Cooley has shipped GO Public, a proprietary IPO-workflow product built on ChatGPT Work, and the only public evidence in the fetched material is OpenAI’s own case study, published 2026-09-18. Every efficiency claim in it is vendor-framed with no independent verification. For legal-ops leads deciding whether to build now or wait, the useful finding is architectural: GO Public is designed around mandatory lawyer review, and that review burden, not drafting, is the cost center you would actually be staffing.
What Cooley shipped, and what is only vendor-claimed
The documented facts are narrow. According to OpenAI’s published case study, “Cooley developed GO Public, a proprietary AI product offering built on ChatGPT Work.” The same page reports that in 2025 the firm advised on 180 deals globally, totaling more than $51.5 billion in deal volume, and claims Cooley has advised on more venture-backed IPOs than any other firm over the past two decades.
Those figures deserve a label before any other analysis: they are Cooley-reported numbers inside OpenAI’s marketing. The case study quotes no accuracy metric, no turnaround-time comparison, and no cost figure of any kind. Nothing in the fetched material independently verifies an efficiency outcome. The 180 deals and $51.5 billion describe Cooley’s practice, not GO Public’s effect on it. Treating them as product outcomes would be a category error, and any coverage that does so is repeating the vendor’s framing.
What the case study does establish, in the vendor’s own words, is the architecture. That turns out to be the most transferable part of the announcement.
The harness is the product
OpenAI describes GO Public’s core as an “agentic harness” that “analyzes information so lawyers can review and validate the work.” The case study goes further: “The harness provides a controlled workflow for agents that lays out which steps agents can perform automatically, where lawyers must review or validate the work, and how it all comes together.”
Read that sentence as a design specification rather than marketing. It encodes three properties any firm could implement on any model platform:
- Explicit step classification. Every step in the workflow is assigned in advance to either automatic agent execution or mandatory human validation. Nothing is left to the model’s discretion about whether a lawyer needs to see it.
- Validation as a hard gate, not a suggestion. Lawyers “must” review at defined points. The workflow cannot complete through the agent path alone.
- A defined assembly step. The harness specifies “how it all comes together,” meaning the relationship between agent outputs and the final work product is documented, not emergent.
This is the part of GO Public that survives scrutiny, and it is platform-independent. A firm could build the same controlled workflow on a competing model, or on-premise, because the defensible asset is the assignment of steps to review gates, not the underlying model. That distinction matters for the ownership question below: if the workflow logic is the valuable IP, the firm’s negotiating position with the vendor should reflect that it, not OpenAI, supplied the legal process knowledge. The case study notes that Cooley “partners closely with OpenAI,” combining subject-matter expertise with AI engineering; which party owns the resulting workflow definition is not addressed anywhere in the source, and it should be settled in writing before a firm invests in codifying its own deal processes into a vendor’s platform.
Build now or wait: a four-gate checklist
The decision is not really “adopt ChatGPT Work or not.” It is whether to invest in a review-gated workflow now while gating what data flows through it. Four gates, each grounded in what the announcement does and does not provide:
| Gate | What to demand before proceeding | Status in the Cooley announcement |
|---|---|---|
| Workflow-IP ownership | Written terms assigning the workflow definition, prompts, and process logic to the firm | Absent; “proprietary product” is asserted, ownership terms are not shown |
| Verification-loop design | A step-by-step register of automatic vs. lawyer-validated actions, with named owners for each gate | The pattern is described, but no firm’s register is public; you build your own |
| Confidentiality gates | Contract-level data-handling terms before any deal-confidential material touches the model | Absent; the fetched passages show no contract terms |
| Measured metrics | Pre/post baseline: drafting hours, review hours, error catch rate, cycle time per deal | Absent; no efficiency, timing, or cost metric appears in the source |
The confidentiality gate deserves specific attention because the platform’s own marketing defines the exposure. OpenAI’s product page states that ChatGPT Work can “connect to your internal work files and apps like Gmail and Slack to create polished, finished deliverables and automate tasks.” That is precisely the data surface an IPO practice cannot expose casually: draft registration statements, diligence files, and client communications carry confidentiality obligations, and public-offering work adds securities-filing sensitivities on top. The same page offers generalized assurances that ChatGPT “is built with your security and privacy in mind” and that users choose how their data is used. Those are marketing statements, not contract terms, and they settle nothing about privilege, retention, or professional-responsibility compliance in any jurisdiction.
A reasonable build sequence follows from the table: develop the harness on non-confidential or synthetic deal material now, run it long enough to produce your own baseline metrics, and let confidential data in only when the data-handling and ownership terms exist in writing.
The economics: review is the cost center you are staffing
The intuitive pitch for drafting automation is that work disappears. The evidence points somewhere else. When drafting collapses into a generated starting point, the drafting hours shrink but the validation obligation does not; someone with a law license still has to read every generated sentence against the source material, and that reading is skilled, billable, and slow. The cost center shifts from junior drafting to lawyer review.
That shift reprices capital-markets staffing in ways a firm should model before claiming savings. Review-heavy workflows concentrate hours in more senior, more expensive timekeepers. Staffing models built on large junior drafting pyramids do not obviously survive it. And the billing question follows: clients paying for a generated first draft will reasonably ask why review time costs what drafting time used to. None of this appears in the Cooley announcement, which is exactly why the metrics gate in the checklist matters. A firm that instruments drafting hours and review hours separately, before and after deployment, can answer these questions from its own data. A firm that does not will be quoting the vendor’s narrative to its own partnership.
