AIAI for BusinessInsurance

Sumitomo Life's ROI table and five other AI rollouts that started in the C-suite

Sumitomo Life's president started using ChatGPT during a small pilot, and weekly usage at the insurer is now above 80%. Here are six executive-first AI rollouts, what each one teaches, and a simple grid for deciding whether to lead your own rollout by example or by mandate.

Office tower at dusk, amber light spreading from the top-floor executive suite down through most floors.

Most companies judge an AI rollout by how many licences they bought. A better question is what the leaders did with the tool themselves before they asked anyone else to change how they work. Sumitomo Life gives a clean answer. OpenAI reported on November 18 that employees at Sumitomo Life who used ChatGPT more than 100 times a month reported an estimated 524% return on investment. The bigger lesson is in the story of how the insurer got there. Here is the thesis this piece argues: in an AI rollout, the executive's main job is to be seen using the tool on real work and then to measure what happens. A written mandate is optional, and it backfires when it arrives before the example.

The six entries below run from the quietest leadership signal to the loudest. I chose them for what leaders publicly did and what came of it, not for revenue or market share. This ranking is my judgement, not an established fact. Most of the figures come from the companies themselves or from OpenAI, which sells the product, so read them as self-reported results from a party with a commercial interest.

Six executive-first AI rollouts worth studying

Sumitomo Life: the president joined the pilot

Sumitomo Life is a mutual life insurance company in Japan, founded in 1907 and headquartered in Osaka. According to OpenAI's case study (vendor source), in 2025 about 100 to 150 people, including executives and members of planning teams, piloted ChatGPT Business, and in June 2026 the company moved to ChatGPT Enterprise. What makes it notable is who was in that first group: Sumitomo Life's president began using AI during the ChatGPT Business pilot. Kakeru Tateyama of the planning department explained why that mattered: "When senior leaders don't understand a new technology, it can be hard to get it off the ground." Company leaders used the platform during early trials and encouraged staff adoption through internal portals and management meetings. The use cases are everyday ones. Planning teams use it to review source material and organize their findings when they prepare proposals for senior management.

BBVA: trained its top leaders, then let demand set the licence count

In May 2024, BBVA entered into a partnership with OpenAI and provided its employees with 3,300 ChatGPT Enterprise licences. The bank's own newsroom says it then moved to 11,000 and gave the reason: "The expansion of our licences is a product of heavy in-house demand". In OpenAI's account (vendor source), the rollout included 250 senior leaders trained, including the CEO and chairman, and produced 20,000+ Custom GPTs created across the bank, with around 4,000 GPTs used frequently. That ratio of about one in five is the detail to remember. If you let people build assistants freely, you need rules for which custom GPTs get shared and maintained. In December 2025, BBVA and OpenAI announced a multi-year program to bring ChatGPT Enterprise to all 120,000 employees across 25 countries.

Morgan Stanley: earned advisors' trust with tests before scale

Morgan Stanley's contribution to this list is evaluation. The firm designed Evals, test scenarios and question sets drawn from real advisor inquiries, to systematically grade the AI's performance. Morgan Stanley reported in June 2024 that 98% of its financial advisor teams had adopted its OpenAI-built assistant, and the same figure still held in October 2025. The guardrail stayed in place. Its chief analytics officer said "Even tools like Debrief require advisor validation before sending outputs to clients." Note that the 98% counts advisor teams, not individual users.

Moderna: a CEO who put a number on daily use

Moderna started with its own tool. It launched an internal AI chatbot tool, mChat, at the beginning of 2023; built on OpenAI's API, mChat was adopted by more than 80% of employees. OpenAI's case study (vendor source) says that within two months of the ChatGPT Enterprise adoption, Moderna had 750 GPTs across the company and 40% of weekly active users created GPTs. The leadership signal was a number: in 2024, Bancel set a goal for employees to use ChatGPT "no fewer than 20 times a day" (Wall Street Journal), and the CEO and executive committee role-modelled it. One independent tracker adds a caveat. It notes that Moderna has published almost no results, and the executive who led its AI rollout left for OpenAI in January 2026.

Shopify: tied AI to reviews and headcount, then admitted the side effect

Tobi Lütke told Shopify staff in a memo he published on Apr 7, 2025 that reflexive AI use is a baseline expectation, that AI use would count in performance reviews, and that teams must show why AI cannot do the work before asking for more headcount. The memo connected AI use to two decisions managers make every quarter, which explains why it carried weight. About 17 months later, Lütke described the cost on a podcast: "The failure case now of lazy work is not lack of output." His advice was to use the model to make a point shorter, not longer. A leader who points out the failure mode of his own mandate is rarer than the mandate itself.

