Tracelight

Every core job runs through the agent: building and extending models, reviewing them for errors, answering questions about them and turning them into slides and documents. Strip out the models and what remains is a set of Office add ins and a project workspace with nothing to do. The company's own engineering sits in how spreadsheet structure is translated for the model and in the checks applied to every change, and both exist to make the inference usable.

Oversight is built into how changes land. Users can plan a task with the agent before any calculation is done, switch between an automatic mode and a control mode for reviewing what it proposes, and inspect every edit in a changeset that checks formula health, flags unintended hard coded values and preserves the surrounding structure and formatting.

AI actions are logged, a precedent tree shows how any cell is linked, and a workbook comparison shows every change between two versions with a summary of its impact. The vendor's benchmark also counts how often an agent edits cells it was not asked to touch, which is the failure a spreadsheet owner most needs to control.

The company publishes its own benchmark with the method laid out. Agents were scored on exact formula correctness, cell by cell, across 240 prompts on 80 real Excel models in three difficulty tiers, and a correct number typed in as a value scores zero. A second test planted 480 realistic errors in the same models and measured how many each system found and how many of its findings were real.

Tracelight scored 75 percent on modeling and found 68 percent of the planted errors, figures that also tell a buyer a quarter of cells and nearly a third of errors were missed. The test is run by the vendor rather than an independent party, but the method is specific enough to repeat, and changesets, cell level citations and precedent tracing give a team the means to check the output on its own models.

Named customers appear on the company's site, including Arma Partners, Puma Growth Partners, Vistra, RSM and Armanino, but no case study ties a named firm to a measured result. The figures on offer, such as decks built 80 to 90 percent faster and early users at banks and private equity firms reporting time savings above 90 percent on standard modeling tasks, are attached to use cases or unnamed users. The company's own benchmark is the strongest performance evidence available and is covered under model risk.

A contractual guarantee states that customer data is never used to train any model, its own or a third party's, and the model providers behind the service operate under zero data retention terms. Customers set their own retention, enterprise spreadsheet data is deleted immediately after processing, and the free audit tool keeps no spreadsheet data at all. Together these answer both questions a bank or fund will ask: what is kept, and what it is used for.

Customer data is defined broadly, covering uploads, connected files, prompts and outputs, and is stored only to run the service. Customers decide which data is processed, set retention, delete at any time and can export everything when an agreement ends, after which their data and any dedicated resources are permanently deleted. A data processing agreement is published and a subprocessor list is kept in the trust center. Nothing addresses the Gramm Leach Bliley Act or how a bank's client information should be handled inside a model.

A SOC 2 Type II audit by an independent party is complete, and a trust center run on Vanta holds the documentation. Regular third party penetration tests cover the whole platform, data is encrypted with TLS 1.2 or later in transit and 256 bit AES at rest, and single sign on runs over SAML 2.0 so firms can enforce multifactor authentication. ISO 27001 is described as an alignment with the standard rather than a certification, and buyers should read it that way.

No license or supervised status is held, and none would be expected for a modeling tool. Customers use it for valuations, diligence and investment committee materials that sit inside their own regulatory and professional obligations, and nothing published addresses that boundary.

No governance framework or bias evaluation is published. The judgment calls that matter here are how the agent proposes assumptions and which figure it uses when sources conflict. The benchmark measures formula accuracy and error finding, not those choices, and while firm specific style guides and skills shape the output, how they are governed is not described.

No liability position, accuracy warranty or remediation route for errors in agent built work is published. The company's own benchmark shows the agent misses some errors, and the product is used on live deals and client deliverables, so who bears a missed error is the open question.

The model providers are not named on the public site. Pricing tiers promise premium and latest model access, the founders describe the product as bringing frontier models into Excel, and the security page says model providers work under zero data retention terms, all of which confirm third party models without saying whose. A subprocessor list kept in the trust center may name them.

The agent runs inside Excel, PowerPoint and Word through add ins approved for Microsoft 365, which request only the minimum permissions each task needs, and on a web platform that keeps chats, files and outputs together by project. Paid plans include integration access through the Model Context Protocol, and comparable company work pulls in Capital IQ tickers. Decks and documents stay tied to the model, so numbers remain consistent when the model changes. It is not itself a system of record, and the core working environment belongs to Microsoft.

Enterprise customers choose where their data is stored and processed, with the EU, the United States and Australia offered and other regions available, and the company commits to keeping data in the chosen region. EU residency and processing also anchor its GDPR position. That is a residency control the customer holds, not a statement of where the vendor happens to host.

Prices are published for the individual plans. A free trial gives limited platform access, limited chat and three model review runs. Pro costs $40 a month with higher usage limits, a confidential mode and access to premium models and integrations. Max costs $200 a month for five times the usage of Pro plus priority support. Enterprise is quoted and adds unlimited usage, a firm wide skill library, single sign on, configurable retention and an admin hub for tracking and billing. The usage limits themselves are not stated, so a team cannot tell from the page how far a plan stretches.

Published use cases cover private equity underwriting and portfolio reporting, mergers and acquisitions, valuations, corporate development, and financial and commercial due diligence. Named customers span an investment bank advising technology companies (Arma Partners), a growth investor (Puma Growth Partners), a fund and corporate services provider (Vistra) and two accounting and advisory firms (RSM and Armanino), so financial institutions sit alongside advisory practices. Hosting regions in the EU, the United States and Australia point to an international base. No customer count is published.