WSJ Events Sep 18, 2026
With Sastry Durvasula, COO of TIAA · Janelle Sallenave, COO of Chime
Janelle Sallenave said the standard operators should now hold themselves to is a fall of ten percentage points a year in the cost of serving a customer, with the customer's experience improving at the same time.
Cost and quality have been treated as a trade-off for as long as operations has been a job. Her claim is that the trade-off has stopped applying, and that leaders who still price it in are working to an old standard.
"We're no longer at a point where those two things have historically been in conflict."
Sallenave scaled Uber Eats before joining Chime, where she runs operations; Sastry Durvasula runs operations at TIAA, a firm managing more than $1.5 trillion in its asset management business alone, and has spent his career at American Express, Marsh McLennan, McKinsey and TIAA. Both hold the chief operating officer title, which is the seat this work now sits in.
The full interview is covered here so you can skip it.
Here are the 12 lessons that matter.
Key Takeaways
Durvasula's term for what blocks a large firm is enterprise complexity debt — technical, data, process, skill and product debt accumulated over decades
The analogy he reaches for is electricity, not the internet or mobile
Sallenave's objection to that analogy: electricity was one before-and-after, while this changes every quarter
Chime's first attempt at AI optimized inside functional silos, which made each team faster without making the company faster
The fix is an AI factory, where each function becomes a production line and AI builds the connective tissue between them
Judgment, taste and accountability for outcomes are what Sallenave now hires and organizes for, rather than mastery of the task
A missing space after a full stop took Chime's agent 25 minutes and 15 Slack messages to fix
TIAA printed every screen a call-center rep must read and stitched them together: 81 inches of interface
Close to 200 agents go into TIAA production this year, each one graded on a four-stage autonomy framework
Durvasula expects the orchestrator, not the agent, to be the valuable layer every firm ends up needing
1. Enterprise Complexity Debt
Asked why this transformation is harder than the ones before it, Durvasula started with the kind of company he has always worked for.
He has never worked anywhere young
for the record, I've only worked for 100-year-old companies for my whole career.
Sastry Durvasula
The host picked up a phrase from their conversation the previous week, and Durvasula ran with it.
Tech debt is only one of the debts
So that's why I talk about enterprise complexity debt — be it the technical debt, data debt, process debt, skill debt, product debt, over the course of years, acquisitions that companies have done — now we're going to be forced to do rewiring at a different level, and that's what makes it really exciting.
Sastry Durvasula
He has seen the internet, COVID, social media and the mobile phone go through firms including American Express, Marsh McLennan, McKinsey (which he noted is marking its 100th year) and now TIAA. His claim is that none of them reached the same layers.
What is different about the rewiring
it's getting into the layers of the firm that we need to rewire, that we never had a chance to rewire from a cognitive-level point of view.
Sastry Durvasula
2. The Electricity Analogy
The comparison Durvasula uses is electrification, and he is reading the history rather than reaching for the phrase. The book is "Electrifying America" by David Nye, on the social effects of electricity's arrival, and he said the period may have to be relived.
He also described the forces around AI as pointing both ways at once: data centers are good and bad, China is good and bad, and security cuts in two directions, because the same technology defends you and attacks you.
The job is interesting for exactly that reason
So it's quite an interesting time to live, and — the best part of the job.
Sastry Durvasula
3. Inside The Washing Machine
Sallenave accepted the electricity comparison and then said what she thinks it misses.
Electrification happened once
The electricity example, to me, is very pre-and-post — there wasn't electricity, and now there is.
Janelle Sallenave
This one resets every quarter
And the thing that I have found with this new world that we live in is every quarter there's something so fundamentally different — there's something that we understand that last quarter we didn't, about how we could or would or will or won't use AI to transform the way that we work.
Janelle Sallenave
Her image for it was a washing machine, tumbling while you look for something to hold. What makes it harder than earlier cycles, she said, is that the rate of change itself has to be sustained.
The leadership question is how long you can hold it
It's like running the fastest speed you can and saying to yourself, how long can I hold this?
Janelle Sallenave
Changing how the company works, how it organizes and parts of its culture all at once is what she called the hardest and most interesting part of the past year.
4. Chime Has No Sacred Cows
Before describing the framework, Sallenave named her advantage.
A 14-year-old company starts clean
I do want to say a caveat that we start from a really lucky place — this company is 14 years old, which means we don't have any sacred cows. We don't have tech debt. It's such a gift.
Janelle Sallenave
Durvasula offered to swap seats. The underlying point she drew from it was that the method that got the company here will not be the method that takes it forward.
5. Right But Wrong
Chime's first move, about three years ago, was the obvious one: customer support, chat bots and declarative bots, under a rule of doing no harm to the customer experience. A year of that taught the company what had to be true internally about how data is organized and how governance works.
What the first year actually established
gave us confidence that, one, we didn't totally know what we were doing; two, nobody seemed to know what they were doing, and we could figure it out together.
