SB StartupBasket
All ideas
77 /100 GO Medium complexity

BookTriage — ANOC risk ranker for Medicare agents

Reads every client's ANOC against their drugs and doctors, then tells the agent which fifty to call first.

InsurTech SaaS US AI-agentSMBSolo-builderSeasonalRetentionMedicare
— views
Evaluation Scores
77/100

GO

Overall Score

17
Problem
13
Demand
11
Build
13
Distrib.
12
Revenue
6
Time
5
Defense

BookTriage

1. One-liner

Reads every client’s ANOC against their drugs and doctors, then tells the agent which fifty to call first.

2. Trend signal — why now?

Three things collided in the last twelve months, and the third one is what makes this buildable.

The Medicare Advantage market is convulsing. Plan exit rates surged to 6.9% in 2025 and then leaped to 10% for the 2026 plan year. Nearly 3 million seniors were forced off their Medicare Advantage plans in 2026 due to insurer exits, with top carriers withdrawing from over 850 counties between 2025 and 2026. CMS itself projected MA enrollment would fall in 2026 as insurers cut unprofitable plans. This is not a normal year of tweaks. Networks are being redrawn, formularies re-tiered, and whole plans deleted underneath people who don’t read their mail.

The money at risk is precisely quantified. CMS set the 2026 national maximum broker compensation at $694 initial / $347 renewal per member per year for MA, and $114 / $57 for PDP. An agent with 600 clients is sitting on roughly $208K of annual renewal income that silently evaporates if clients get orphaned, churn to a competitor, or land on a plan that stops paying. Separately, carriers cut broker commissions on 15–20% of plans to discourage enrollment in unprofitable ones — so some of the book is now actively worth less, and the agent can’t tell which part without checking plan by plan.

The data to solve it just became free and machine-readable. CMS publishes Monthly Prescription Drug Plan Formulary and Pharmacy Network Information files — plan info, geographic locators, formulary down to NDC level, cost-share tiers, and flags for step therapy, quantity limits, and prior authorization. And from 1 October 2026, CMS requires every Medicare Advantage plan to publish machine-readable provider directory files for Plan Finder, crawled daily as published files. The two hardest inputs — is my client’s drug still covered, at what tier and is my client’s doctor still in network — stop being phone calls and become a download. That is a genuine unlock, and it lands eight weeks before this product’s first selling season.

The timing is not subtle: ANOCs must reach beneficiaries by 30 September, AEP runs 15 October – 7 December. This proposal is dated 31 July 2026. There are roughly eight weeks to build before the exact moment the pain peaks.

Provenance:

3. The opportunity

The incumbents built for the wrong unit of work.

Every tool in this category — MedicareCopilot, Sunfire Matrix, ConnectureDRX, the CRMs — is built around one client you are already sitting with. MedicareCopilot’s pitch is explicit: enter a client’s medications and providers, and its 12-factor engine scores every plan and returns a ranked recommendation “in under 60 seconds.” That is excellent, and it is the wrong shape for the actual bottleneck.

The agent’s problem in October is not what should Marjorie do — given an hour and Marjorie on the phone, any competent agent solves that. The problem is which of my 600 clients is Marjorie. Which forty people in this book are on a plan that’s disappearing, or had their maintenance drug pushed from tier 2 to tier 4, or whose cardiologist just fell out of network. The agent has one 54-day window and cannot call everyone.

And the honest ones admit they don’t call everyone. From an agent forum: agents with large books avoid proactively contacting all clients because it creates “a flood of unnecessary contact” — if the existing plan is still best, the client doesn’t need to do anything, so agents “prefer to be selective and focus on clients who actually need their help.” That selectivity is currently a guess. There is no tool that ranks the book by risk. The industry’s own advice is to “pull a list of all Medicare clients by carrier and plan type” and “flag anyone who moved plans last year” — a crude proxy for the thing that actually matters, which is whether this specific person’s drugs and doctors survived the plan year.

So the gap is: a triage layer that runs once, across the whole book, before the season starts, and outputs a ranked call list with a reason attached to each name. Incumbents own the last mile. Nobody owns the first one.

There’s a second, nastier version of the same gap. When a plan goes non-commissionable or a carrier exits, the agent’s income changes without any notification tied to their client list. Agents currently discover this by reading carrier bulletins and cross-referencing manually. Mapping commission changes onto an actual book is pure data work.

