Our read
Start with a closed set of agreements your team has already reviewed by hand, write your real diligence questions as extraction prompts, and check each output through its source link. Hebbia's home page shows a financial example with color flags and a deal-conviction rating; the law-firm pages are where source links and review appear.
What you load into Matrix
Hebbia describes Matrix as a tool for reading large document sets such as filings, data rooms and contracts. Its home-page example is a financial screen of one company: 23 listed source files, among them SEC filings, earnings call transcripts, investor presentations and market data, feed a table of metrics.
The legal examples look more like a diligence room. At Ice Miller, a second-year associate uploaded a 30-gigabyte compressed file of deal documents. Hebbia kept the folder structure and built a Matrix spanning every agreement, which he then customized for his team's workflow. At Troutman Pepper Locke, the input was more than 1,000 legacy natural gas contracts spanning decades, with terms customized by locality, all needing review before a migration.
Reading the grid and writing the questions
On the home page, the example lists attributes such as LTM revenue, EBITDA margin, market cap and net leverage, with values and a basis column naming the period or data source, for instance LTM or Capital IQ estimates. It shows one company, so it does not tell you how a legal matrix is laid out across many contracts.
Orrick lists corporate due diligence, M&A deal points analysis, lease abstraction and governance document synthesis as uses. It describes Matrix breaking complex workflows into step-by-step AI agent actions and organizing results in a structured grid that colleagues can review and verify.
Troutman Pepper Locke's team shows the question-writing step. Its eMerge group refined the prompts on a sample set of contracts before scaling to the full pile. The output was the exact provision language, a plain-language summary of key terms and standardized coding, with data points hyperlinked to the source text.
Source links and the conviction rating
The home-page example uses green, red and yellow flags next to metrics, such as a green flag on net leverage and a red one on the maturity profile. Its only confidence-style indicator is a deal-conviction rating running from Low to High, set at Monitor Only in the example; the page ties it to transaction likelihood and actionability, so read it as a deal signal, not a score for how reliable an answer is.
Source links are described on the law-firm pages. Ice Miller's associate says the platform gives citation-backed answers and he can jump straight to the citation to check each point. Troutman Pepper Locke describes extracted data hyperlinked to source text so reviewers can validate with a click.
Where the lawyers sit in the loop
At Ice Miller, an M&A partner runs incoming redlines through Hebbia to separate material changes from cosmetic ones, and the firm says associates spend less time on busywork and get more reps on legal work. Troutman Pepper Locke says freed capacity moved from extracting information toward synthesis and risk analysis.
Orrick's page describes a grid designed for lawyers to review and verify, but it does not describe citations or a review procedure, so ask for both in a demo. Troutman Pepper Locke also says attorneys reuse prior prompts and past extractions from earlier engagements with the same client.
Editorial recommendation for a pilot: run one closed sample by hand and through Matrix, ask for quote, summary and label as separate outputs so each can be checked alone, compare cell by cell through the source links, and load the full room only afterward.
Data handling, as Hebbia states it
Hebbia's security page states that user documents, prompts and outputs are never used to train its models or those of any model provider it works with; Orrick's page also says Hebbia does not train on user data.
The same security page says Hebbia operates across the US and EU, with regional processing for storage and inference and a dedicated tenant if needed.
The data processing agreement says that when the agreement ends, Hebbia must destroy or return all processor data, at the customer's election, without undue delay. If law requires retention, Hebbia isolates the data from further processing until deletion is lawful. Hebbia gives at least 10 days' notice before engaging a new subprocessor, and the customer may object in writing on reasonable data-protection grounds; without an objection in that window, the subprocessor is deemed authorized.
The security page lists ISO 42001, SOC 2 Type 2 and ISO 27001 for Hebbia's own program. Your privacy team decides whether that fits a given matter.
Limits of the public examples
The product example is financial, with SEC filings and call transcripts, and the legal workflows come from customer pages. The home page does not show a legal matrix, so confirm the layout on your own files before assuming one row per contract.
Outcomes are vendor-reported and measure different things: Orrick reports diligence in minutes instead of days, Ice Miller a 30-hour task cut by half, and Troutman Pepper Locke a project estimated at three to four times longer without Hebbia. They cannot be ranked against each other and say nothing about accuracy.
Method and limits
2026-10-19
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.
