ISSUE 07 / OCT 26, 2026
This week in enterprise AI: what happened, why it matters, what I'd do.
The tool is the easy part
RBC, Twilio, Ramp and Amazon got their AI results from what they built and fixed around the tools.
|
■ FROM THE PEOPLE DOING IT “The business leaders themselves need to become very well versed in artificial intelligence.” Bruce Ross, group head of artificial intelligence, Royal Bank of Canada |
|
PATTERN WATCH If your rivals can buy the same tools, the tools can't be your edge. The work around them can, and it's the part that's hard to copy. |
|
|
Every company in this issue could buy the same AI tools as its rivals. The gains came from what each one built or fixed around them. RBC built its own models. Twilio cleaned its data and brand rules first, Ramp rebuilt a tool nobody used, and Amazon put the work in the hands of a much smaller team.
— Manny
|
| THIS WEEK'S THREE STORIES |
|
|
■ CASE STUDY · BANKING
RBC: 70% of its AI return comes from models it built.
What happened Bruce Ross, group head of artificial intelligence at Royal Bank of Canada, says the bank's cyber department built its own AI practice four years ago, and it has kept growing. The team uses AI and graph technology to find synthetic accounts. It also tests attack paths most people never try, because "bad actors work at machine speed with technology." RBC built some of its models and got others from industry. Ross says this has "reduced our fraud losses" significantly, across businesses, not just within one. He adds that 70% of RBC's economic return from AI comes from machine learning it built itself.
Why it matters Any bank can buy the same AI tools, so those tools alone don't set RBC apart.
WHAT I’D DO Have finance work out what share of your AI return comes from capability you own and what share comes from tools anyone can license. If nobody knows, start counting it this quarter. |
TO MAKE THIS WORK First: List every AI model in use and mark whether you built it or got it from a vendor. As you do it: Report the figures by business line, and retest models as attackers change their paths. Then: Bring the built-versus-licensed split to each quarterly budget review. |
Read the source →Bloomberg Tech Disruptors, "RBC on Turning AI Into a Competitive Advantage", Oct 20, 2026
BY ROLE Executive sponsor: Fund a standing AI team inside security, as RBC did in cyber, and judge it on the losses it prevents. Technology and data: Connect account, device and location data across business lines, so fraud models can work across businesses rather than inside one. Finance and risk: Ask for fraud-loss figures before and after each AI change, measured the same way each quarter, because "significantly" won't pass an audit. | BEHIND THE NUMBERS Ross describes the goal as deflecting bad traffic "at the atmospheric level" before it lands. The same approach helps with anti-money-laundering reporting, where demands are far higher than two years ago. The machine learning behind the 70% was built more than five years ago, and RBC's in-house model, Atom, sits under work like credit decisions. | HOW SOLID IS THIS Executive interview on Bloomberg Intelligence's Tech Disruptors podcast. The fraud reduction and the 70% figure are RBC's own claims. Ross gave no loss figures, and none of it was independently checked. |
|
|
|
■ THE VENDOR ON ITSELF · SOFTWARE
Twilio: two weeks of campaign work now takes under a day.
What happened Twilio's CMO writes in Fortune that account research and personalized content for one campaign used to take the marketing team more than two weeks. Now it takes under a day. The team runs as many as 15 account-based campaigns a quarter, up from a standard three. The CMO says the model didn't do it alone. It took getting brand standards consistent and data clean, plus governance and integration work. People stayed in the loop to judge taste. When AI localized a French campaign, it put berets on everyone, and the campaign couldn't go live.
Why it matters The CMO credits the work done before automating, not the model. The beret campaign is the kind of output that work is there to catch.
WHAT I’D DO Pick one content workflow and write down the rules a person would use to reject its output. Then run ten AI drafts against those rules and count how many fail. |
TO MAKE THIS WORK First: Before redesigning anything, have marketing and data leaders confirm that brand standards are written down and campaign data is clean. As you do it: Design the workflow with the people who do the work, and decide at each step what stays human, especially judgment on taste. Then: Run it with a few trusted marketers on real campaigns, and scale only once output passes brand review without rework. |
Read the source →Fortune, Oct 23, 2026
I wrote about this: Bolting AI onto a broken process →
BY ROLE Function heads: Write down what "on brand" means for your function, with examples of rejected work, so AI output can be checked against something. Technology and data: Check that the account data your agents will pull is clean, connected and easy to reach before anyone promises faster campaigns. Finance and risk: List which AI tools each team uses, because governance has to cover every model in use and not only the one you approved first. | BEHIND THE NUMBERS Twilio's teams have ChatGPT and Gemini for content creation, and the company is working to open Claude to the broader organization for marketing work. The CMO says the connective tissue between models is still immature. The CMO cites Deloitte, which found only 25% of organizations have moved 40% or more of their AI pilots into production, and a survey in which only 8.4% of data leaders say their data is ready for production AI. | HOW SOLID IS THIS A vendor describing its own use, in a Fortune commentary piece by Twilio's CMO. The campaign figures are Twilio's own and weren't independently checked. |
|
|
|
■ WHAT WENT WRONG · FINTECH
Ramp's first AI coding tool went unused. Its rebuild: 75% of merged PRs.