4. Target market

  • Primary customer: Independent Medicare agents and small agency principals in the US running 200–2,000 client books — typically 1–10 licensed agents, often an agent plus an assistant. Includes small FMO downlines and captive-adjacent independents. The sweet spot is the 400–1,200 client solo or duo agent: big enough that manual review is impossible, small enough to have no analytics staff.
  • Why they buy: Their renewal income is a function of retention, and retention is a function of calling the right people before someone else does. At $347/member/year, losing 30 clients costs about $10K a year, recurring. One prevented churn pays for the product for a year. They also carry genuine fiduciary anxiety — an orphaned client on a dead plan is a client whose doctor stops being covered in January.
  • Rough TAM reasoning: Annual payments to MA brokers may now exceed $10 billion. At an average blended commission in the low hundreds per member-year, that implies millions of commissioned member-years intermediated by a large but finite population of licensed agents. I don’t have a defensible count of active independent agents with 200+ books, and I won’t invent one — this is a stated assumption to validate. Directionally, a market that pays out $10B in commissions supports tens of thousands of such agents; capturing 2,000 of them at $1,200/yr is a $2.4M ARR business.
  • Why now for them: 2026 was the worst disruption year on record (10% plan exits, 3M displaced). AEP 2027 preparation starts the moment ANOCs land on 30 September 2026. The pain is calendared.

5. Product sketch (MVP)

  • Book import — upload a client roster (CSV export from Sunfire, Connecture, or any CRM) with current plan ID, ZIP, and where available, medication list and preferred providers.
  • Book-wide ANOC diff — for every client, compare their current plan’s 2026 benefits against the 2027 plan year: premium, MOOP, drug tiers, network changes, and whether the plan still exists in their county at all.
  • Ranked call list — the core output. Every client scored by disruption severity, sorted, with a one-line human-readable reason: “Plan withdrawn from Lee County”, “Eliquis moved tier 2 → tier 4, est. +$1,840/yr”, “Primary care physician no longer in network.”
  • Non-renewal and orphan sweep — flags every client whose plan is terminating, and every client whose plan went non-commissionable, so the agent sees income risk and client risk side by side.
  • Talk-track per client — a short generated summary the agent reads on the call: what changed, what it costs this person, what the two best alternatives look like.
  • Outreach sequencing — export the ranked list to the agent’s CRM, or generate compliant email/SMS/letter drafts in priority order.
  • Season dashboard — how many high-risk clients contacted vs. remaining, so an agency principal can see coverage across a team of agents.

Deliberately out of scope for v1: enrollment submission, Scope of Appointment capture, call recording, and quoting. Those are regulated, well-served by Sunfire/Connecture, and integrating them turns a 10-week build into a 10-month one. BookTriage ends at “here’s who to call and why,” then hands off.

6. AI angle — what’s load-bearing

Remove the AI and roughly half this product survives — which is the honest answer, and it’s why the AI has to be pointed at the half that genuinely needs it.

Structured diffing is not AI, and shouldn’t pretend to be. Comparing formulary tiers across CMS files, detecting a terminated contract, checking a provider NPI against a directory file — that’s data engineering. Doing it well is the product’s spine. Claiming an LLM does it would be a lie and would make it less reliable.

Where AI is genuinely load-bearing:

  1. Messy client-roster ingestion. Every agent’s book is a differently-shaped CSV with free-text medication columns — “eliquis 5mg BID”, “Eliquis”, “apixaban”, “blood thinner”. Mapping that mess to RxCUI/NDC codes reliably across thousands of rows is exactly what LLMs are now good at and what previously required either a clean data-entry process (agents don’t have one) or a paid drug-terminology vendor.
  2. Provider name resolution. “Dr. Chen at Baptist” → an NPI in a directory file. Fuzzy, contextual, high-volume.
  3. Plan-document reading where structured data runs out. Supplemental benefits (dental, OTC allowance, transportation) are the differentiators seniors actually care about and are described in prose across ANOCs and Evidence of Coverage PDFs. Extracting comparable structure from those is a language problem.
  4. Per-client explanation. Turning a diff into a sentence an agent can say out loud to an 78-year-old, with the dollar impact computed, at 600× scale.

Without AI, this is a rigid tool that works only for agents with pristine data — a rounding error of the market. With it, it works on the book the agent actually has.

7. Localization angle (if any)

N/A — this is a US-only play. The product exists because of CMS regulatory structure: the ANOC mandate, the 15 Oct–7 Dec AEP window, CMS-published formulary files, and CMS-set commission caps. None of that is portable. There is a structural cousin in any country with privatized, annually-renewing, broker-intermediated health plans (the Netherlands’ annual switching window, Switzerland’s basic insurance) — but those are different products, not localizations, and neither has the US’s combination of broker commissions and public plan data.

