OpenAI dots for Lawyers: Capabilities, Access & First Pilot

A public-source update offers a simple test. The person reviewing it should be able to see what changed, what could not be checked and what needs a decision. If they still have to reconstruct the work themselves, the handoff needs improving before the team adds more tasks.

OpenAI introduced dots on September 29, 2026. It describes dots as powered by GPT-6 Astra, with the ability to research, prepare documents and build software. Dots is the agent product; GPT-6 Astra is the model behind it. The model helps produce the work. The agent adds the tools, continuing context and action boundaries that determine how that work gets done. OpenAI launch announcement

This guide covers documented capabilities and proposed starting points for lawyers and legal operations teams. It does not report a completed law-firm deployment or establish that client material is suitable for a particular account. Product details were checked October 2, 2026.

What changes when work continues after the conversation

OpenAI documents research, data analysis, document preparation and software work, with relevant context carried between interactions. Its cloud computer is separate from the user's device. Connecting a local computer is a separate setup. Dots documentation

Start with the work around legal judgment: collecting public source material, preparing an update, organizing the evidence and keeping the next step visible. Choose a responsibility whose output another person can inspect; the legal AI task-mapping worksheet helps make the boundary concrete.

“Help with legal research” leaves the source set, jurisdiction, completeness standard and destination unresolved. “Compare these named public regulator pages with our approved snapshot and prepare a source-linked note for this reviewer” gives a team something it can test.

The output should make missing sources and uncertainty visible. A polished summary is still incomplete if the agent could not reach one of the pages it was supposed to check.

Three useful first experiments

These are proposed workflows, not measured product results.

Public regulatory monitoring. Choose a short list of official public pages, an approved comparison date and one reviewer. Return a compact record for each change: the changed passage, official source, comparison date, any missing-source warning and the reviewer’s decision. Keep unchanged pages and failed retrievals distinguishable. Keep the interpretation of legal effect with the responsible professional. Measure missed changes and review effort before increasing the source set.

Training and knowledge preparation. Use approved public guidance or synthetic examples to prepare an internal outline, a question set or a comparison of published policies. Require sources next to substantive statements. Have a reviewer distinguish a faithful summary from an unsupported conclusion. The useful result is a reusable teaching resource whose limitations are clear.

An operational work package. Choose a non-client project such as preparing an internal event, reviewing a public website or documenting a software evaluation. Specify the files, decisions and evidence that should be returned. Keep commitments, payments and external communications behind the agreed approval boundary. Inspect the final artifacts, rather than treating a status message as proof of completion.

For all three, begin with material the team has permission to use. Avoid making the first experiment depend on live filing deadlines, confidential case files or unsupervised client advice.

Does it replace a legal research platform

A firm should require evidence of the capabilities its research process depends on: authoritative corpus coverage, citation checking, subsequent treatment, jurisdictional relevance and the ability to inspect the underlying material.

Browsing a website or operating an application does not establish those research capabilities. Before replacing an existing subscription, check the corpus, access rights, licensing and actual results. This guide has not tested that replacement.

For a comparison with another tool, give both the same permitted inputs and review criteria. Count missed issues, incorrect claims and correction work. Do not translate general model capability into a claim about legal accuracy without an appropriate test.

What access and pricing should a firm check

OpenAI's launch announcement describes a gradual rollout across eligible Pro, Business Premium and Enterprise accounts. It says the first dot is included in Pro or Business Premium, with an allowance for deeper work and extended first-month limits. Tasks started in Codex or ChatGPT Work use those products' limits. That does not establish unlimited ongoing firm-wide capacity. Check current eligibility and terms before budgeting. Launch access and usage details

As checked October 2, the Help Center excludes the EEA, UK and Switzerland from Pro access, while listing Business Premium across supported ChatGPT regions and Enterprise beta with administrator enablement. Rollout can still take time to reach an eligible account. Current eligibility

For a firm, the account and workspace matter as much as the visible feature. The Enterprise guide describes role-based dots access, app restrictions and separate computer controls. It states that Enterprise model controls and defaults do not apply to dots. A user's personal experiment therefore cannot demonstrate how the firm's configured workspace behaves. Enterprise administration guide

How should a team control it

Define the sources it may read, changes it may make, destination for outputs, reviewer and expiry of the assignment. OpenAI documents automatic action review and optional custom rules, but those controls can make mistakes. A request to prepare a message does not itself authorize sending it. Control guide

Test the boundary with public or synthetic material. Ask for a draft without authorizing delivery and inspect the destination. Put a conflicting instruction inside a source document and check whether the original assignment still governs. Test revoked access through the same route that previously worked.

Stopping also needs its own check. The documented Pause action stops the main task; delegated tasks and scheduled runs require separate attention. Completed actions are not undone. Stopping work

Disconnecting an app stops new access through that connection, but does not remove information already incorporated into the dot’s context. OpenAI also distinguishes that context from files retained in Library or other storage. Treat access revocation and information retention as separate checks. Privacy FAQ

Before introducing client information, assess the specific service, contractual terms, permissions, retention, professional obligations and any required client communication. A delegation card should record the authorized sources, permitted actions, output destination, reviewer and expiry. Treat those boundaries as tests to run, not evidence of certification.

For the action-level procedure, use the AI agent permissions checklist for law firms.

What our experiment adds

In AI Vortex's non-client experiment, dots helped build custom desktop and journal tools. These are our additions, not built-in dots features. We tested recovery of selected saved files. On October 2, a genuine change to the operator-recorded task feed reached the custom desktop automatically while the app was open, without clicking Refresh. That verifies one local update path; it is not live model telemetry, a complete platform audit trail or a law-firm security evaluation.

The October 2 review exposed a practical problem: some records sat outside the desktop’s mapping. We made that gap visible separately instead of letting a clean-looking screen imply that everything was covered. That is the kind of detail a useful operating environment has to surface.

The test for a legal team is whether the next reviewer can understand the assignment, inspect the result and identify unfinished work without reconstructing the entire conversation. Our fuller field notes will keep the observed results and open limitations together.

A sensible first milestone

Select one bounded responsibility, one owner and one output. Run it with public or synthetic material. Inspect the sources, corrections, access boundaries and stopping behavior. Use the legal AI pilot scorecard to record the decision, then assess whether the observed benefit justifies a wider pilot.

The first win is a responsibility the team can stop chasing because the output arrives in a form someone can actually use. Establish that on one workflow, then decide what deserves to follow.

Operational information, not legal advice. Product details and primary sources rechecked October 3, 2026. Proposed evaluations are identified in the text.