What happened In 2025, Ramp's first in-house AI coding tool was a Chrome extension that let designers prompt AI to make minor website changes. Two frontend engineers, Zach Bruggeman and Jason Quense, built it. People liked it, but it wasn't adopted. Engineers already knew how to edit a line of code, and the tool needed a local development environment that was too complicated for non-developers. In November 2025 the team rebuilt it as Inspect, a centrally configured remote environment with a coding agent on top. Today 75% of Ramp's merged pull requests come from Inspect.
Why it matters People skip a tool that does a job they already do well, or that makes them set something up first. Ramp's rebuild spread without a mandate once the setup disappeared and the tool did work engineers couldn't do on their own machines.
WHAT I’D DO Ask five people who dropped your last stalled AI pilot what got in the way. List every step they took before first use, and remove one this month. |
TO MAKE THIS WORK First: Before the pilot, map the work as people do it today, including any setup the new tool adds to their day. As you do it: Let a small group use it on real work, and fix what slows them before widening access. Then: Once use holds, widen access to the rest of the team. |
Read the source →The Pragmatic Engineer, Aug 25, 2026
BY ROLE Executive sponsor: Let people choose their tools, as Ramp does, and read low use of yours as a problem with the tool. Technology and data: Configure the environment centrally so a new user needs no local install, and test it with someone who has never set it up. Function heads: Pilot on a task people do often and find slow, not one they already finish in seconds. | BEHIND THE NUMBERS By January 2026, two months after the November launch, about 60% of pull requests at Ramp were authored by Inspect, rising to 75% by May. Inspect passed one million total sessions in July, and engineers have built more than 200 agents on top of it. Inspect is run by a team of 5.5 people, and more than 150 Ramp engineers have contributed code to it. | HOW SOLID IS THIS Company-reported in interviews with The Pragmatic Engineer by CTO Rahul Sengottuvelu, head of engineering Hamid Dadkhah and principal engineer Zach Bruggeman. The figures are Ramp's own and were not independently checked. |
|
| THE PROOF IS IN THE PUDDING |
|
15 of 84 15 of 84 companies showed a number for what AI did for their organization: 18%. Four-week average: 25%. The rest said they use AI but showed no result: no cost cut, no hours saved, no error rate. One that did: Uber's CTO runs 'Agentic Pods': AI-proficient engineers are embedded with business teams in finance, legal and HR for two-week engagements, spend the first days observing how employees work across systems, then build AI agents around those workflows — one financial pacing report fell from two days to 10 minutes. (Harvard Business Review) |
|
Amazon: six engineers, 76 days, work sized for 40 people. On The a16z Show, host Ben Horowitz cited Amazon CEO Andy Jassy's account that six engineers on the Bedrock team rebuilt its inference engine in 76 days, work that would have taken 40 people a year. Jassy called the Bedrock experience "very eye opening" and said Amazon spent time sharing the story internally to remind people what's possible.
The a16z Show, "Amazon CEO Andy Jassy on Why Amazon Is Investing $220 Billion in AI", Oct 23, 2026 Act: Exploit attempts on a decade-old PHP flaw jumped from thousands to millions AI found the flaw in PHP, the code behind most websites. Exploit attempts rose from thousands to millions after disclosure, and governments are urging site owners to patch PHP or take sites offline. Separately, hackers used Claude and Grok to automate attacks on more than 400 financial institutions in Latin America, and Axios reports that unvetted AI agents are also accessing US bank accounts. Ask your technology team this week which of your sites run PHP and whether they are patched. Also ask for a list of every AI agent that can reach accounts or customer data, and what each is allowed to do.
MIT Technology Review, Oct 22, 2026 Act: DHS bans DeepSeek in federal agencies, including tools built on it The US Secretary of Homeland Security ordered all DeepSeek products and services out of federal civilian agencies within 60 days. The ban also covers third-party tools built using DeepSeek. It does not apply to private consumers, state and local governments, or Defense Department national security systems. Ask each AI vendor which models run underneath its products, and add the answers to your AI inventory and procurement terms. Erik Avakian, technical counselor at Infotech, says organizations may be using DeepSeek indirectly through vendors without knowing it.
CIO.com, Oct 21, 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
|
|
