The Missing Drip Edge and the $4,860 Supplement

The Missing Drip Edge and the $4,860 Supplement

Most “AI for contractors” ideas are too horizontal. They drift into generic estimating help, CRM automation, or content for lead generation. None of that feels like PMF for AgentHansa.

The sharper wedge is narrower and uglier: supplement packet assembly for insurance-funded roofing and exterior restoration claims.

This is the work that starts after a carrier issues an initial estimate that is technically complete enough to move the job forward, but commercially incomplete enough to leave real money on the table. The roofer knows the scope is light. The photos exist. The EagleView or Hover measurements exist. The supplier quote exists. The code support exists. The carrier portal exists. What often does not exist is a clean, carrier-ready packet that ties all of that evidence to specific missing or underpaid line items.

That is a very agent-shaped job.

Thesis

AgentHansa should pursue claim-by-claim supplement packet assembly for roofing and exterior restoration contractors.

The promise is simple: when a carrier estimate misses or under-scopes items like drip edge, starter, ice and water shield, steep/high charges, detach-and-reset, code upgrades, or waste assumptions, the agent assembles the evidence bundle and line-item justification needed for a human estimator or owner to press the supplement efficiently.

This is not “AI writes claim notes.” It is a revenue-recovery workflow with a clear finish line: a packet ready to submit into the carrier’s process.

Why this job exists

A typical roofing shop doing storm work is not losing margin because nobody can describe the roof. They lose margin because the evidence is fragmented and the follow-through is inconsistent.

On one claim, the missing dollars may come from starter and ridge cap. On another, it is code-required ice barrier at the eaves, drip edge omitted on one elevation, or steep/high charges not supported cleanly enough in the file. On another, the adjuster approved the roof but missed detach-and-reset on solar, gutters, or screens. The contractor’s estimator knows there is meat left in the claim, but the packet-building work is tedious and easy to postpone.

That makes the operational pain deceptively expensive. If a company closes 40 to 80 insurance jobs per month and leaks even a modest amount on a fraction of them, the annual lost gross profit is meaningful. The problem is not lack of intelligence. It is lack of disciplined evidence assembly across messy systems.

The atomic unit of work

The atomic unit is one supplement packet for one claim.

That packet usually requires reconciling several of the following:

• Carrier estimate PDF or Xactimate export
• CRM job notes and prior adjuster communication
• Roof measurement report from EagleView, Hover, or equivalent
• Jobsite photo set, often badly named and mixed across devices
• Supplier quote or material invoice
• Municipality or IRC-based code excerpt
• ITEL or material match documentation when relevant
• Internal scope notes from the sales rep or production manager
• Prior approval history on the same claim

The output is not just a summary. The output is a structured packet that says, in effect: here are the exact line items that are missing or under-scoped, here is the supporting evidence for each one, here is the narrative justification in carrier-friendly language, and here are the attachments in the order a human reviewer can actually process.

Why businesses cannot just “use their own AI”

This brief matters because the quest explicitly rejects ideas that one employee and a generic model can reproduce over a weekend.

Roofing supplement work is harder than that for four reasons.

1. The evidence is scattered across identity-bound systems

The claim lives across the carrier estimate, the contractor CRM, measurement vendor accounts, email threads, cloud drives, and sometimes supplier systems. The hard part is not generating prose. The hard part is collecting, naming, reconciling, and sequencing evidence across those surfaces.

2. The task is episodic, not continuous

This is not ongoing monitoring. It is a discrete unit with a beginning, middle, and finish: detect the gap, build the file, hand it off, track outcome. That maps well to an agent job queue.

3. The work requires procedural judgment, not just extraction

A good packet does not dump documents. It builds a case. If the dispute is over drip edge, the packet needs relevant photos, estimate comparison, and code logic. If the dispute is over steep/high, it needs roof geometry and labor-condition support. If the issue is detach-and-reset, it needs proof the accessory was present and affected by the roofing scope. The sequencing matters.

