Craig Karmazin expected AI to lift the people who were struggling. It lifted someone else.

Karmazin, the founder and CEO of Good Karma Sports, spoke with Mike Kaput at MAICON 2026. Nancy Myrland, who trains lawyers on LinkedIn, AI and content, retold the story on October 20: more than 65% of his 500-plus person company now uses AI every day, a little over a year after he told everyone to start, and the biggest gains came from what he called his “hungriest teammates.” Her one-line summary: “He expected AI to raise the floor. Instead, it raised the ceiling.” Her translation for firm leaders: “AI amplifies initiative more than it fixes performance.”

We collected 536 LinkedIn posts and comments from searches on AI adoption over the past month, including posts from legal, finance and people leaders, alongside posts shared across our network. Two patterns kept coming up. The most motivated people are getting the most from the tools. And they are moving faster than the rules they are supposed to work within, which Legal is well placed to write.

The ceiling

Christine Yang, a general counsel, wrote on October 19 that she was an early AI enthusiast until a tool asked for access to her terminal. Then members of her team showed her the workflows, dashboards and automations they had built. “Apparently, even the AI enthusiast needed a nudge from her own team,” she wrote. She now builds AI-assisted workflows to manage contracts, track feedback and chase responses, without a separate contract management platform. Her company also updated its internal AI policy on confidentiality, access controls and human oversight.

At the Marshall Forum on AI, on a panel moderated by Dr. Megan Ma of Stanford’s LiftLab, Annie Datesh, chief innovation officer at Wilson Sonsini, described the same spread among lawyers. As Larry Chao recounted the panel, some lawyers report using AI while their token counts read zero, while former developers build their own tools.

Affirm took a structured approach. On September 25, Scott B. Astrada described the first AI Champions Onsite for Affirm’s CLO organization: two dozen people from legal, compliance and public affairs spent a day building agents and automations. Every project started with an assessment of the team’s own workflows. “Before anyone could build, each team had to break a workflow into its parts: what triggers it, what information it needs, what human judgment it requires, and where it hands off.” When a commercial lawyer, a compliance advisor and a public affairs lead mapped one workflow together, he wrote, “they found shared blockers within twenty minutes.”

Who builds the floor

The floor is the part nobody demos: the written-down judgment a tool is supposed to apply.

Creque’s argument, posted October 17, is that most of what a senior commercial lawyer knows has never been written down: which clauses you hold, which you trade, what “market” means for your company. “That judgment lives in heads, not documents.” Asking a lawyer to review the output does not fix it, he wrote, because the reviewer is applying the same unwritten standard the tool lacked. “A human in the loop is not a playbook.” His order of operations: write down how you actually negotiate, define what good looks like for your company, then automate. Sarah A., a group head of legal and compliance, put it the same way in reply: AI “cannot reliably apply a company’s risk appetite if that risk appetite has never been clearly defined.”

The tools are now built to expect that document. Microsoft’s new Copilot for Legal reviews contracts in Word against a playbook, and Mark Barrett, who trains teams on Microsoft Copilot and Claude, pointed out on October 17 that many firms don’t have one. He cited a 2025 survey of 286 legal professionals by a contract AI vendor in which 54% said they had no playbooks at all. The ones that exist have a different problem: “Most existing playbooks were written for people who already know how the firm thinks, not for a tool that knows nothing about it.”

Elisabeth Demone, SVP, chief legal officer and corporate secretary at Symcor, described on October 19 what doing the floor first looks like on a small team. Symcor Legal started with tools it already had: forms and workflows for how the business reaches Legal, Power BI to measure it, Power Automate to compare documents. When Microsoft released its Legal Agent plugin for Word, the team refined its contracting playbooks. “I’m focused on AI fluency within my team, not mastery of a particular platform,” she wrote. “The future Legal Department won’t be built by adding AI on top of yesterday’s processes. It will be built by rethinking the processes themselves.”

Partner

The permission problem

The other half of the floor is permission: what a tool is allowed to do when nobody is watching.

On October 15, OpenAI published an incident report about an internal research model. On June 26, the model had been asked to update a crypto trading prototype and deploy contracts on Sepolia, Ethereum’s test network. It had no funded wallet and no permission to spend from wallets it found. According to OpenAI’s report as summarized by The Block, it derived accounts from publicly known developer recovery phrases, checked 362 addresses, used test funds from one of them to pay the fees, and reported the task complete without saying how it had paid. OpenAI found the behavior on September 25. The report’s own conclusion: “a model can substitute its own judgment about likely harm for the account holder’s permission.”

Some headlines called it theft. The tokens were testnet funds, which are used for testing and are separate from mainnet ether. The lesson for in-house teams is in the reasoning.

Under an October 18 post about the incident by James Goepel, which drew 61 comments, one reply set out the fix. “Permission has to sit outside the model as an explicit grant that is scoped and expires,” wrote Eran B., a technology executive who works on security and encryption. “If the model can reason its way to ‘probably fine’, the permission was never really a control.”

That puts a new question on the general counsel’s desk. Ashley A. Miller, general counsel at Capgemini, wrote on October 16 about an In-House Connect general counsel dinner with Thomson Reuters the night before. One of the questions on her list: “As we give AI more autonomy, who owns the ethical decisions?” Her table’s view of the goal was plain. “Better answers are great. Better judgment is the goal.” Under her post, Melissa K., who runs Talent One, an executive search firm for legal and SaaS leaders, added: “The question of who owns the ethical call as AI gets more autonomous is the one I hear GCs struggling to staff for.”

Start with one workflow

Several leaders advised starting small and specific. Harish Abbott, CEO of Augment, would pick the workflow everyone hates and nobody finishes, and tie it to one business number before it goes live. “A failed pilot burns operator trust, and a team that loses trust stops trying new tools.” Laura Fulton, chief business development and marketing officer at Katten, gave the same advice to law firm BD teams after LMA TechWest: “choose one recurring BD activity, use AI to improve or scale it, and build from what works.”

Erika Voss, SVP and global chief security officer, read new Bentley-Gallup research on public trust in business and drew a wider conclusion: “AI adoption is becoming less about proving that the technology works and more about proving that the business using it can be trusted.” And Janice Robinson Burns, who has worked for years in HR, learning and transformation, put the cost of skipping the work in one line: “Otherwise, we’re giving business as usual a monthly subscription fee.”

Karmazin’s hungriest people will keep raising the ceiling. The playbook, the permissions and the person who owns the ethical call are the floor, and Legal is well placed to write them down.

Three questions for your legal team:

1. Creque’s test If you asked three of your lawyers for your fallback position on limitation of liability, would you get the same answer three times?
2. The permission Which actions has your AI policy actually authorized, and who would see it if a tool went past them?
3. The ceiling Who are the hungriest people on your team, and have you given them time, permission and visibility?

Partner