It’s been nearly a year since we launched Leaders of Code, a section on the Stack Overflow Podcast the place we curate candid, illuminating, and (dare we are saying) inspiring conversations between senior engineering leaders.
A powerful roster of visitors from organizations like Google, Cloudflare, GitLab, JPMorgan Chase, Morgan Stanley, and extra joined members of our senior management group to match notes on how they construct high-performing groups, how they’re leveraging AI and different quickly rising tech, and the way they drive innovation of their engineering organizations.
To kick off 2026, we needed to gather some overarching classes and customary themes that lots of our visitors touched on final yr, from the significance of high-quality coaching knowledge to why so many AI initiatives fizzle to what the belief/adoption hole tells us and tips on how to bridge it.
Learn on for a very powerful insights we heard final yr.
Poor knowledge high quality undermines even essentially the most refined AI initiatives. That was a unifying theme of our present all through 2025, starting with the inaugural Leaders of Code episode. In that dialog, Stack Overflow CEO Prashanth Chandrasekar and Don Woodlock, Head of International Healthcare Options at InterSystems, explored how and why a sturdy knowledge technique helps organizations notice profitable AI tasks.
An out-of-tune guitar is an apt metaphor right here: Regardless of how expert the musician (or superior the AI mannequin), if the instrument itself is damaged or out of tune, the output can be inherently flawed.
Organizations speeding to implement AI usually uncover that their knowledge infrastructure is fragmented throughout siloed programs, inconsistent by way of format, and devoid of correct governance. These points stop AI instruments from delivering significant enterprise worth and proving their worth to skeptical builders.
Within the episode, Prashanth and Don emphasised that sustaining a human-centric strategy when automating processes with AI requires constructing belief amongst customers, which, in flip, begins with clear, well-organized knowledge that AI programs can reliably interpret and successfully use.
Too many organizations rush into AI implementation with out correctly assessing whether or not their data infrastructure can assist it, defined Ram Rai, VP of Platform Engineering at JPMorgan Chase. This overconfidence stems from a basic misunderstanding: Having knowledge isn’t the identical as having AI-ready knowledge. A centralized, well-maintained data base is crucial for getting AI initiatives off the bottom efficiently, but most organizations uncover this requirement solely after launching poorly conceived pilot tasks.
Organizations usually fail to guage whether or not their AI tasks align with core enterprise values. This could result in wasted investments in instruments that can’t entry the inner context essential for significant outcomes. In extremely regulated environments with heavy compliance necessities like banking and finance, Ram says his group can’t ignore the productiveness advantages provided by AI. On the identical time, he says, they need to “be surgical about it,” significantly when coping with important infrastructure the place “we will not solely belief probabilistic AI.”
Enterprise AI fashions often hallucinate as a result of they lack entry to inside firm data, as Ram factors out: “Why does AI hallucinate? As a result of it lacks the proper context, particularly your inside context. AI does not know your IDP configuration, token lifetimes, your authentication patterns or your load stability settings, so the coaching knowledge is skinny on this proprietary data.”
This hole between basic coaching knowledge and particular organizational data leads AI instruments to make convincing-sounding however essentially incorrect options. Grounding AI instruments in verified, inside documentation considerably improves accuracy and reliability, serving to enterprise customers notice the worth they want from these new instruments.
The conversation with Ram highlighted how Stack Overflow’s structured Q&An information supplies excellent fine-tuning materials for next-generation AI fashions by providing the form of community-driven, verified data that may bridge this context hole. Organizations that put money into strong inside data programs create a basis for AI instruments that builders can truly belief.
To study extra about how Stack Inner can assist you construct smarter, extra reliable AI programs, try this webinar.
Stack Overflow’s 2025 Developer Survey revealed a hanging paradox: extra builders actively mistrust the accuracy of AI instruments (46%) than belief it (33%), whereas solely a tiny fraction (3%) report “extremely trusting” the output.
This belief deficit has actual penalties for adoption and productiveness. The number-one frustration, cited by 66% of builders, is coping with “AI options which are virtually proper, however not fairly,” which frequently leads on to the second-biggest frustration: “Debugging AI-generated code.” Many builders discover themselves losing time reviewing and fixing AI-generated code slightly than experiencing the promised productiveness beneficial properties.
