AI
Legal Chiefs Get an AI Mandate Without the Authority
Chief legal officers are told to govern enterprise AI, yet only 8.7% own it while legal teams already run the tools.
Only 8.7% of general counsel own AI governance at their companies, a 2026 Plexus survey of 150 legal chiefs found. Generative AI use inside those teams has reached 87%, according to FTI Consulting and Relativity’s General Counsel Report.
Boards still want the chief legal officer to answer for hiring algorithms, automated pricing, and agents that act without a person in every step. The systems themselves usually report to IT.
Only 8.7% of General Counsel Own AI Governance
Plexus, the legal operating-system vendor, surveyed 150 general counsel across Australia, the United States, New Zealand, and EMEA for its Future-Ready General Counsel 2026 report. The study found that only 8.7% of general counsel own AI governance in their organisation, while IT or the CIO owns it in 28.7% of organisations, more than three times as often.
That gap is the second move in a story that is usually told as a promotion. Legal is being asked to design the rules for systems it does not run, then to defend the outcomes when those systems harm a candidate, a customer, or a counterparty. Ownership of the stack has already drifted toward technology, and in some banks toward banks paying millions for AI chiefs whose job is to fade into the operating model.
THE GOVERNANCE VACANCY IN 150 LEGAL DEPARTMENTS
- GC ownership: Only 8.7% of general counsel own AI governance in their organisation.
- IT ownership: IT or the CIO owns it in 28.7% of organisations.
- Active adoption: 58.7% of GCs are piloting, scaling, or running AI in legal work.
- Full rollout: Only 6.7% have fully operationalised AI-enabled operating models.
Among teams that already use AI, 43% report a 21% to 40% drop in manual legal work, so the productivity case is not theoretical. The same survey still has 69% of GCs spending less than 40% of their time on strategy, and 32% of organisations still treat legal as a cost centre. A title change to architect of AI governance does not, by itself, move either number.
Plexus also found that 35.3% of teams are still in pilots, the largest single group. Another 21.3% are exploring and have implemented nothing. Ten-point-seven percent have no AI strategy at all. Most of the profession is a year or two from an operating model that would survive a regulator’s file request.
High-Risk EU Rules Slip to December 2027
The legal clock did not wait for that operating model, and then it did. The EU AI Act entered into force on 1 August 2024. Prohibited practices have applied since 2 February 2025. Transparency duties in Article 50 applied from 2 August 2026 and were not postponed.
What moved is the hard end of the statute. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was dated 8 July 2026, published on 24 July, and entered into force on 27 July 2026. Stand-alone high-risk systems in Annex III, including tools used in hiring, credit, education, and access to essential services, now face duties from 2 December 2027 instead of 2 August 2026, a 16-month deferral. High-risk AI built into regulated products under Annex I moves to 2 August 2028.
THE EU AI ACT CLOCK AFTER THE OMNIBUS
- 1 August 2024: The AI Act enters into force, with duties staged over following years.
- 2 February 2025: Prohibited practices apply and stay in force.
- 2 August 2026: Article 50 transparency duties apply and were not deferred.
- 27 July 2026: The Digital Omnibus enters into force, rewriting later dates only.
- 2 December 2027: Annex III high-risk obligations apply for stand-alone systems.
- 2 August 2028: High-risk AI embedded in Annex I products must comply.
For a CLO in a multinational, the delay is not a holiday. Article 50 is already live. So are the bans. New York City’s Local Law 144 still requires bias audits of automated hiring and promotion tools, published results, and notice to candidates. China still layers scenario rules on top of its cybersecurity, data-security, and personal-information statutes, watched by the Cyberspace Administration of China. Voluntary yardsticks such as NIST’s AI Risk Management Framework and ISO/IEC 42001:2023 fill the gaps between those regimes. A company that waits for December 2027 still has to explain a hiring model, a chatbot, or an agent that is already in production.
Who Gets Named When an Agent Fails?
The essay that launched this debate says legal should not seize every AI decision from the business. It should write the rules by which those decisions are made: who may deploy a system, which uses need extra scrutiny, what evidence must be kept, when a human has to step in, and who answers when harm follows. That is a clean sentence on a whiteboard. It is a mess in an org chart.
