AI agent document reading can help organise a due diligence room, locate clauses and build a source-linked issue list. It cannot prove that every page has been interpreted correctly, nor can it replace qualified legal, financial or technical judgement. The useful workflow gives the agent bounded access and a narrow question, requires citations to the actual documents, and sends every material conclusion to an accountable reviewer.
Define the job before opening the files
“Read the data room” is too broad. A more useful instruction is to identify all contracts with a change-of-control clause, list the clause text and page reference, and flag documents where the relevant section was not found. Another task might compare a disclosure schedule with the underlying agreements. These are retrieval and organisation tasks. They create a map for human review without pretending that the agent has made a transaction decision.
Start with an inventory: document title, date, version, source, owner, confidentiality level and whether it is searchable text or a scan. If a file cannot be read, mark it as unreadable. If there are two versions, do not silently choose the newer-looking one; record both and ask which controls. The quality of an agent's output is constrained by the files it actually receives and the permissions under which it works.
Set a deadline and scope for the pass. A first pass could identify missing documents and unusual terms; a second pass could verify those findings against source pages. This staged approach is easier to audit than asking for a polished investment memo in one prompt. It also makes the system's limitations visible early, when a human can redirect the work.
Why are citations essential?
A plausible summary without a source is not a due diligence finding. Require each material statement to identify the document, version, section or page and, where possible, a short supporting passage. Then have a reviewer open the original page. The citation is a route to verification, not proof by itself. A model can cite the wrong page, misread a table or combine statements from different versions.
NIST's Generative AI Profile describes “confabulation”: confidently stated content that is erroneous or false. It notes that generated citations can also be misleading. This is particularly relevant to long-form document review, where a fluent paragraph can hide a missed exception. Ask for direct evidence and an uncertainty label, then verify the high-impact items yourself.
Keep a distinction between “found”, “not found” and “not reviewed”. A search that does not locate a clause does not prove it is absent, especially in scans, image-based annexes or unusual wording. If the agent has not opened an attachment, the report should say so. A due diligence team can then allocate a human reviewer to those gaps rather than reading the agent's silence as assurance.
What should the human reviewer check?
Review every item that could change price, obligations, closing conditions or risk allocation. Also sample-check lower-risk findings to estimate whether the workflow is reliable. For a contract, compare the summary with the exact defined terms and exceptions. For financial figures, trace the number to the original statement, period and units. For dates, check whether the document was amended later. The agent may accelerate finding the page, but it does not accept responsibility for the conclusion.
A useful issue log has columns for question, agent finding, source location, confidence or uncertainty, human verification, owner and resolution. Do not fill the “verified” column automatically. A reviewer should be able to reject a finding or request a second source. If the model repeatedly misses a type of table or appendix, change the workflow and assign that category to manual review.
Keep final legal and financial judgments with the people authorised to make them. An agent can draft questions for counsel or highlight a covenant, but it cannot determine the legal effect of the clause in all relevant jurisdictions. It can identify an unusual revenue pattern, but it cannot independently validate the accounting treatment. Treat its output as a reading aid, not a professional opinion.
How should confidential files be handled?
Before connecting an agent to a data room, confirm that the organisation permits the tool, the data location, retention terms, user access and any downstream integrations. Limit access to the documents needed for the task. A prompt should not include personal identifiers or trade secrets merely because they are available. The FTC's discussion of AI privacy commitments highlights the sensitivity of internal documents and user data supplied to AI services.
Use the approved business account and permissions, not a personal account copied into the workflow for convenience. Record whether the agent can only read, or can also export, share or alter files. Read-only access is an important boundary for an initial pilot. If the task requires sending findings to colleagues, review the recipients and redactions before transmission. The device on which an agent runs does not by itself settle the service's data-handling terms.
Where does Agent Q fit?
VERTU Agent Q presents Ruby Talk as access to AI agents, with VERTU describing a broader agent ecosystem. The Agent Q product page provides a current device example. That positioning makes document-reading workflows relevant, but it does not establish that a particular data room, document app or enterprise account is integrated. Confirm the exact software version, authorised services and permissions before putting confidential material into a workflow.
Begin with a non-sensitive sample set. Ask the agent to identify titles, dates and one clause type, then compare each result with the source. If accuracy is poor or citations are absent, do not scale to a live transaction. If the pilot works, document the approved scope, reviewer and escalation path. A phone can make reading more mobile; it does not remove the need for a controlled review process.
A realistic first-pass checklist
Count the files received and processed. List any scans, encrypted items or missing annexes. Request source-linked extracts for a small number of defined questions. Have a human verify the highest-impact answers and a random sample of the rest. Record corrections, then decide whether the next batch is worth processing. This checklist creates a measurable test of the workflow without promising a percentage accuracy that has not been demonstrated.
The output should end with open questions, not just a neat summary. Which documents are missing? Which findings rest on ambiguous language? Which figures differ across versions? Which items require counsel or finance review? A useful agent-assisted process makes those uncertainties easier to see. If it turns uncertainty into polished certainty, it has made the diligence task harder. Preserve the source files and the reviewed issue log so another authorised team member can reproduce the finding later.
Frequently asked questions
Can an AI agent complete due diligence alone?
No. It can assist retrieval and triage, while qualified people verify sources and make the final decisions.
Are page citations enough?
No. Open the cited page and check context, definitions, dates and exceptions.
Can I upload a confidential data room to Agent Q?
Only after your organisation approves the precise service, access and data-handling terms. Device ownership alone is not authorisation.
What should the first pilot measure?
Measure coverage, citation quality, missed documents and reviewer corrections on a small approved sample.
