Job

Can an AI deep research agent write your reports?

Researched November 2026

Can an agent do this job today?

Supervised

Deep research agents are in production use, according to vendors: OpenAI reports that Hebbia's multi-agent platform, used by finance and legal firms, can draft investment committee memos and runs deep research over their documents, and Focused reports building production deep research agents with Weber Shandwick for strategic research and media analysis. Both accounts are vendor-published. The output cannot be trusted unchecked: the DeepTRACE audit (ICLR 2026) found deep research configurations still include large fractions of statements unsupported by their own listed sources, with citation accuracy of 40–80% across the systems tested, and OpenAI's own launch post says its agent can hallucinate facts and struggle to tell authoritative information from rumors. A person must review each report before it informs a decision, so this is a supervised hire, not a ready one.

Abstract illustration of stacked rings like a well with small blank pages drifting down to a glowing green stack at the bottom

On this page

Can an AI agent do deep research reports?

Yes, as a drafter you check, not as an analyst you trust blind. An AI research agent takes a question, reads many sources and returns a long report with citations. OpenAI documents its deep research proposing a plan you can edit first, and OpenAI's deep research and Google's Gemini Deep Research can both read your own documents as well as the web. Vendors report production use: OpenAI describes Hebbia's agents, used by finance and legal firms, as able to draft investment committee memos, and Focused describes deep research agents it built with Weber Shandwick for strategic research and media analysis.

The catch is the citations. A peer-reviewed audit presented at ICLR 2026 found deep research systems still produce large fractions of statements their own sources do not support, with citation accuracy between 40% and 80% across the systems it tested. So hire it for the searching, reading and first draft, and keep a person on the checking and the conclusions.

This page covers general research reports. For narrower jobs see market research, competitive intelligence, literature review and financial research.

What the agent does and what stays with you

Agent with review
- Turn a question into a research plan
Agent
- Search the web and chosen sites for sources
- Pull in your own files and documents
Agent with review
- Draft a structured report with citations
Human
- Check that cited sources exist and support the claims
- Fill gaps from paywalled or subscription data
- Decide what the findings mean and recommend action

7 tasks: 2 agent, 2 agent with review, 3 human

Task	Who does it	Why
Turn a question into a research plan	Agent with review	OpenAI's help center says ChatGPT proposes a research plan you can review and modify before the research begins.
Search the web and chosen sites for sources	Agent	OpenAI documents searching the public web or only the sites you specify, and the source list is visible for checking.
Pull in your own files and documents	Agent	OpenAI documents uploaded files and connected apps as sources, and Google documents Gmail, Chat and Drive content in Gemini Deep Research.
Draft a structured report with citations	Agent with review	OpenAI says every report includes citations or source links so you can verify it; audits show those citations need checking.
Check that cited sources exist and support the claims	Human	Published audits find hallucinated URLs and unsupported statements in agent reports, so the check cannot be left to the agent that wrote them.
Fill gaps from paywalled or subscription data	Human	OpenAI says authenticated data services work only where they are connected and supported, so anything outside them needs a person with access.
Decide what the findings mean and recommend action	Human	OpenAI reports its agent often fails to convey uncertainty accurately, so judgment about confidence and next steps stays with the person who owns the decision.

OpenAI's help center describes ChatGPT's deep research workflow in five steps: you describe the outcome, choose sites or files, review a proposed plan, follow progress and interrupt if needed, then get a structured report "with citations or source links so you can verify the information." Google says Gemini Deep Research can draw on Gmail, Chat and Drive files in addition to the web, if your admin allows it.

Abstract illustration of a row of blank page icons with green threads running to a single glowing point, and a small amber glow where some threads cross

What stays with you is anything the agent cannot check about itself. OpenAI's launch post lists the limits plainly: the agent "can sometimes hallucinate facts in responses or make incorrect inferences," "may struggle with distinguishing authoritative information from rumors," and "currently shows weakness in confidence calibration, often failing to convey uncertainty accurately." Those are exactly the parts of a research job a reader relies on.

Independent work points the same way. Rao, Wong and Callison-Burch checked 53,090 citation URLs from 10 models and agents on one benchmark and found 3–13% were likely hallucinated (no record in the Wayback Machine) and 5–18% did not resolve. They also report that deep research agents produce more citations per query than search-augmented models but hallucinate URLs at higher rates. That paper is a preprint. Its practical lesson: a long source list is not evidence that the sources say what the report claims.

