ISSUE 08 / NOV 2, 2026
This week in enterprise AI: what happened, why it matters, what I'd do.
AI worked for one team. Now what?
This week: Stellantis, Stanford Health Care and Credit Saison.
|
■ FROM THE PEOPLE DOING IT “The best thing that you can do for your career is to make yourself redundant.” Ned Curic, Chief Engineering & Technology Officer, Stellantis |
|
PATTERN WATCH Before you scale one team's AI win, set your targets and write your own test first, then count the checking work every other team still does. Tools that fit routines people already run spread fastest. |
|
|
In Fortune's reporting, Stellantis engineers estimate one team built a hands-free driving system in six weeks instead of a year or more. At Stanford, Heidi Porter said, human triage missed half of eligible surgery patients before its AI tool. And OpenAI published Credit Saison's own figure: a 954% pilot return against a 500% target. Which of your teams still checks the AI's work? Who wrote the test your tool must pass? Did you set your targets before the pilot started?
— Manny
|
|
|
■ CASE STUDY · AUTOMOTIVE
Stellantis: three-minute AI drag estimates, but still one wind-tunnel car daily
What happened Ned Curic, Stellantis' chief engineering and technology officer, says some teams show "phenomenal successes". But "as a whole organization, we are not there." Engineering manager Joydip Saha's team built a hands-free driving system in six weeks instead of a year or more. Yet he bars engineers from using AI to architect systems so they don't forget how to debug. AI returns a drag estimate in about three minutes instead of 14 to 24 hours, yet the wind tunnel still tests "typically, one car a day." Infotainment director Mel Roberts checks all 10 to 15 sources her AI research tool returns: "I don't think it saves time."
Why it matters Stellantis set its target on the whole job: four years down to 24 months. Its technology chief still says the company is not where it needs to be, despite "phenomenal successes" on some teams. Many leaders would have announced the fast team as the win.
WHAT I’D DO Measure net time saved across representative teams, counting the time spent checking and reworking AI output, before the result goes into an enterprise business case. First: Finance agrees with each team lead whether the net time saved cuts a cost line, raises throughput or improves another measured outcome. As you do it: The sponsor adds the net figure to the monthly review and names the work people must still do without the AI. Then: Team leads whose numbers hold up coach the next teams, and finance retests the figures every quarter as the tools change. |
Read the source →Fortune, Nov 1, 2026
BEHIND THE NUMBERS Nearly every Stellantis engineer who described an AI speed-up also described checking the machine's work, and the speed-ups are the engineers' own accounts to Fortune. If your AI business case counts headline speed-ups but not review time, ask whether the released capacity cuts cost, raises throughput or improves another measured outcome before approving spend. | HOW SOLID IS THIS Independent reporting, results self-estimated Unconfirmed claim ○───○───●───○ Outside proof Fortune reporting from two days of on-site interviews. The speed-ups and workload figures are the engineers' own estimates and were not independently measured. |
|
|
|
■ CASE STUDY · HEALTHCARE
Stanford: human triage missed half of eligible surgery patients.
What happened Stanford Health Care's Surgical Co-management Navigator screens next-day elective surgeries for patients who need surgical co-management, where clinicians share care before, during and after surgery. Before launch, routine human triage missed half of eligible patients. The AI reads pre-op notes and medication lists and applies Stanford's own rubric, answering only yes, maybe or no. Stanford tested five models against consensus chart review and picked o3-mini, which performed competitively at lower cost. Every case appears in the electronic health record list, and clinicians can confirm, decline or override any of them. A review of missed cases, a small sample, found more errors in Stanford's process than in the AI.
Why it matters Stanford could pick a cheaper model because it owned the test: its own rubric and chart review. When results are poor, your team may look only at the AI output. Then you may overpay for a better tool to make up for errors in your process.
