedback on early reports
PDF export from HTML needs a mechanism. Frontend-slides produces HTML. Converting to PDF reliably (preserving layout, animations flattened to static) needs a tool - likely headless Chrome/Puppeteer or similar. Not hard, but needs to be part of the pipeline.
Connections¶
- CRM Stack Migration - if Attio properly models Gen H's broker and network relationships, the data feeding into network/corporate account reports gets significantly better. CRM migration is upstream of report quality.
- Underwriting CLI - the funding reports pipeline is now the first shipped Mode 3 tool at Gen H (see the three modes framework in the underwriting CLI doc). It proves the pattern and demonstrates the handover challenge that applies to underwriting too.
- Sales Intelligence CLI - complementary. Report pipelines automate recurring reports; sales intelligence enables bespoke, ad-hoc report creation by salespeople themselves. They share the same infrastructure (LightDash CLI, Claude Code, frontend-slides). The network/corporate account reports already planned here are the templated version of what salespeople would create on the fly there.