Layer 01 — Model
The reasoning engine
The underlying LLM: Claude, GPT, Gemini, or a vendor's fine-tuned financial model. Replaceable as benchmarks evolve. Often the layer with the least durable advantage.
The platform layer for AI in finance shifted materially between January and May. This memo evaluates whether a mid-market PE firm should buy a vertical platform, build on Anthropic's stack, or compose the two — and recommends an architecture, with honest cost ranges and a tagged decomposition of every workflow in the source document.
The single most important fact for this discussion: on 5 May 2026, Anthropic launched a financial services agent suite that maps almost one-to-one onto the workflows in the source document. It is a signal that the platform layer is consolidating around exactly the work the document describes — and that the build-or-buy question is no longer "wait and see."
Four developments since January reshape the decision:
The plugins Anthropic released are organized into two categories — “research and client coverage” and “finance and operations” — and ship as folders of plain markdown in a public GitHub repository. They are reference architectures, not finished products. Any firm can fork them and customize without writing code. This format matters for the build-or-buy decision in ways the rest of this memo unpacks.
Treating "Rogo vs. Claude" as a single decision is what creates the lock-in feeling. It is in fact three layered decisions, and bundling them is precisely what vertical platforms offer and what platform vendors unbundle. Each layer can be evaluated — and, in principle, swapped — independently of the others.
The underlying LLM: Claude, GPT, Gemini, or a vendor's fine-tuned financial model. Replaceable as benchmarks evolve. Often the layer with the least durable advantage.
What defines “when a CIM arrives, do X, then Y, then route to Z.” This is where firm IP — templates, voice, criteria, prior judgments — lives and compounds.
What an analyst or partner actually uses. Web application, Office plug-in, chat surface, dashboard. Where adoption is won or lost in week one.
Vertical platforms bundle all three. That is their value proposition (fast time-to-value) and their structural risk (the bundle becomes harder to leave as it is configured). Anthropic's stack unbundles them: Claude is the model, plugins are the workflow, Cowork is the interface. Any layer can be swapped without rebuilding the others.
For a mid-market firm, the unbundling matters more than it would at a thousand-banker investment bank. Why is the subject of the next section.
A 30-person firm does not see 5,000 CIMs a year. It sees, generously, 800. The industrialization of analyst work — the use case most vertical vendors lead with — is a real but secondary prize at this scale. The primary prize is something different:
Giving every partner the leverage of a great senior associate — on demand, in the partner's own voice, against the firm's prior thinking, all the time.
That kind of leverage requires the AI to know things that are not in any vendor's training set, and will not be:
None of that lives in a vendor's model. It lives in prior memos, CRM, board decks, email, and partners' heads. The real lock-in question becomes: where does the firm's institutional memory get fed — into a vendor's system, or into a stack the firm owns?
The recommendation is a hybrid, weighted heavily toward Anthropic ownership rather than the conventional 50/50 split. Cost is not the argument — as §VI shows, the hybrid is the most expensive of the three options. The argument is durability: the firm's institutional memory and voice live in the firm's files, not a vendor's database.
Anthropic provides foundation, workflow engine, and interface. One vertical tool is purchased only where vertical depth genuinely wins — dataroom and contract interrogation with cell-level citations. Firm memory is exposed via private MCP and never leaves the firm's environment. The architecture is composable, not bundled, which preserves optionality as the platform layer continues to evolve.
ic-memo, source, screen-deal, value-creation. Add firm-specific plugins for LP DDQ, post-mortem, and thesis update.Ranges below are directional — reflective of mid-market PE pricing observed in the field, not vendor list prices or enterprise quotes. Implementation effort is excluded.
| Approach | Annual Recurring | Lock-In | Time to Value | Primary Trade |
|---|---|---|---|---|
| Rogo only | $20–30K | High | Weeks | Firm IP lives in vendor templates. Difficult to compose with other systems. |
| Claude + Cowork only | $20–40K | Low | 2–3 months | Less polish on commodity workflows (e.g., document-heavy dataroom interrogation). |
| Hybrid — Claude-weighted | $40–70K | Medium | 4–6 weeks | Most expensive of the three. Requires one capable internal owner. Buys durability and composability. |
All three options sit comfortably below 1% of management fee for a fund of this scale. The decision is not driven by cost; it is driven by where the firm wants institutional memory to live and how much optionality the firm wants to preserve.