Skilled builders are essentially the most skeptical of AI, with the bottom “extremely belief” fee (2.6%) and the very best “extremely mistrust” fee (20%). As Ram Rai of JPMorgan Chase acknowledged, “Many builders mistrust AI accuracy—that’s the present actuality, and there’s a battle with adoption of AI.”
This decline in belief—down from over 70% constructive sentiment in 2023 and 2024 to only 60% in 2025—is a crimson flag. Organizations should handle builders’ legitimate accuracy and reliability issues earlier than anticipating widespread adoption and the belief of precise enterprise worth.
Builders flip to Stack Overflow for human-verified, trusted data, with about 35% reporting that their visits to Stack Overflow are a results of AI-related points a minimum of a number of the time. This sample reveals an important perception: when AI instruments fail or produce suspicious outcomes, builders search validation from community-driven platforms the place actual people have vetted the solutions by collective scrutiny. By “grounding AI in our inside actuality utilizing [a] strong group data system like Stack Overflow,” says JPMorgan Chase’s Ram Rai, his group can transfer past purely probabilistic AI towards programs that incorporate verified, battle-tested data.
As we talked about above, the structured nature of group Q&A—with voting, peer assessment, and iterative refinement—supplies precisely the form of high-quality coaching knowledge that AI fashions must generate reliable outputs. Organizations that construct or entry community-driven data layers present their AI instruments the verified context they should transfer from “virtually proper” to constantly dependable.
Organizations want to acknowledge what AI can and can’t do effectively. That was the large takeaway from our conversation with Dan Shiebler, Head of Machine Studying at Irregular AI.
Leaders who handle expectations and deploy AI strategically—the place it supplies real worth slightly than the place it is merely stylish—see higher outcomes. Understanding limitations means acknowledging that AI excels at sample matching and producing code for well-defined issues however struggles with novel architectural choices, advanced trade-offs, and conditions requiring deep contextual judgment.
Essentially the most profitable AI implementations rigorously scope the place AI can add worth whereas sustaining human oversight for choices that require accountability, area experience, or artistic problem-solving that goes past current patterns.
Within the two–part dialog between Peter O’Connor, Stack Overflow’s Director of Platform Engineering, and Ryan J. Salva, Senior Director of Product at Google, Developer Experiences, we explored how AI is remodeling group buildings. From enabling engineering groups to function successfully with only a handful of individuals to decreasing collaboration overhead and accelerating decision-making, there’s no denying the fact that AI is reshaping how growth groups work.
As AI automates routine duties like boilerplate code technology, bug triage, and primary testing, the function of builders is shifting towards structure, important judgment, and cross-functional collaboration.
This transformation does not eradicate the necessity for builders; as a substitute, it elevates the abilities that matter most. The 2025 Developer Survey added a brand new function, “architect,” now the fourth hottest function for respondents. That change displays how the business is recognizing the rising significance of systems-level considering, design choices, and integration work. With the good thing about their human expertise, senior builders will more and more deal with technique, mentorship, and guaranteeing that AI-augmented groups keep high quality and reliability requirements as they decide up much more momentum
Abhinav Asthana, CEO and cofounder of Postman, defined how APIs are the key to enabling LLMs to perform as true brokers by connecting them to dwell knowledge and workflows.
Nicely-designed APIs allow AI brokers to work together with programs successfully, remodeling AI from purely conversational instruments into action-oriented programs able to executing real-world duties. Within the episode, Abhinav shared how Postman makes use of AI brokers to mixture and summarize developer suggestions, offering organizational readability, whereas additionally detailing how the corporate scaled from simply three founders to over 400 individuals.
The important thing lesson from all that? Organizations should prioritize API high quality, documentation, and developer expertise to attain widespread adoption of AI instruments. Postman’s 2025 State of the API report discovered that 89% of builders use generative AI of their day by day work, but solely 24% actively design APIs with AI brokers in thoughts.
This mismatch creates a important hole: AI brokers require exact, machine-readable indicators—specific schemas, typed errors, and clear behavioral guidelines—but most APIs are nonetheless designed primarily for human consumption.
The report made the robust case that “APIs have to be designed with AI brokers in thoughts” as a result of “APIs designed with machine-readable schemas, predictable patterns, and complete documentation will combine sooner and extra reliably than these constructed just for human consumption.” Organizations that put money into API-first growth practices, treating APIs as merchandise with correct governance, versioning, and documentation, subsequently place themselves to capitalize on the AI agent revolution whereas opponents battle with integration challenges.