The only “localization” that matters here is state-level: commission rates are higher in CT, PA, NJ, CA, and DC (California initial rose to $864, renewal to $432), and plan availability is county-level. The product must be county-aware from day one. That’s a data requirement, not a market strategy.

8. Business model — path to $1M–$5M ARR

  • Pricing: Tiered by book size, billed annually because the value is seasonal and monthly billing invites Nov–Jan churn.
    • Solo (up to 500 clients): $79/mo billed annually — $948/yr
    • Agency (up to 2,000 clients): $199/mo billed annually — $2,388/yr
    • Multi-agent (2,000+, per-seat dashboard): $399+/mo — $4,788/yr
  • Benchmark sanity check: MedicarePRO CRM runs $45–55/user/mo; Medicare CRMs generally $50–135/mo; Connecture direct signup is ~$100/mo. So $79–199/mo sits inside the existing category norm — agents already pay this range for tooling. Critically, the ROI story is single-digit: at $347/member renewal, retaining three clients pays for the Solo tier for three years.
  • ACV: blended ~$1,200.
  • Rough math to $1M ARR: 840 customers × ~$1,200 = $1.008M.
  • Rough math to $5M ARR: 3,200 customers at a higher blend ($1,560) as agencies skew the mix — realistically requires winning small FMOs as channel resellers rather than only direct agents, plus year-round product (see expansion).
  • Expansion path: (a) seats, as solo agents hire during AEP; (b) year-round SEP monitoring — Special Enrollment Periods and mid-year formulary changes create off-season triggers, which converts a seasonal tool into a 12-month one and is the single most important product move after v1; (c) FMO/upline licensing — an FMO buying for its downline of 300 agents is a five-figure contract and a distribution multiplier; (d) adjacent lines (ACA under-65 books have an analogous open-enrollment problem).
  • Margin note: CMS data files are free. Inference is the main variable cost — a 600-client book run is a bounded batch job, not a chat product, so per-customer COGS is low single-digit dollars per season. Gross margin should sit comfortably above 85%.

The seasonality risk is real and I won’t paper over it. Sept–Dec is when this product is indispensable and Feb–Aug is when it isn’t. Annual billing and year-round SEP monitoring are the two mitigations, and both need to be in place by the second season.

9. Go-to-market wedge — first 100 customers

The calendar does the selling. Everything below is timed to land in the six weeks before ANOCs arrive.

  • The “free book audit” wedge (primary). Offer a free one-time triage of the agent’s book: they upload a roster, get back the ranked list of their 25 highest-risk clients, and see the rest behind the paywall. This is the entire sales pitch, it demonstrates value in ten minutes on their own data, and it’s a natural fit for a market that is deeply skeptical of vendor claims. Target: 500 free audits → 100 paid in the first season. A 20% conversion is aggressive; 10% still yields 50 customers, and the audit itself generates the case-study numbers (“we found 41 at-risk clients in a 600-person book”).
  • Agent forums and Facebook groups, where this audience actually lives. insurance-forums.com has active Medicare agent threads (agents there discuss CRM tooling, uplines, and AEP survival in detail). There are large private Facebook groups for Medicare agents, plus r/InsuranceAgent. These are communities where a genuinely useful free tool posted in August — not a pitch — travels. Rule: show the audit output, not the landing page.
  • FMO and upline partnerships. FMOs exist to make their downline agents productive and they already distribute tooling. A single FMO with 200–500 downline agents is a distribution channel, not a customer. Approach 20 mid-size FMOs with a revenue-share or white-label offer before September. This is the highest-leverage channel and also the slowest — start it early, don’t depend on it for the first 100.
  • Ritter, PSM Brokerage, Senior Market Sales and the AEP-content ecosystem. These organizations publish AEP readiness guides and readiness timelines that agents actively read in August–September. Guest content, tool inclusion, or sponsorship puts the product in front of the exact reader at the exact week they’re planning their season.
  • Direct outreach to agency principals. Small agency principals (5–20 agents) are findable via state DOI licensee lists and agency directories. The pitch to a principal is different and stronger: “your agents are each guessing which clients to call; here’s the same ranked list for all of them, with a coverage dashboard.”

10. Build complexity — justification

Medium. The AI layer is off-the-shelf — entity resolution on drugs and providers, document extraction from ANOC/EOC PDFs, and per-client summary generation are all standard API work. Standard web stack, no novel infra.

The real work is data engineering against CMS file formats: ingesting monthly formulary/pharmacy files, plan benefit data, county-level plan availability, and (from 1 Oct 2026) MA provider directory files, then keeping year-over-year plan identity stable so a diff is meaningful. Plan contract/PBP identifiers change shape across years and that mapping is fiddly, unglamorous, and exactly where a competitor would give up. Add compliant-messaging care on any generated outreach, since CMS marketing rules govern agent communications.