An 8×8 survey of 2,501 CIOs and CTOs, run by Censuswide in July 2026 across the UK, the United States, France, Australia, and Ireland, asked who is on the hook when an AI agent makes an error. The CIO was named far more often than legal.
WHO TAKES THE BLAME WHEN AN AGENT ERRS
| Function named | Share of respondents |
|---|---|
| CIO | 52% |
| Customer service leadership | 16% |
| Legal or compliance | 6% |
Legal is told to own the aftermath and, in this sample, is almost never named in the incident. That is not protection. It is a delay. A discriminatorily trained hiring model, a hallucinated filing, or a leaked privileged memo still becomes a legal problem the moment a regulator, a court, or a plaintiff arrives. The CLO then inherits a system whose logs, vendors, and approval chain were built by someone else.
Policies written for employees do not automatically cover agents acting for those employees. Access rights, audit trails, and retention schedules that were designed around people leave a hole when software drafts, files, or decides. The hole is already visible to practitioners who spend their week on agent onboarding rather than on another set of principles.
Generative AI Has Already Rewritten Legal Work
While the enterprise argument rages over who owns the rules, the legal department has already changed its own workflow. FTI Consulting and Relativity’s General Counsel Report, released in March 2026, found 87% of general counsel reporting generative AI use on their teams, nearly double the 44% in 2025 and more than four times the 20% recorded in 2023. Thirty-nine percent now list AI among strategic priorities for efficiency. Departments with a formal technology roadmap rose to 53% from 25%, and about 70% plan to spend on new technology in the next 12 months.
A 2023 Goldman Sachs analysis by Joseph Briggs and Devesh Kodnani put 44% of U.S. legal tasks in reach of generative AI, with the equivalent of almost 40% of legal-industry jobs exposed to replacement. Mass job loss has not arrived. What arrived is a rebase of how the work is done, and of what a junior lawyer’s week is for.
WHERE LEGAL TEAMS PUT GENERATIVE AI FIRST
- Summarisation: 83% of GCs say their teams use or are testing it.
- Meeting notes: 67% use generative AI here, which can turn meeting notes into searchable legal records.
- Contract clauses: 63% use it to find specific terms.
- Privilege review: 34% say they are most uncomfortable using it for this task.
Comfort is highest on research, e-discovery, document review, and contract drafting. It collapses on privilege, investigations, and data-breach response. Confidentiality is the live fault line. A model that ingests a deal room, a litigation hold, or a board pack can spray that material into a vendor’s context window. Courts have already sanctioned lawyers for fabricated citations, which is a reminder that human review is still the control, not the footnote.
The Thomson Reuters Institute has estimated that AI could free 240 hours a year per professional, worth about $19,000 each, up from a 200-hour estimate the year before. That figure is a forecast of redirected time, not a payroll result. If the hours come out of first-year review, the apprenticeship that used to teach judgment shrinks with them. If they come out of outside-counsel invoices, the billable-hour bargain with law firms has to be rewritten. Outcome-priced legal agents are already being sold into in-house teams on that premise, which is a different business than a department that bills time and a firm that sells hours.
Harvey, ContractMatrix, and the Seven-Hour Review
Gartner has pointed corporate legal departments at six uses that combine value and feasibility: extracting and classifying contract data, automated review and redlining, summarising documents, intake and triage, transcribing and summarising meetings, and scoring contract risk. Those are the jobs that used to eat the week. Purpose-built products such as Harvey, Thomson Reuters CoCounsel, Lexis+ AI, Legora, and Eudia now sit on top of them, while Microsoft, Anthropic, and OpenAI push general models into the same workflows.
A&O Shearman put Harvey into enterprise use in December 2022. The firm has nearly 4,000 lawyers. Harvey now helps 4,000 of its staff across 43 jurisdictions. Staff save 2 to 3 hours a week on summarisation, analysis, and translation, a different metric from the contract-review figure that followed. With Harvey and Microsoft, the firm built ContractMatrix in 2023, grounding output in the firm’s own precedents so a tracked-changes loop keeps a lawyer in the file. About 2,000 of its lawyers use that tool daily. Harvey’s customer page says attorneys cut contract review time by 30%, saving up to an estimated 7 hours on an average review.