Janelle Sallenave
The next move was to ask every function, finance and customer support among them, to find its own applications. Sallenave's assessment of that is the phrase she applies to most of it.
The pattern she keeps hitting
we realized that was right but wrong — which is how frankly every AI thing we do is right but kind of also wrong
Janelle Sallenave
Why functional gains did not add up
we realized that we were optimizing inside of functional silos, that we were finding ways to use AI to operate more efficiently, to do the task more efficiently, but it wasn't helping the company be faster, it wasn't helping the company build better, and we realized we were missing, in essence, the connective tissue.
Janelle Sallenave
6. The AI Factory
The framework that followed treats each function as a production line and puts AI between the lines rather than inside them. Sallenave was candid that the language came from elsewhere.
The objective is not doing tasks faster
It is the idea of recognizing that the objective isn't using AI to do tasks — it's actually to fundamentally reimagine how we build, how we work.
Janelle Sallenave
What that costs the humans is the thing they were good at. The quality of a product requirements document, a technical design document or a marketing brief stops being the measure of the person who wrote it, because those artifacts are now inputs the factory connects.
What she organizes for instead
It is about helping our teams think about how we organize and lean into talent — that's really around judgment, taste, the accountability for the outcomes, and not the mastery of actually doing the task.
Janelle Sallenave
7. Archie And The Missing Space
The factory has a name, Archimedes, shortened to Archie, and Chime routes employee-reported bugs into it. An employee posts in a Slack channel that something in the app looks wrong, and the system works the ticket with oversight. Sallenave said small items are the point, because they show what the thing can actually do.
Her example was from the previous Friday. Someone reported that on one page of the app, a space was missing after a full stop between two sentences.
A ten-second fix, done the long way
us humans could have solved that problem in ten seconds — you found the space in the code, in the app, you'd find the place, you'd put your cursor, you'd add a space bar, we'd move on with our lives.
Janelle Sallenave
Instead it ran to 15 Slack messages and about 25 minutes, after which she had the system write its own retrospective. She said she got a good laugh out of it, and treats the episode as information about controls and organization rather than as a failure.
How she describes the phase
it's like falling forward into success.
Janelle Sallenave
8. Rewiring TIAA's Workflows
Asked about cross-functional goals, and about the risk that five teams aligned on day one revert to their own priorities by day two, Durvasula described a transformation running on three axes at a firm with retirement, wealth management and asset management businesses, the last managing more than $1.5 trillion.
The first axis is the workforce of the future, asked function by function: what a marketer does, what a financial analyst does, what an investment analyst in asset management does, what a developer does. The second is the workloads inside those jobs, which he said are becoming clearer and more prescriptive as the technology changes.
The third axis is where the company actually changes
But then you have the workflows of the future — that's where you're actually rewiring the company.
Sastry Durvasula
His example is record keeping for the retirement business, a workflow nearly every business and function inside TIAA contributes to. Rewiring it means a developer has to imagine their own job with AI and contribute to the record-keeping redesign at the same time.
The other example was the contact center that reports to him. To see what a representative faces when a retired teacher or healthcare worker calls with a question, the team printed every screen the rep has to read and stitched them together.
The interface, measured
it's 81 inches, it's as tall as LeBron James
Sastry Durvasula
Every one of those screens exists because a different unit contributes a different artifact, and the rep has to read across all of them to answer one question.
How the goals are set and governed
we're going to have to bring all these teams together and set goals at the enterprise level.
Sastry Durvasula
TIAA has prioritized around 100 workflow-of-the-future use cases, governed by an enterprise steering committee that includes Durvasula and the rest of the C-suite. Agent deployments run through that same workflow-rewiring process rather than alongside it.
9. The Empathy Agent
One deployment surprised him. A representative handling a complaint is trying to resolve it and close the transaction, which leaves no room to compose a careful reply.
He did not expect this to be the use case
I didn't think that it would be the best use of AI, but honestly — a complaint comes in from a participant, and now you're dealing with the complaint.
Sastry Durvasula
The agent works inside the call and drafts empathetic wording the rep can accept, the way autofill suggests the rest of an email, rather than making them switch to Outlook and start again.
10. Experience, Not Efficiency
Pressed on whether the standard moves past efficiency, Durvasula said the word he would use is experience, and gave the case that makes it concrete.
Who is on the other end of the call
when you have an 82-year-old calling, who is experiencing cognitive decline, and they're calling our rep, the responsibility is dramatically different.
Sastry Durvasula
Sallenave interrupted to push the point further, and this is where the number in the headline comes from.
The trade-off has stopped applying
We're no longer at a point where those two things have historically been in conflict.
Janelle Sallenave
So she proposes a standard
And so I would challenge all of us to say our expectation should be something like a 10-percentage-point decrease in our cost to serve every year, while the quality of what our customers experience goes up.