Realistic estimate: 10–14 weeks to a credible v1 for a pair — one strong backend/data engineer and one full-stack. A solo builder can hit the free-audit MVP (import → diff → ranked list) in 7–9 weeks by deferring outreach drafting and the agency dashboard. That is tight against a 30 September ANOC date but not impossible if started now; the honest fallback is to launch the free audit in September and monetize the full product in AEP 2028, using season one purely for data, testimonials, and a waiting list.

11. Gating checklist

GatePass?Note
Legal in target market✅Analytics/decision-support tool for licensed agents. Uses public CMS data. Does not enroll, quote-to-bind, or market to beneficiaries directly — the agent remains the regulated actor. Generated outreach must respect CMS marketing rules; product ships drafts for agent review, never auto-sends. Client roster data is PII/PHI-adjacent — requires proper handling, BAA posture, and encryption from day one.
Ethical — no harm / dark patterns✅Aligns agent incentive with client interest: it surfaces clients being harmed by plan changes. One real hazard — ranking must be driven by client disruption, not agent commission, or the product becomes a churn engine that moves seniors for the agent’s benefit. Commission impact is shown as a separate, clearly-labelled field and must never enter the client-risk score.
Market exists (evidence above)✅10% plan exit rate, 3M displaced seniors, $10B+ in broker commissions, agents openly describing selective-outreach guesswork.
1–5 person team can build this✅Pair in 10–14 weeks; solo MVP in 7–9.
Launchable with <$50K / ₹40L✅Public data, off-the-shelf APIs. Main costs are two people’s time and modest inference/hosting.

All five pass.

12. Feasibility score

AxisWeightScoreNotes
Problem intensity2017/20Directly tied to recurring income at $347/member/yr, felt acutely in a fixed 54-day window, with a documented current workaround that is explicitly a guess. Not quite 18+ because it’s seasonal — intensity is enormous for four months and mild for eight.
Demand evidence1513/15Multiple independent hard signals: record 10% plan exits, 3M displaced seniors, $10B commission pool, CMS commission tables, agents on forums describing the exact unsolved triage problem, and an existing paid tooling category at $50–135/mo proving willingness to pay. Short of 14–15 because I found no one asking for this specific product by name.
Build feasibility1511/15Pair in 10–14 weeks on off-the-shelf AI. Docked for genuinely fiddly CMS data engineering and year-over-year plan-identity mapping, plus PHI-adjacent handling requirements. Not a weekend project.
Distribution clarity1513/15Named channels with a self-demonstrating wedge (free book audit on the customer’s own data), named communities, named FMO/content partners, and a calendar that creates urgency without paid ads. Docked because conversion on the free audit is unproven.
Revenue mechanics1512/15Pricing sits inside an established category range, ROI is single-digit clients retained, 840 customers to $1M, >85% margin. Docked for seasonality risk and because annual-billing acceptance in this market is an assumption.
Time to first revenue106/10This is the weak axis and the score reflects it honestly. The buying window is real but narrow: if you miss ANOC season, the next genuine urgency is twelve months out. Revenue is plausible within 8–12 weeks only if the September launch lands. Miss it and you’re looking at a long, cheap winter.
Defensibility105/10Execution moat plus accumulating book data and CRM-adjacent workflow lock-in. Realistically copyable — the CMS files are public to everyone. The durable edge is the year-over-year plan-mapping layer and being the tool agents already have loaded when the next season starts.
Total10077/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required

The data engineering is the product, so this needs a real backend engineer. And Medicare distribution is a relationship business with its own vocabulary — an outsider will be spotted instantly in agent forums. Strongly recommend a licensed-agent co-founder or advisor; without one, the FMO channel is close to inaccessible and the product will make credibility-destroying errors about how agents actually work.

Key assumptions to validate (3–5)

  1. Assumption: Agents will hand over a client roster (PII/PHI-adjacent) to an unknown vendor. How to test: Run 20 free audits in August. Measure how many agents actually upload versus stall at the data-sharing step. If they won’t upload, the entire wedge collapses — pivot to a local/desktop processing model or a de-identified upload format.
  2. Assumption: The ranked list materially beats the agent’s own intuition. How to test: For 10 agents, have them list their top-25 riskiest clients from memory before seeing output, then compare. If overlap exceeds ~70%, the product is a nice-to-have, not a must-have.
  3. Assumption: Willingness to pay ~$948–2,388/yr, billed annually. How to test: Pre-sell at a founding-member discount in August–September, before the product is complete. Ten prepayments is a real signal; ten “sounds great, ping me in October” is not.
  4. Assumption: Client medication/provider data exists in agents’ books at usable coverage. How to test: Inspect the first 20 real rosters. If fewer than half of clients have drug lists, the drug-tier diff — the highest-value signal — degrades badly, and the product leans on plan-termination and network data instead.
  5. Assumption: CMS provider directory files (from 1 Oct 2026) arrive usable and reasonably complete. How to test: Pull the first published files in October and assess coverage before promising network-change detection.