A worked example: one report

This is an illustration, not a customer story. Assume, for this example, that a 20-person software company is weighing a move into the Nordic market.

The head of growth writes the brief: market size and main competitors for HR software for small businesses in Sweden, Norway and Denmark, sources from the last 24 months, official statistics first. The agent proposes a plan with five sections. She removes one (pricing, which sales already knows) and approves it. Twenty-five minutes later there is a 14-page report with 60 citations and a one-page summary.

A reviewer spends an hour on it. She opens every source behind the summary's six numbers. Five match. One market-size figure links to a vendor blog post that quotes it without a source, so she marks it unverified and moves it to the open questions. She samples 20 body citations: 18 support their sentences, one link is dead and one cites a 2019 page outside the date range. The report goes back with three corrections. The head of growth writes the recommendation herself. The agent saved the reading; the people decided what was true and what to do.

What you need to provide

The agent touches whatever you let it read: public web pages, uploaded files and, if connected, internal documents, email and chat. That can include customer and financial data, so decide first what it may see.

A written question tied to a decision, with scope limits and a date range.
A source policy: sites to prioritize or restrict to (OpenAI documents both options), and sources to exclude.
Read-only access to the files the report needs. OpenAI says deep research uses read actions from connected apps and does not use write actions, and that connected apps do not override provider permissions or workspace restrictions.
Access to any subscription data the answer depends on. Where no connector exists, plan for a person to pull that data.
Your retention rules. OpenAI says deep research conversations follow the same data handling settings as regular ChatGPT conversations; check your workspace settings before uploading anything sensitive.

Job brief template

Use this as the brief for an operator or for yourself when you start a run. Replace the bracketed values.

Copy-ready brief

Goal: Each research question gets a cited report within two business days in which every number and named claim in the summary traces to a live source that says it.
Deliverable: A report in a shared document (Word, Google Doc or PDF) with a one-page summary, findings by section, inline citations, a source list with access dates and a list of open questions the agent could not answer.
Inputs: The question, the decision it informs and who will read the report
Scope limits: date range, geography, industries, competitors or products in and out of scope
Preferred and excluded sources (for example, official statistics first, no undated listicles)
Files or internal documents the agent may read, with read-only access
Logins or exports for any subscription data the report must use
An example of a past report in the format you want
Acceptance criteria: Every number in the summary links to a source that contains that number; zero exceptions on a full check of the summary.
On a sample of 20 citations from the body, all URLs resolve and at least 19 support the sentence they are attached to.
No source is older than the date range set in the brief unless the report says why it was kept.
Claims the agent could not verify are listed as open questions, not stated as findings.
The research plan was approved by the requester before the research ran.
Systems: deep research tool (for example ChatGPT deep research or Gemini Deep Research)
document store (for example Google Drive or SharePoint), read-only
subscription data services, where connected
Constraints: Uploaded files and connected apps follow the tool's data handling and your workspace retention settings; do not give it documents your policy keeps out of third-party tools.
The agent gets read access only; OpenAI says its deep research uses read actions from connected apps, not write actions.
Reports are internal drafts until a named reviewer signs them off.
Volume: Example: 2–10 reports a month
Cadence: Example: on request, with a two-business-day turnaround
Human checkpoints: Research plan: the requester approves scope and sources before the run.
Citation check: a reviewer opens every source behind the summary and a sample from the body.
Conclusions: the person who owns the decision writes or approves the recommendation.

Keep the zero-exception criterion on the summary. Readers act on the summary, so that is where an unsupported number does the most damage.

Risks and human checkpoints

Citations point to pages that do not exist or no longer resolve.
How you notice it
Click every link in the summary and a sample from the body; Rao and colleagues found 3–13% of agent citation URLs were likely invented.
A claim is cited, but the source does not say it.
How you notice it
Read the cited passage for each number in the summary; DeepTRACE found large fractions of unsupported statements in deep research output.
The report presents one side of a contested question with confidence.
How you notice it
Ask the brief to list the strongest opposing evidence and check that section is not empty; DeepTRACE found deep research answers highly one-sided on debate queries.
Stale or low-authority sources stand in for official data.
How you notice it
Check source dates against the brief's date range and look for rumor sites or undated pages; OpenAI says its agent may struggle to tell authoritative information from rumors.
The report skips paywalled or private data and fills the gap with whatever it could reach.
How you notice it
Compare the source list with the data sources you expected; a key database missing from it means that part of the report needs a person.