The hybrid is the most expensive option of the three — not the cheapest. The argument for it is durability rather than savings. Anthropic-only is the cost-minimizer; Rogo-only is the speed-maximizer; the hybrid trades a moderate cost premium for ownership of the firm's most valuable layer (its memory and voice) and the ability to swap vendors at any layer without rebuilding the others.
The risk worth pressure-testing in §VIII is whether the firm has the capacity to absorb the implementation work the hybrid requires. If the answer is no, the recommendation collapses toward Rogo.
Sector maps and white-space identification are where the firm's proprietary view lives. A vendor will produce a generic sector map; the firm's edge is its specific view of where the gaps are. This belongs in a firm plugin that ingests sell-side research, BamSEC filings, and the firm data lake, then writes in the firm's voice.
The extraction work is commodity (Hebbia or Rogo will read a CIM well). The scoring against thesis fit and routing is firm-specific. Compose: a vendor MCP tool extracts the metrics; a firm plugin scores fit and writes the one-pager in the firm's voice.
Pulling public comps and trading multiples is a pure data-pipeline problem. CapIQ plus PitchBook MCP, or a vendor's pre-built feature. Not a place to reinvent.
Two pages on business model, market, financials, and key questions. This is where partner voice and firm pattern-matching across prior deals matter most. A firm plugin fed by prior memos in the data lake.
Industry research draws mostly from external sources (IBISWorld, Third Bridge, expert transcripts). A generic LLM produces a competent primer; the "so what for the firm" overlay is firm-specific. Compose.
Asking the right questions is the firm's edge. A plugin that knows the firm's thesis gaps, prior expert calls in the space, and what each partner cares about will produce better question lists than any vendor.
Transcription and topic-extraction are commodity. Surfacing quotes for or against the thesis is firm-specific. Compose: Tegus handles transcripts; firm plugin handles thesis-mapping and updates the knowledge base.
Quick valuation ranges from comps and precedents are mechanical. Anthropic's financial-analysis:comps and dcf plugins handle this directly.
The Hebbia sweet spot — cell-level citations across thousands of dataroom documents. Buy it. Expose to Claude as an MCP tool.
Reconciling management financials to GL is computationally heavy and benefits from a tool that traces every cell. Rogo's Offset acquisition targets exactly this. Compose: vendor handles the spreading; firm plugin frames the normalization narrative.
The math is standard; the judgment about which signals matter for this sector is the firm's view. A unit-economics plugin, using Anthropic's private-equity:unit-economics as a starting point.
Contract-extraction with traceable citations is exactly where vertical tools earn their keep. Hebbia, Spellbook, or similar.
Mapping the stack uses standard scanners; the AI/automation upside scoring is firm-specific (and tied to the firm's portfolio AI playbook). Compose: external scan plus firm plugin for upside assessment.
Highest stakes, lowest volume, hardest to recover from a hallucination. Tax counsel still owns this. AI's role is drafting questions for counsel and preparing briefing packs. Not the first workflow to automate.
Broker continues to own this. Useful for surfacing claims-history patterns once a corpus exists, but not v0.1.
The synthesis layer across every workstream. The firm's own framework for what counts as a red flag, what severity means, who owns mitigation. Highest-leverage build target after the IC memo itself.
Anthropic ships an lbo-model plugin already. Fork it to the firm's templates and capital-structure conventions. Compose with Daloopa or a similar vendor for historicals.
The highest-leverage workflow in the firm. Fork private-equity:ic-memo and rebuild around the firm's section structure, voice, mandatory disclosures, and partner-by-partner objection patterns from prior IC meetings.
Anticipating which partner asks what, in what order. Only works against the firm's IC archive. A pure build — and arguably more valuable than the IC memo itself for partner adoption.