Risk flags

  1. Seasonality (highest structural risk): Value is concentrated in Sept–Dec. Annual billing plus year-round SEP monitoring are mandatory mitigations, not nice-to-haves. A product that’s only useful four months a year has ugly churn at renewal.
  2. Narrow launch window / timing: Missing 30 September doesn’t cost a quarter, it costs a year. This is the difference between the 6/10 on time-to-revenue and a 9/10.
  3. Platform and data dependency: Entirely dependent on CMS continuing to publish plan, formulary, and directory data in machine-readable form, and on file formats staying stable year to year. A format change lands mid-season, at the worst possible moment.
  4. Incumbent encroachment: MedicareCopilot, Sunfire, or Connecture could add book-wide triage — they already hold the client data and the enrollment relationship. This is the most likely way the idea dies. The counter is speed, plus being CRM-agnostic (working across whichever platform the agent uses) rather than locked to one ecosystem.
  5. Regulatory/compliance: CMS marketing rules govern agent-to-beneficiary communication, and rules for third-party marketing organizations are actively enforced. Generated outreach must stay draft-only and agent-approved. Additionally, handling client health data demands proper security posture from day one — a breach here is existential.
  6. Ethical drift: A ranking that quietly optimizes for commission rather than client harm would be both a compliance hazard and a betrayal of the product’s premise. Keep the two scores structurally separate and auditable.

14. Structured verdict

Score:                  77/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder (data engineering) paired with a licensed
                        Medicare agent or ex-FMO operator for domain and distribution
Time to revenue:        8–12 weeks IF the September ANOC window is hit; otherwise ~12 months
Capital to launch:      $10–20K (two people's time, public data, off-the-shelf APIs)
Top 3 assumptions to validate first:
  1. Agents will upload a client roster to a new vendor — run 20 free audits in August,
     measure upload completion rate, not interest
  2. Ranked output beats agent intuition — blind top-25 comparison with 10 agents,
     kill if overlap >70%
  3. Annual prepay at $948–2,388 — pre-sell 10 founding members before the product ships
Kill criteria:
  - Abandon if fewer than 10 of 50 approached agents will actually upload a roster
    (the wedge requires their data; without it there is no product)
  - Abandon if <5% of free audits convert to paid in season one
  - Abandon if Sunfire, Connecture, or MedicareCopilot ships book-wide ANOC triage
    before your v1 — they own the data and the enrollment relationship, and a
    standalone tool cannot out-integrate them from behind
  - Abandon if agents' rosters routinely lack medication data for >50% of clients
    AND plan-termination-only signals fail to impress in pilot

15. Next step — 1-week validation sprint

The whole sprint runs on one real book of business, because everything here is falsifiable only against real data.

  • Day 1–2 — Get one roster. Recruit a single friendly agent with 300+ clients (via insurance-forums, a Medicare agent Facebook group, or a warm intro to an FMO). Offer a free manual audit in exchange. If nobody will share a roster in two days, that is itself the answer — the data-access assumption is the make-or-break, and it should be tested first, not last.
  • Day 3–4 — Do the triage by hand. No product. Manually diff that book against 2026→2027 plan data: terminations, county exits, drug tier moves, network changes. Produce the ranked call list as a spreadsheet. Time yourself — the hours this takes are the value you’re selling.
  • Day 5 — The blind test and the ask. Before showing results, have the agent write down their own top-25 riskiest clients. Compare against yours. Then ask for $948 upfront for the coming season.

Falsifiable go/no-go: Proceed only if (a) the manual audit surfaces at least 15 genuinely at-risk clients the agent had not identified, meaning overlap with their intuition list is under 70%, and (b) the agent either prepays or gives a specific, dated commitment to prepay. Interest, enthusiasm, and “this is really cool” are not results. A signed check or a surprised agent staring at fifteen names they missed — those are results.

Interested in a detailed proposal?

Get a deep-dive with market research, competitive analysis, and implementation roadmap.

Contact us

info@startupbasket.ai