Human checkpoints

Research plan: the requester approves scope and sources before the run.
Citation check: a reviewer opens every source behind the summary and a sample from the body.
Conclusions: the person who owns the decision writes or approves the recommendation.

The failure that matters most is not a wrong answer that looks wrong. It is a fluent paragraph with a real-looking citation that does not support it. We recommend making citation checking a separate step with a named reviewer, not something the requester does while skimming. Budget the time: in the example above, an hour of review for a 14-page report.

How to run a trial

Test the agent on questions you already know the answer to before you rely on it for new ones.

Pick three past research questions your team answered well, with the final reports on file.
Write each as a brief with the same scope and date range, and run the agent.
For each report, check every summary number and 20 random body citations. Record dead links, unsupported claims and anything it missed that your team found.
Compare its findings with your team's report. Note what it added that your team missed, too.
Time the review. If checking takes nearly as long as the original research, the agent is not saving you work on this kind of question.
Run two new questions with the full checkpoints before you widen use.

Pick three past research questions
That your team answered well, with the final reports on file
Write each as a brief and run the agent
With the same scope and date range
Check every summary number and 20 random citations
Record dead links, unsupported claims and anything it missed that your team found
Compare its findings with your team's report
Note what it added that your team missed, too
Time the review
If checking takes nearly as long as the original research, the agent is not saving you work
Run two new questions
With the full checkpoints, before you widen use

For more on structuring the test, see how to run a paid trial with an AI agent. For the human roles involved, see research assistant and market research analyst, and for the wider function, AI agents for research.

Questions buyers ask

What does it cost compared with an analyst? The human side is published: BLS lists a May 2025 median wage of $78,760 a year ($37.87 an hour) for market research analysts (SOC 13-1161), wage only. The agent side is harder to pin down. ChatGPT's pricing page lists deep research as limited, expanded or maximum depending on the plan, but it did not show prices or task counts when we read it. Ask for the number of reports per month at the plan you would buy, then add the review hours.

How long does a report take? OpenAI said at launch that a deep research task takes 5 to 30 minutes. Review time comes on top.

Can it use our paid databases? Only where the tool supports a connector for that service and your workspace enables it, according to OpenAI's help center. Otherwise someone with access pulls the data.

Sources

Deep research in ChatGPT
OpenAI Help CenterPrimaryAccessed Nov 14, 2026
Introducing deep research
OpenAIVendorAccessed Nov 14, 2026
Gemini Deep Research now integrates with your Workspace content
Google Workspace UpdatesPrimaryAccessed Nov 14, 2026
Hebbia's deep research automates 90% of finance and legal work, powered by OpenAI
OpenAIVendorAccessed Nov 14, 2026
Scaling Deep Research with Multi-Agent Systems (Weber Shandwick)
FocusedVendorAccessed Nov 14, 2026
DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence
International Conference on Learning Representations (ICLR 2026)PrimaryAccessed Nov 14, 2026
Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents
arXiv (Rao, Wong and Callison-Burch)PrimaryAccessed Nov 14, 2026
Market Research Analysts: Occupational Outlook Handbook
U.S. Bureau of Labor StatisticsPrimaryAccessed Nov 14, 2026

What we could not verify

Both production accounts on this page are vendor-published: OpenAI's story about Hebbia (whose customers are not named) and Focused's case study about Weber Shandwick (which gives no figures). No independent account of a company running deep research reports in production was found.
Hebbia's time savings (30–40 hours per deal for investment bankers, 20–30 hours per deal for private equity) are vendor-reported; the page does not say how they were measured.
The Rao, Wong and Callison-Burch figures on hallucinated citation URLs come from an arXiv preprint (April 2026) that had not been peer-reviewed when read.
Agent pricing is not shown. ChatGPT's pricing page lists deep research as limited, expanded or maximum by plan but did not render prices or monthly task counts when read, and the per-plan limits OpenAI published in April 2025 may no longer apply. So the page has no cost comparison.
No source was found measuring how often deep research reports misstate figures from paywalled or private sources the agent could not open.

Need this job done?

The AI agent job board isn't open yet. Tell us about the job and Nikolas will reply personally.

I build or operate AI agents