Duolingo: the memo that went public without its context

CEO Luis von Ahn posted an internal memo on LinkedIn on April 28, 2025, announcing Duolingo would become an 'AI-first' company. It said the company would "gradually stop using contractors to do work that AI can handle." After weeks of pushback, he wrote: "When I released my AI memo a few weeks ago, I didn't do that well." "Internally, this was not controversial," von Ahn told the Times. The lesson is about who reads the memo. An internal instruction posted publicly is read by contractors, users and the press, none of whom share the internal context. The quieter fix, dedicating every Friday morning to AI experimentation, is the part most worth copying.

Do executives need to use AI first for teams to adopt it?

Yes, as a signal of permission. No, as a source of usage volume. A study of ChatGPT Enterprise by OpenAI-affiliated researchers (a vendor source, though built on workspace data rather than surveys) found that at the average organisation managers and directors represented approximately 24 per cent of observed weekly active users, while executives were around 10 per cent. Early-career workers sent roughly eight to nine more messages per week than the average user inside the same firm, while executives generally sent fewer messages.

So the people who use the tool most are not the leaders. The leaders' contribution is legitimacy: the visible proof that using AI on real work is allowed, safe and noticed. Every case above that worked combined that visible example with something measured. Sumitomo Life tracked usage tiers, BBVA watched demand, and Morgan Stanley ran tests. Duolingo shows the other way it can go: a rule that reached readers before the context or the example did. Sumitomo Life also did the unglamorous work first. As a financial institution, it worked with OpenAI to assess the Enterprise tier against its own standards, covering workspace access controls, feature permissions, and OpenAI's policy of excluding business data from model training by default. Executives can only use the tool openly if those data controls are settled before the first prompt.

The model-or-mandate grid for an AI rollout

The rule fits on a whiteboard: model it before you mandate it, and measure it before you expand it.

To place your own rollout, ask two questions. Do senior leaders visibly use the tool on their own work? Is AI use tied in writing to a decision such as reviews, headcount or contractor spend?

  • Leaders use it, no written rule: pull. Sumitomo Life's pilot and BBVA's first phase sit here. Adoption spreads because people copy what they see.
  • Leaders use it and a rule exists: pull with consequences. Moderna and Shopify sit here. This works when the leader also names the downside, as Lütke did with over-long output.
  • A rule arrives before the example or the context: push. This is how Duolingo's memo landed outside the company. Expect backlash and a clarification within weeks.
  • Neither: drift. Staff use personal accounts, nobody measures anything, and the first real data you get is an incident.

Most teams should start in the first box and move to the second only after they have numbers.

Reading Sumitomo Life's usage tiers: a worked example

The useful part of the Sumitomo case is how it splits return on investment by usage. The overall estimated ROI averages 203%, based on employee-reported time savings, and weekly usage of ChatGPT Enterprise has passed 80%. Infrequent users, by contrast, logged an estimated 55%, nearly a tenfold gap inside one company.

Here is a hypothetical, arithmetic only, using those two tier figures. Picture a 100-seat pilot where every user sits in the infrequent tier at 55%. If 10 of those people move up to the heavy tier at 524%, the average becomes (90 × 55 + 10 × 524) ÷ 100, or about 102%. Converting a tenth of the seats roughly doubles the blended return. That is why an executive using the tool in a management meeting can matter more than a hundred extra licences: it moves people between tiers. One caveat applies. These are self-reported time savings published by the vendor, and heavy users may simply be the people whose work suited the tool in the first place.

What to do on Monday to start an executive-first rollout

  • Ask each member of the leadership team to bring one real task from this week that they ran through the approved tool, and to show the prompt and the output at the next management meeting.
  • Export licence-level usage and sort users into monthly-use bands, with 100 or more as the top band, as Sumitomo did. Write down how many sit in the bottom band.
  • Before anyone senior uses the tool openly, list four controls: access, feature permissions, training exclusion and retention. Get written answers from IT.
  • Count your custom GPTs and how many were opened in the last 30 days. BBVA's 20,000 to 4,000 ratio is a realistic benchmark.
  • Before sending any AI memo, have someone outside the team read it as if it had leaked, and cut every line they would take as a threat to jobs.
  • Set one norm on output length, borrowed from Lütke: use the model to make messages shorter.

Who to watch: Sumitomo Life, which plans to expand AI access to a larger group of employees during the second half of fiscal 2026, will show whether a pilot led from the top still works when the wider workforce joins.

The number to copy from Sumitomo is the tier table, not the 203%. If you can't say how many of your users sit in the bottom band, you don't yet know what your rollout is worth, and the leaders using it visibly is the cheapest way to change that.

Go deeper

The lessons that take this article further, free to read.

  1. 1Sharing, governance, and custom gpts at workChatGPT & the OpenAI ecosystem
  2. 2Security, privacy, and data controlsChatGPT & the OpenAI ecosystem
  3. 3What custom gpts are and the GPT storeChatGPT & the OpenAI ecosystem
  4. 4Meetings, email, and admin: reclaiming hours each weekAI in daily work
  5. 5Privacy and confidential data: what not to pasteResponsible & trustworthy AI

Sources

Finished reading?

Validate your read to earn XP and feed your radar.