WHAT I’D DO Before you buy or swap an AI tool, ask your team lead to build a test from your own real cases. The people who do the work should agree on the right answer for each. Run each candidate tool against it, then choose on the results and the cost. |
Read the source →Practical AI Podcast, "Abundant Intelligence at the Midwest AI Summit", Oct 29, 2026
I wrote about this: Test new models on your own work →
HOW SOLID IS THIS Third-party account, not independently verified Unconfirmed claim ●───○───○───○ Outside proof Retold by Heidi Porter, principal at Moyer Group AI, in a Midwest AI Summit talk aired on the Practical AI Podcast, not by Stanford itself. The missed-case review was a small sample, and the figures were not independently checked. |
|
|
|
■ CASE STUDY · FINANCIAL SERVICES
Credit Saison: 954% pilot ROI against a 500% target, rollout approved in a month
What happened Looking back on 2025, credit card company Credit Saison piloted ChatGPT Enterprise from late June to early August with 315 employees, every executive included. That was just under 10% of its about 3,700 staff in Japan. It reports a 954% return and 14 hours saved per person a month, against targets of 500% and 7.5 hours. It approved a companywide rollout one month after the pilot ended. At its card counters, a custom GPT turns the audio of each day's team debrief into results, notes and handover items. Teams run the meeting as usual, so it is easy for staff new to AI. It is now used at every location.
Why it matters The debrief tool reached every counter because it fit a meeting teams already held. Credit Saison set its return and hours targets before the pilot. The pilot drew volunteers, nominated staff and every executive.
WHAT I’D DO Pick one meeting your teams already hold every day or week. Ask one team lead to try an AI summary of its recording, and count the minutes of paperwork it removes. |
Read the source →OpenAI, Oct 29, 2026
HOW SOLID IS THIS Customer-reported, not independently verified Unconfirmed claim ●───○───○───○ Outside proof A vendor naming a customer's result: Credit Saison's own figures, published by OpenAI. The pilot group included volunteers and staff nominated by the project team. The source does not say how ROI was calculated, and none of the figures were independently checked. |
|
|
Microsoft offers 60,000 buyouts while J&J helps people opt out of AI and find another path Microsoft offered voluntary buyouts to 60,000 people, a quarter of its staff. Four tech giants are projected to spend $700 billion on AI infrastructure. J&J helps staff who don't want to work with AI plan their departure or find jobs that don't use AI.
HBR, Oct 30, 2026 Act: An open model turned Firefox fixes into working attacks for about $1,000 each. Z.ai's open-weight GLM 5.3 reverse-engineered Mozilla's published fixes and built working attacks against six Firefox versions before users had the update. The fastest attempt took 12 hours. No provider safeguards were triggered. Shorten the time between a fix's release and installing it, and keep an inventory of the software you run, including open-source software.
Epoch AI, Gradient Updates, Oct 30, 2026 Act: Top agent frameworks lack persistent agent IDs and records of human approvals Before agents touch sensitive systems, ask each framework vendor whether it records an ID that stays with each agent and which person approved each task. A study of the four most-used frameworks found major gaps in these records, so you may not be able to reconstruct an incident.
MIT Technology Review, Oct 29, 2026 |
|
In Boston last month, I talked about an insurer that used AI to make it faster for a customer to report an accident. Six months later, the claim wasn't resolving any faster. Reporting an accident is a step in the process. Getting through the claim is the outcome. That's how an AI program can look successful for months: we measure what the technology completes, rather than what changes for the customer. Stellantis could have told that kind of success story. One team built a hands-free driving system in six weeks. But the company's target is the whole job, taking vehicle development from four years to 24 months, and its technology chief says it isn't where it needs to be. You can't measure AI productivity only at the point of production. You have to look at the whole workflow. So before you celebrate a speed-up, ask whether the step got faster or the outcome did. |
|
What are you being asked to decide about AI, or where is a change stuck? Send me your question. I’ll answer the best ones in future issues, and I never use your name. |
|
I spent more than 25 years in technology, 17 of them at Microsoft, helping large organizations change how they work. Now I run Transnova.ai, helping executives turn AI and technology change into operating results. A talk for your team: Check the workflow before counting savings |
|
■ TRANSNOVA.AI
Where is your team stuck?
AI isn’t the transformation. Redesigning the work is. Transnova.ai, my firm, finds the workflow that keeps slowing your team down, redesigns it, and measures the result.
|
|
Connect on LinkedIn
|
|