Janelle Sallenave
She said that is starting to become the new standard rather than an ambition, and that it changes how operations is organized from end to end.
11. What The Startups Teach
Asked what large incumbents can learn from AI-native companies two or three years old that have grown faster on leaner headcount, Sallenave began by lowering the pedestal: those companies are making it up as they go too. Then she named two things worth copying.
The first is structure.
Flatter, and deliberately so
there's just a flattening in how those companies have built, in terms of org structure
Janelle Sallenave
Chime calls its version the builder model, championed by its chief technology officer, and Sallenave said the work structure has to flatten as the factory lines go in.
The second is culture, and she framed it as the part operators neglect.
The change lands unevenly on the staff
But for other people, this is a massive amount of change coming at them way faster than they're comfortable with.
Janelle Sallenave
For a data scientist this is the best time to be doing the job, she said; for others it is not, and leaders focused on the technology skip the culture work that determines whether any of the return arrives.
Durvasula agreed and then turned it around. A firm that has operated for a century in a regulated industry has something the young companies lack.
The learning runs both ways
at some point these companies will have to aspire to that level of maturity and impact
Sastry Durvasula
He expects more regulation and more scrutiny of risk and compliance, and said incumbents have already survived every previous technology cycle's version of that. TIAA's venture arm is how he gets both directions at once: the firm invests in these companies, learns from them, and gives them access to the enterprise. Sallenave added that the chance to shape how those companies build their roadmaps is one of the rewards of engaging with the period rather than waiting it out.
12. Four Stages Of Autonomy
An audience member asked where TIAA draws the line on how much autonomy an agent gets. Durvasula's answer is a framework with four stages: AI-assisted, AI-augmented, AI-adaptive and AI-autonomous. The assisted end is a colleague-facing tool he called MyGate, a model-agnostic generative and agentic tool set; the autonomous end is a true agent.
Why every agent has to be graded
our belief system is that every agent has to go through this framework, from a responsible AI and governance point of view, because we are making some heavy-duty financial decisions
Sastry Durvasula
The volume going into production
this year alone we are implementing close to 200 agents in our production environment across this framework.
Sastry Durvasula
Most of them sit in the first three stages. He said TIAA has nothing at the level of a self-driving system yet, and named one deployment at stage four, carrying more governance than he expects will eventually be necessary, with a human in the loop at what he called a higher dosage as the stages climb.
The architectural claim he closed on is about which layer matters.
The orchestrator, not the agents
I also believe that the future belongs to the orchestrator more than the agents.
Sastry Durvasula
He described three layers: a system of action operated through the orchestrator, a system of intelligence, and a system of record, which is everything a firm already runs on. Build, buy and partner are all live for the orchestrator, and he expects specialist firms to supply much of what sits underneath it.
Bonus Insights
Agents get reviewed like employees
An earlier panel had suggested performance-managing agents the way companies manage staff, and the audience asked whether either of them does it. The answer from the stage was that voice bots and chat bots are already the entry point into customer support and are "doing 70, 72% of all of our interactions", and that every one of those interactions runs through the same quality-assurance process used on human agents, because the panel treats the two as no different.
Asked whether any agent had been put on a performance improvement plan, the reply was that they are constantly on one: it is never good enough, which is the nature of iterative improvement.
The host's own worry
The host raised a failure mode from an earlier session on stage — that a cross-functional group aligned on day one can then run fast in the wrong direction, with each team reverting to its own norms by day two. That framing is what produced Durvasula's account of enterprise-level goal setting and the steering committee above it.
The host also pressed on governance directly, noting that Durvasula had used the word only once, and that trust and regulatory compliance become the binding constraint as deployments move up the autonomy stages.
Their shared bottom line is that the constraint is organizational rather than technical: the returns show up when the workflow is redesigned across functions, not when each function automates its own tasks.
Products, Companies & Tools Mentioned
TIAA (Retirement, wealth management and asset management, the last running more than $1.5 trillion; around 100 workflow use cases prioritized and close to 200 agents going into production this year)
Chime (14 years old and, on Sallenave's account, carrying no tech debt; its AI factory is named Archimedes and takes employee-reported bugs from Slack)
Slack (Where Chime employees report app bugs into the factory, and where the 25-minute exchange over a missing space played out)
Uber Eats (The business Sallenave scaled before Chime, cited as the pre-AI comparison for a fast-growing operation)
American Express and Marsh McLennan (Two of the century-old firms Durvasula worked at before TIAA, alongside McKinsey)
McKinsey (Where he worked before TIAA, and which he noted is marking its 100th year)
Waymo (His reference point for full autonomy, and what TIAA does not yet have an equivalent of)
Books & Resources Mentioned
Electrifying America – David Nye (The history of electricity's social effects that Durvasula is reading, and the basis for his claim that AI is the closer analogy to electrification than to the internet)
Watch the full episode:
More Sastry Durvasula interviews and podcast appearances
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: