Most enterprise AI projects are built backwards. A vendor sells the platform, an executive picks a use case, and engineers spend twelve months wiring a model into a workflow nobody actually wanted. The result is a glossy pilot, a press release, and a dashboard that quietly gets archived. According to McKinsey research cited by MIT Technology Review, organizations capture less than one-third of the value expected from their digital investments — and the leaders quoted in that reporting are blunt about why: companies start with technology and bolt customers on at the end. The firms getting outsized returns from AI are doing the opposite, and the discipline behind it has a name: customer-back engineering.
Why Most Enterprise AI Programs Underdeliver
The MIT Technology Review piece featuring Capital One’s Ashish Agrawal, managing vice president of business cards and payments tech, makes a quiet but damning point: engineers in big companies often have no direct line to the customers they’re building for. They get problem statements filtered through three layers of product managers, sales notes, and stale survey data. By the time a feature ships, the original customer pain has been abstracted into something that looks reasonable on a roadmap and feels useless in the field.
This matters more in the AI era than it did in the SaaS era. AI features amplify whatever assumption you start with — a misread customer need doesn’t just produce a clunky form, it produces an automated workflow that quietly makes the wrong decision at scale. If your team is shipping an AI-integrated product without engineers ever sitting in on a support call, you’re not building software anymore — you’re industrializing your blind spots. Our take: the next wave of AI failures won’t be technical. They’ll be empathy failures dressed up in model cards.
How Capital One Forces Engineers Into the Customer’s Seat
Agrawal describes a specific, almost old-fashioned set of practices Capital One uses to keep engineers close to customers. Every engineer in his organization has an annual goal to hit several customer touchpoints, including digital empathy sessions to observe user journeys, embedded rotations through customer support, engineering ride-alongs with sales and success teams, and hackathons built around real customer problems. None of this is exotic. What’s notable is that it’s mandatory and measured.
Why it matters: engineers are problem-solvers who sit closer to systems and data than anyone else in the company. When they see a friction point in person — a customer fumbling through a five-step form, an agent re-reading the same chat thread — they tend to fix it at the layer where it actually lives, not at the UI layer where product teams usually patch it. Agrawal calls this “sideways innovation,” and argues it produces a multiplier effect because engineers approach problems from a dimension sales or product simply can’t see.
Imagine you run a mid-market fintech and your fraud team is drowning in manual review queues. The customer-back move isn’t to buy a fraud AI platform. It’s to put two engineers next to the fraud analysts for a week, watch where the queue actually slows down, and then decide whether the answer is a model, a smarter API integration, or simply a better screen. Expect the firms that institutionalize this kind of rotation to quietly outship the ones that don’t.
What Agentic AI Looks Like When the Data Layer Is Real
The Chat Concierge example in the source is worth studying because it’s not a chatbot story — it’s a data story. Capital One built a multi-agent framework that lets car buyers compare vehicles, schedule test drives, and book appointments with salespeople in a single conversation, while dealers can take over the chat through the Navigator Platform. Multiple logical agents coordinate to mimic human reasoning across the flow.
The point Agrawal hammers is that the agents are not the hard part. The hard part is the data ecosystem underneath: governed, unified data that lets the agentic loop perceive, reason, and execute reliably. “A clean data layer is what orchestrates the agentic loop,” he says, and that’s exactly the constraint most enterprises hit when they try to copy the playbook. If your customer data sits in seven systems with three definitions of “active account,” no amount of model tuning will save the agent.
For a banking or lending platform, the AI roadmap is mostly a data roadmap. Before you scope the agent, scope the master data, the lineage, and the governance. Our prediction: within eighteen months, RFPs for enterprise AI will be evaluated less on model selection and more on data readiness scoring — and vendors who can’t speak fluently about governance will lose deals they would have won in 2024.
The Agentic AI Numbers Buyers Should Actually Care About
The MIT Technology Review Insights survey cited in the source reports that 70% of leaders say their firm uses agentic AI to some degree. Roughly half of executives say agentic AI is highly capable of improving fraud detection (56%) and security (51%), with 41% citing cost and efficiency gains and 41% citing customer experience improvements. More than half of banking executives surveyed expect continued improvement in fraud detection (75%), security (64%), and customer experience (51%).
Why this matters for buyers: the expectation gap between today’s capability and tomorrow’s outcome is widest in customer experience, which jumps from 41% confidence today to 51% expected improvement. Translated for a non-technical buyer, that means the easy wins in CX from agentic AI aren’t fully here yet — and the vendors promising them today are mostly selling futures. If you’re evaluating a platform, weight the fraud and security claims more heavily than the CX claims, because that’s where current capability is strongest. Our take: the firms that quietly deploy agentic AI inside compliance, KYC, and identity verification workflows in 2026 will create the operating leverage that funds their flashier customer-facing rollouts in 2027.
FAQ
Q: What is customer-back engineering? A: It’s an approach where product and engineering teams start with a clearly defined customer experience and work backward to the technology choices needed to deliver it. Agrawal describes it as the opposite of starting with a platform capability and hunting for somewhere to apply it. The discipline forces engineers to validate the problem before committing to a stack.
Q: Why does agentic AI need stronger data governance than older AI projects? A: Because agents don’t just predict — they act. Agrawal points out that agentic systems require rigorous oversight, and that responsible AI standards and a governed data layer aren’t optional. A bad recommendation is a nuisance; a bad autonomous action is a liability event.
Q: How should a non-technical buyer evaluate enterprise AI vendors? A: Ask how the vendor’s engineers interact with end customers, how their data layer is governed, and what their “crawl, walk, run” deployment path looks like for new adopters. Agrawal recommends that approach explicitly for teams new to AI. Vendors who can’t answer those three questions are selling models, not outcomes.
Key Takeaways
- Treat your data governance roadmap as your AI roadmap; agentic systems collapse without a unified, clean data layer underneath.
- Build mandatory engineer-to-customer touchpoints into annual goals before you build your next model — the empathy gap is the real bottleneck.
- Weight vendor claims on fraud and security capability higher than CX capability today; the survey data suggests the CX wins are still maturing.
- Expect enterprise AI RFPs to shift from model-centric to data-readiness-centric scoring within the next eighteen months.
- If you’re new to AI, adopt Agrawal’s crawl-walk-run cadence and invest in cross-functional teams spanning data science, engineering, product, and design before scaling any single use case.