Our goal is to liberate lawyers from mundane, routine tasks, so that they can concentrate on what matters: strategic thinking, providing expert advice, and exercising sound judgment.
David Wakeling, Global Head of the Markets Innovation Group and AI Advisory Practice, A&O Shearman
The next step is agents that plan multi-step work. A&O Shearman and Harvey have been rolling out agents for antitrust filing analysis, cybersecurity obligations, fund formation, and loan review, used inside the firm and sold to clients and other firms. Once a system drafts, decides, and acts with less direct human involvement, the CLO’s new problem is no longer theoretical. Liability has to be allocated among the vendor, the company, the business owner, and the executives who authorised the tool. An audit trail that cannot reconstruct that chain is not a trail.
HSBC Screens About 980 Million Transactions a Month
Scale makes the same problem concrete. Jennifer Calvery, HSBC’s group head of financial crime, has described a Google-built system the bank calls Dynamic Risk Assessment, piloted in 2021. The bank checks about 980 million transactions each month for signs of financial crime.
We’re finding two to four times more financial crime than we did previously, with much greater accuracy.
Jennifer Calvery, Group Head of Financial Crime, HSBC
Calvery also wrote that the bank now has 60% fewer false positive cases, and that analysis of billions of transactions across millions of accounts has gone from several weeks to a few days. For a CLO, the performance claim is not the end of the file. Someone has to say how the model is validated, how missed cases and false hits are handled, what human oversight remains, and what evidence can be produced if a regulator later challenges a decision. The larger the operational win, the heavier that file becomes.
Accountability Needs a Named Owner
The traditional CLO brief, regulation, litigation, contracts, intellectual property, investigations, and board advice, has not gone away. It has been stretched over hiring tools, pricing engines, customer bots, fraud models, coding agents, and autonomous workflows, many of them running on law that is still being written. Legal should not own every deployment. It does have to own the question of whether the company can reconstruct a decision after it blows up.
That means an inventory of significant systems and agents, named owners, risk-based approval thresholds, testing and monitoring, incident paths, and documents that would survive discovery. It also means treating agents as a class of worker with onboarding, access control, and a person who remains answerable for what they do. Change still has to start in the legal department itself, because a team that cannot govern its own drafting tools will not be believed when it writes rules for everyone else.
The 8.7% who already own the mandate are having a different conversation with the chief executive than the 28.7% who watch IT hold the pen. The 87% who already use the tools are not waiting for that conversation to finish. December 2027 is a date on Annex III. It is not a date on the next agent that files, prices, or hires in the company’s name.
Disclaimer: This article is news reporting and analysis of corporate legal roles, published surveys, and public regulation. It is informational only and is not legal advice, compliance advice, or a recommendation that any company adopt, delay, or abandon a particular AI system or governance model. Readers who need to act on AI regulation, privilege, employment tools, or vendor contracts should consult a qualified attorney licensed in the relevant jurisdiction before making decisions. Figures and legal dates reflect the surveys, company pages, and official texts cited here and can change as new reports and implementing rules appear.
-
AI5 months agoFable 5 Came Back Under a Commerce On-Off Switch
-
AI5 months agoGoogle’s SpaceX GPU Lease Has a Sept. 30 Deadline
-
CRYPTO5 months agoPlasma One’s XPL Locks Face a 1.81 Billion Cliff
-
APPS5 months agoDGO’s Rs 549 World Cup Pass Cost Fans Sleep and Data
-
AI5 months agoMoonshot AI’s $30 Billion Ask Became a $35 Billion Close
-
NEWS6 months agoColorOS 17 Device List Spans Oppo, OnePlus and Realme
-
GAMING5 months agoThe RTX 4050 Under Rs 70,000 Hides a Wattage Gap
-
AI5 months agoBoris Cherny’s Loops Still Demand a Higher Production Bar
