Crunched

The agent is the whole product. Building, linking, error checking and slide making all run through models working inside the customer's own files, and no data asset or workflow system sits underneath that would survive their removal. The proprietary layer is a context search that finds the right cells and tabs in a workbook with dozens of sheets, plus a dedicated system for catching finance specific errors, and both exist to direct the inference.

Oversight is designed into how changes land. The agent flags the mistakes it finds directly in the sheet for a person to review, every number stays live linked from source to model to slide, and any change it makes can be undone and reapplied later. Customers use the error check as a review layer between analysts and senior bankers rather than as a replacement for that review. No approval step before edits, confidence threshold or list of actions the agent will not take is published.

Traceability is a design principle rather than an add on. Every number is live linked from its source through the model to the slide, assumptions come with written rationales and links to their sources, and a dedicated checking system targets finance specific errors such as net working capital logic, missing add backs and mislinked revenue assumptions. Those are behaviors a buyer can test on its own models. No accuracy rate, evaluation method or policy for monitoring quality as models change is published.

Results are reported in some detail but not tied to named customers. A case study published by its model provider reports time savings above 50 percent on modeling work, financial spreads falling from more than eight hours to one, company write ups from eight hours to twenty minutes, and a real estate investor's deal task dropping from ten hours to one, with the firms unnamed. The company's hiring materials name professionals at firms including Pareto, PwC, BCG and CBRE as users, with no outcomes attached. The names and the numbers never meet.

Whether customer files, models or prompts are used to train or improve any model is not stated, and no retention or isolation terms are published.

No privacy terms, retention schedule, subprocessor list or data processing commitments specific to the product are published. Confidential deal documents and client models pass through it and on to a named external model provider, so the gap is a real one for a bank or fund.

No certification, audit report or trust portal is named. The founders describe building with the security needed for clients to trust them with highly confidential data, which is a commitment rather than a credential.

No license or supervised status is held, and none would be expected for a modeling and review tool. Its finance customers are regulated, and models it builds or checks feed deal decisions and client materials inside their obligations. Nothing published addresses that boundary.

No governance framework, testing regime or bias evaluation is published. The judgment calls that matter here are which assumptions the agent proposes and how it rates them, since one marketed workflow has it label a business plan's assumptions as optimistic, realistic or pessimistic. How those labels are reached and checked is not described.

No liability position, accuracy warranty or remediation route is published. The product is positioned as the first reviewer of models headed to senior bankers and clients, so a missed error is the failure that matters, and nothing published says who bears it.

The model supplier is named down to the version, with a reason for each. A case study published by Anthropic, with the founders quoted, states that the product runs on Claude Opus 4.5 and Sonnet 4.5. Sonnet powers the context search that gathers the right information from large workbooks, using parallel tool calls and a long context window, while Opus makes the deeper and more costly deal analysis economical. The company says it evaluated several models before choosing, and describes keeping customers current as providers change underneath. No failover or customer choice of provider is described.

The agent works inside Excel and PowerPoint through add ins listed on the Microsoft marketplace, so models and decks stay in the formats teams already review. It pulls data from PDFs, images, web sources, third party financial data providers and a customer's own internal databases, and links it into existing templates rather than replacing them. Workflows are tailored to each firm's methodology by a team working alongside the customer. It is not a system of record for any function, and the environment it works in belongs to Microsoft.

Nothing published says where customer workbooks, documents and decks are processed or stored, or whether any regional or private deployment is available. That matters because the inputs include information memorandums and client models from live transactions.

No price is published. The Microsoft marketplace listings state that a paid license is required for full access, and the company sells through demos and firm specific tailoring, which points to quoted enterprise pricing. Nothing indicates whether charging is per seat, per firm or per workflow.

Buyers named in the company's launch, accelerator and model provider materials are investment banks, private equity firms and investment firms, alongside management consulting, audit and corporate finance teams. Its hiring materials name professionals at firms including Pareto, Nordover Capital, BCG, PwC, BDO and CBRE as users, and it reports customers in six European countries and the United States. The most recent positioning leads with advisory and audit firms, so financial institutions are one of two core audiences rather than the only one. No customer count is published.