The Cooley arrangement is that repricing viewed from the buyer’s side: the value capture depends on terms the public evidence does not show.
Counter-evidence: the model layer is the weak link
The strongest reason to keep the harness, rather than the model, at the center of the design comes from outside the vendor ecosystem. An independent evaluation of six frontier AI models on Legal Zero-Days, defined as previously undiscovered vulnerabilities in legal frameworks whose exploitation could cause immediate societal disruption, found that “the best-performing system achiev[ed] 10.00% accuracy,” according to the arXiv paper.
Legal Zero-Days is an adversarial benchmark, and routine IPO drafting is not an adversarial task; the 10.00% figure should not be read as a prediction that GO Public fails nine times in ten. What it establishes is narrower and still important: at the model layer, novel legal reasoning, the kind without established templates, remains unreliable in ways a vendor case study will never surface. The failure mode for deal work is a plausible, well-formatted document containing an error that only a lawyer checking against the record would catch. That is precisely the error class a mandatory validation gate exists to intercept, and it is why removing the gate to save review time converts a staffing cost into a professional-responsibility and securities-filing exposure. The burden of proof for loosening any review gate should sit with measured catch-rate data from your own workflow, not with vendor assurances.
The competitive read: platforms are selling products, not seats
Cooley’s own CIO, David Wang, is quoted in the case study saying the legal industry has traditionally been relatively change-averse. Coming from the firm doing the shipping, that reads less like self-criticism and more like a competitive claim: Cooley expects slower peers to fall behind.
The structural signal supports him. Cooley is not a marginal experimenter; per Wikipedia’s profile of the firm, it reported $2 billion in revenue in 2022, and in 2024 it ranked 19th in the AM Law 200 and 25th globally in the Global 200. When a firm of that size builds a named, branded product on a vendor’s LLM platform, the platform is no longer selling seats; it is selling firm-embedded workflow products that carry the client’s own brand. OpenAI has been assembling the delivery capacity for exactly this model, as Groundy covered in its reporting on OpenAI’s deployment company and partner network.
For mid-size capital-markets practices, this raises the cost of the change-aversion Wang names. A competitor’s productized workflow compounds: every deal run through a tuned harness produces the firm’s own metrics, templates, and review-gate refinements, none of which transfer to a firm that waited. Waiting is a defensible position on confidentiality grounds, but it is not a neutral one competitively. The way to hold both positions is the harness-first sequence: build the verification workflow and the measurement baseline now, on safe data, so the only thing left to flip when contract terms arrive is the data connection.
Verdict and open questions
Proceed harness-first, not platform-first. Copy GO Public’s documented pattern, a controlled workflow that assigns every step to either automatic agent execution or mandatory lawyer validation, and begin building and instrumenting it now on non-confidential material. Gate any deal-confidential IPO work on three things the announcement does not provide: written data-handling terms, written workflow-IP ownership, and your own pre/post metrics on drafting hours, review hours, and catch rates. Staff and price for lawyer review as the durable cost center, because the independent evidence, a best-case 10.00% accuracy on novel legal-reasoning tasks across six frontier models, says the model layer cannot yet be trusted to shrink it.
The limitation is the evidence base itself. Everything known about GO Public comes from a single OpenAI-published, Cooley-sourced case study. Its deal counts and dollar volumes are self-reported; no accuracy, timing, cost, pricing, or data-handling terms appear in the fetched material; and nothing establishes that the pattern generalizes beyond a firm with Cooley’s issuer-side deal flow. Those are not reasons to wait on the workflow. They are reasons to treat every efficiency claim about this product, including any made in the next vendor case study, as an open research question until a firm’s own instrumented deployment answers it.
Frequently Asked Questions
What are the four gates a firm should check before proceeding with a review-gated workflow?
| Gate | What to demand before proceeding | Status in the Cooley announcement |
|---|---|---|
| Workflow-IP ownership | Written terms assigning the workflow definition, prompts, and process logic to the firm | Absent; “proprietary product” is asserted, ownership terms are not shown |
| Verification-loop design | A step-by-step register of automatic vs. lawyer-validated actions, with named owners for each gate | The pattern is described, but no firm’s register is public; you build your own |
| Confidentiality gates | Contract-level data-handling terms before any deal-confidential material touches the model | Absent; the fetched passages show no contract terms |
| Measured metrics | Pre/post baseline: drafting hours, review hours, error catch rate, cycle time per deal | Absent; no efficiency |
What is the recommended build sequence for implementing the workflow?
A reasonable build sequence follows from the table: develop the harness on non-confidential or synthetic deal material now, run it long enough to produce your own baseline metrics, and let confidential data in only when the data-handling and ownership terms exist in writing.
What does the independent evaluation of frontier AI models on Legal Zero-Days show?
An independent evaluation of six frontier AI models on Legal Zero-Days, defined as previously undiscovered vulnerabilities in legal frameworks whose exploitation could cause immediate societal disruption, found that “the best-performing system achiev[ed] 10.00% accuracy,” according to the arXiv paper.
Join the discussion
Share a useful perspective or ask a question about this article.