Counsel continues to own these. Useful for redline-vs-precedent diffing eventually, but not v0.1. The blast radius of a missed change is too high.
The firm's playbook for value creation is the firm's edge. Encode it. Anthropic ships private-equity:value-creation as a head start.
Pre-read summarization is commodity; surfacing off-plan KPIs against the firm's specific value-creation framework is firm-specific. Compose.
PitchBook and CapIQ have the M&A history, capacity scoring, and ownership data. Anthropic's investment-banking:buyer-list covers this directly.
Firm voice and firm history. A plugin that ingests prior DDQ responses, current annual meeting decks, and quarterly reports, and writes in the IR team's voice. The highest-ROI IR build target.
Different LPs care about different KPIs (DPI, exits, ESG, sector mix). A plugin that knows each LP's hot-button items and prior questions is uniquely firm-specific.
CRM is the system of record (Affinity, DealCloud, Salesforce). Expose via MCP; firm plugin handles prioritization, follow-up scheduling, conference logic.
Generic vendors produce generic outreach. Founder-specific messaging that references the firm's recent thinking, prior interactions, and shared connections is a firm build — and is exactly what closes proprietary deals.
Vendor handles deal-history and recent-mandate scraping; firm plugin layers in CRM history and firm-specific assessment. Compose.
A searchable knowledge base across every CIM, IC memo, banker interaction, and expert call — in the firm's environment, under firm governance. This compounds. Arguably the highest long-term-value build target in the document.
Anthropic ships operations:kyc-doc-parse and kyc-rules directly. Use them; integrate with ACA workflow.
Numerical backup pulls from firm data; policy drafting uses Anthropic's compliance plugin. Compose.
Cowork ships with native Outlook and Gmail integration. Out of the box for everyone on day one.
Personalized to portfolio, watchlist, and current pipeline. Runs as a Managed Agent on schedule. Pure firm build — the personalization is the whole product.
The pattern: roughly two-thirds of the workflows are either Build or Compose — firm-specific work, or work where the firm's view is the differentiator. That ratio is the strongest single argument against buying a one-stop vertical platform. The work the firm uniquely cares about is the majority of the document, and none of it sits cleanly inside a vendor's product.
The single most important operational question under the recommendation in §V is whether the firm has one capable internal owner for the workflow layer. Not necessarily a coder — someone who can sit with a partner, observe how an IC memo actually gets written, and translate the practice into a plugin (with Claude doing the writing).
Call the role "BizOps for AI." A capable Chief of Staff or Head of Data with the right disposition can do this. So can a thoughtful senior associate who likes systems work and is willing to spend a slice of every week on it. The role does not need to be full-time at a firm of this size, but the responsibility must sit somewhere accountable.
If the firm cannot or will not establish that ownership, the vendor path is more honest about the tradeoff: pay more on a fully-loaded basis, accept less composability, but do not depend on internal capability that does not exist.
Every other decision in this memo flows from this one.
The thesis hinges on the claim that roughly two-thirds of the document is firm-specific. If a partner reads the taxonomy and disagrees on (for example) IC memo or value-creation plan as a build target, the recommended architecture changes meaningfully.
This is the operational fulcrum. If the answer is no, the recommendation collapses toward Path A (vendor).
The IC memo is the natural candidate — highest leverage, most voice-sensitive, most visible to partners. Building it well validates the entire approach; building it badly is recoverable. Arguments exist for the morning briefing (lowest stakes, fastest visible value) or the DDQ plugin (highest IR-team ROI). The choice should be deliberate.
The architecture assumes prior memos and CRM data can be exposed to Claude via MCP. If the firm requires data to remain in a specific region or environment, the model-layer choice tightens and may push toward self-hosted Claude or a vendor with stricter deployment options.
The source document is comprehensive but treats AI workflows as discrete. The compounding value comes from workflows feeding each other — expert call synthesis feeding the IC memo feeding the post-mortem feeding the next thesis. Worth mapping the connections, not just the rows.