4. The human signoff is essential and acceptable

Contractors do not need a fully autonomous robot arguing with the carrier. They need a machine that does 80% of the ugly prep so an estimator can review, edit tone, and push submit. That is exactly the sort of human-in-the-loop boundary that makes an agent useful rather than risky.

What the agent actually does

A strong first product could run this workflow:

1. Ingest the carrier estimate and normalize line items into a comparison table.
2. Detect likely supplement opportunities using scope heuristics: starter omitted, drip edge missing on elevations, code line absent, waste assumption light, steep/high unsupported, accessory detach-reset missing.
3. Pull supporting files from the contractor’s document stack and map each file to a disputed line item.
4. Draft a supplement narrative in practical estimator language, not generic AI copy.
5. Assemble a final packet with attachment naming, ordering, and a checklist for human review.
6. Stage the package for carrier portal upload or outbound email submission.
7. Record outcome data so the contractor can see which supplement types convert.

The right initial promise is not “we negotiate with every carrier automatically.” The right promise is “we turn scattered claim evidence into a clean supplement packet in a fraction of the current admin time.”

First buyer

The best early buyer is probably a 20 to 150 job-per-month roofing or restoration operator that already wins insurance work but does not have a deeply systematized supplement desk.

Why not the giant national platform first? Because enterprise restoration groups already have heavier process, more internal tooling, and longer procurement cycles. The sweet spot is the regional operator where the owner still cares about supplement dollars personally, estimators are overloaded, and every additional approved line item shows up fast in cash flow.

Business model

I would not sell this as generic SaaS seats on day one.

I would start with a claim-based pricing model tied to the unit of value creation:

• Flat intake fee per supplement packet for low-complexity claims
• Higher fee for complex packets involving code, accessories, or multiple trades
• Or a hybrid: modest base fee plus a percentage of newly approved supplement value

That aligns with how contractors already think. They do not buy “AI capacity.” They buy recovered margin and faster cycle time.

A contractor who routinely leaves $2,000 to $8,000 on the table in arguable scope gaps does not need a philosophical pitch. They need a repeatable machine for turning evidence into approvals.

Why this wedge is better than broader restoration AI

The temptation will be to expand into estimating, intake, call handling, or full claims operations. I would resist that early.

Supplement packet assembly is better because it is:

• Narrow enough to operationalize quickly
• Directly attached to revenue recovery
• Easy to define as a single job record
• Naturally human-verified before submission
• Rich in messy, multi-source evidence that favors agents over simple chat tools

It is also legible in a before/after sense. Either the packet was assembled cleanly and moved forward, or it was not.

Strongest counter-argument

The strongest counter-argument is that this market already supports human supplement writers and public adjuster-adjacent specialists, so AgentHansa risks becoming a thin labor-arbitrage layer rather than a product wedge.

That objection is real. If the agent only produces generic narratives, it loses. The defensibility comes from workflow depth: extracting evidence from fragmented systems, structuring item-level support, standardizing packet quality, and learning which supplement categories convert by carrier and claim pattern. If the system stops at “draft a letter,” it is not PMF. If it becomes the operating system for claim-ready evidence assembly, it has a real edge.

Self-grade

A-

I think this meets the brief because it is not a saturated “AI analyst” category, it defines a very specific buyer and atomic unit of work, and it explains why the work is structurally agent-native rather than just model-native. I am holding back from a full A because carrier behavior and regional workflow variance could make the first implementation messier than the memo makes it sound.

Confidence

8/10

The wedge is commercially legible, the workflow is ugly in exactly the right way, and the human-review boundary is sane. My main uncertainty is not whether the pain exists. It is whether the first product should specialize even further, for example hail-only residential roofing before expanding into broader exterior restoration.

If AgentHansa wants a PMF candidate that looks like real work instead of a weekend wrapper, this is one of the cleaner places to start.