ISSUE 01 / SEP 14, 2026
This week in enterprise AI: what happened, why it matters, what I'd do.
Stage it in.
Then widen it.
Four companies that gave AI a small job first, and a bigger one once it had earned trust.
|
■ FROM THE PEOPLE DOING IT “Our back office opex costs are flat, and our employee engagement scores are at or near all-time highs.” Andrew Caprara, CEO, Softchoice |
|
PATTERN WATCH Softchoice and Stripe both found that agents expose weaknesses that were already there. Softchoice let its agent check people's work before it acted; Stripe had to add agent identities and load limits after a near-outage. |
|
|
The best evidence this week came from companies that staged AI in. None of them switched an agent on at full scope. They had it check people's work first, then do the easy parts, then more. That order matters more than the tool.
— Manny
|
| THIS WEEK'S THREE STORIES |
|
|
■ CASE STUDY · CONSUMER GOODS
Estée Lauder is getting M.A.C. ready for shoppers who ask AI chatbots.
What happened Estée Lauder named Brian Franz chief technology and transformation officer, combining global IT, its Shopify and Accenture partnerships and enterprise AI. All 14,000 employees get Microsoft Copilot and ChatGPT as a baseline. In the plants, an assistant called Ella helps operators fix equipment and switch production runs faster. Next month M.A.C turns on Shopify's agentic feature, so the brand can be discovered on AI chatbots, with the aim that shoppers can buy within a click or two.
Why it matters The chatbot is becoming a shop window you don't control. What it says about your product depends on product data you may not trust and content nobody has checked.
WHAT I’D DO Ask marketing to check how ChatGPT, Gemini and Claude describe your top five products. Then ask who owns the product data those answers come from. |
TO MAKE THIS WORK First: Agree who answers for what the chatbots say about you. Marketing, product and customer care each own part of it. As you do it: Trace each wrong answer back to the system it came from. Most of them will be data problems. Then: Give product data one owner and one update path, so a price or ingredient change reaches every system the same day. |
Read the source →Fortune, Sep 9, 2026
BY ROLE Executive sponsor: Do what Estée Lauder did: put IT, transformation and AI under one leader with a revenue target, so AI isn't run as an IT project. Technology and data: Fix product data first: names, ingredients, prices, stock. Agents read it literally, and a mismatch between systems becomes a wrong answer at the point of sale. Champions: Name one AI champion per brand or business unit and give them real time for it. A list of names isn't a community. | BEHIND THE NUMBERS AI champions sit in each of its 20+ brands, and younger staff reverse-mentor senior leaders. Its Formula Navigator tool raised product-development efficiency 20%. In an EY-Parthenon survey cited in the same article, only 33% of consumer and retail leaders trust AI to make growth decisions. | HOW SOLID IS THIS Company-reported to Fortune. The 20% efficiency figure is Estée Lauder's own and wasn't independently measured. |
|
|
|
■ CASE STUDY · IT SERVICES
25–30% more volume. No new back-office staff.
What happened Softchoice put an AI agent on a process that carries about 30% of its business. It went in two steps: first the agent checked people's work, then it did the easier parts while people reviewed. Volumes rose 25–30% with no added capacity. CEO Andrew Caprara says back-office costs are flat after about two and a half years of AI work, and staff engagement is near an all-time high.
Why it matters This is the scoreboard that counts: volume up, back-office cost flat. It came from one high-friction, high-volume process, not from handing out licenses.
WHAT I’D DO Pick your highest-friction, highest-volume process. Put an agent on it that checks the work first, and widen its job only when the team trusts it. |
TO MAKE THIS WORK First: Decide which steps the agent takes, which stay with people, and who signs off on its work. As you do it: Write down the test for widening its job before you start: an error rate you'll accept, and the team's own view of whether the work got easier. Then: When the job widens, give people short guidance on what to check and when to overrule the agent. |
Read the source →AI Proving Ground Podcast, "AI Moves Weekly, Your Company Doesn't", Sep 9, 2026
BY ROLE Executive sponsor: Review AI work weekly, as Caprara does. Quarterly reviews let weak pilots live for months. Finance and risk: Track cost per transaction for that process against volume, not AI license spend. People managers: Bring team leads into the design. Softchoice widened the agent's job once the team said the work was easier. | BEHIND THE NUMBERS Small "flash teams" (an AI-native engineer, a data specialist, a process person and the frontline owner) build working prototypes in days. Caprara moved the CEO's AI reviews from quarterly to weekly and recommends giving each prototype about a week before deciding whether to pursue it. He admits the first year was a lot of motion that didn't add up. | HOW SOLID IS THIS Reported by Softchoice's CEO on a podcast run by World Wide Technology. Not independently measured. |
|
|
|
■ CASE STUDY · PAYMENTS
10,000 Stripe employees use its own AI agent every week.
What happened Stripe's internal AI agent, Kai, is used by more than 10,000 employees every week. Staff have packaged about 2,000 reusable “skills”, around 50 of them used daily across the company, and Kai picks the right one from what a person describes. Each session runs in its own isolated cloud sandbox, and non-engineers use it to build their own reports without filing a ticket. When agents nearly took down production, better prompts didn't fix it: Stripe added agent-specific identities, priority signals and load shedding.
Why it matters Stripe learned that agents don't create new kinds of problems; they make existing weaknesses bigger and faster. Most companies can't yet answer the basic question: if an agent acted wrongly last week, could we prove what happened and who owned it?
WHAT I’D DO Ask your platform lead whether each of your agents has its own identity, or is acting under the account of the person who started it. If it's the second, that's your first fix. |
TO MAKE THIS WORK First: Log every action an agent takes, so you can replay last week if you need to. As you do it: Keep a list of every agent, its owner and what it can touch. Put the most checks on agents that touch money or customer data. Then: Retest agents whenever the model or the systems under them change. What was safe last quarter may not be now. |
Read the source →How I AI (Claire Vo), "How Stripe Built Kai: Data Dashboards, Reusable Skills, and Enterprise AI Governance", Sep 7, 2026
BY ROLE Technology and data: Give every agent its own identity and limits on how hard it can hit core systems, before you scale. Stripe added both after agents nearly took down production. Transformation lead: Track which skills people actually reuse and retire the rest. At Stripe, about 50 of 2,000 are used daily. Champions: The people whose skills others reuse are your real champions. Find them in the usage data, not the org chart. | BEHIND THE NUMBERS Stripe tracks which skills are used heavily, suggests improvements to their authors and retires the rest. Its view: every extra skill adds noise to retrieval, so quantity becomes a quality problem. Kai stops and asks for explicit permission before using a tool such as a person's calendar. | HOW SOLID IS THIS A Stripe engineering manager describing his own system on a podcast. Usage figures are Stripe's own. |
|
|
A one-week pause took daily AI use from 10% to over 50%. Looking back on 2023, when GPT-4 launched, Zapier CEO Wade Foster says one thing worked better than any memo or AI committee: pausing the whole company for a week-long hackathon. It moved daily AI use from about 10% of staff to over 50%.
Pioneers of AI podcast, "Zapier wants to be the AI hub where all your agents connect", Sep 9, 2026 Act: The EU AI Act already has an incident-reporting clock. Article 73 took effect Aug 2. Providers of high-risk AI must report serious incidents within 15 days, 10 if someone died, and 2 for widespread or critical-infrastructure incidents. Draft Commission guidance says an indirect link, such as a loan denial traced to a flawed AI credit score, can be enough. Name who decides when the clock starts.
CIO.com, Sep 8, 2026 PG&E's predictive grid model has avoided 19 million outage minutes. The utility combines data from 5.5 million smart meters and grid sensors to predict and prevent faults, which CIO Ajay Waghray says has also prevented 28 potential ignitions. More than 600 AI cameras tell smoke from cloud, and an AI assistant searches the millions of pages of operating documentation at the Diablo Canyon nuclear plant.
Technovation with Peter High, "PG&E CIO Ajay Waghray on Building Trusted Data for Predictive AI", Sep 10, 2026 |
|
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. |
|
■ 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.
|
|
Invite me to speak · Connect on LinkedIn
|
|
