Never breaking: keeping AI away from the code itself
OutSystems has AI edit the abstract model of an application rather than the code itself, then deterministically generates code; each new application automatically inherits permission controls and compliance settings, making it far less prone to failure than raw "vibe coding."
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
AI edits the abstract model, not the code itself
Most AI code-generation platforms generate production code directly, so vibe coding often breaks. OutSystems does not let Claude Code, Codex, or its own Mentor touch the code—they operate on an abstract layer representing application intent, then deterministically generate code with one button press. Each generated code asset automatically carries permission controls and compliance requirements (GDPR, HIPAA), and reuses existing compliant components, avoiding a fresh compliance review every cycle. He calls this the platform's "secret weapon."
— Woodson MartinTrust in regulated industries is a moat startups cannot buy
Nathan points out a common dynamic: native AI startups find it harder to replicate the foundational capabilities of an incumbent than the incumbent does to add an AI layer on top. Woodson adds a second reason—for regulated enterprises, trust is never a pure technology assessment. It is the sum of this platform's delivery history on similar systems over many years, whether other customers will vouch for it. This accumulated trust, no matter how cool the technology, is hard for startups to earn in a short window.
— Woodson MartinEven a PDF-reading task can get stuck in compliance queues
The fastest-growing companies like Anthropic and OpenAI do not meet traditional uptime standards, yet enterprises still buy them—which blurs what "enterprise-ready" means. Woodson gives a concrete scenario: a customer has built and tested an agentic system but now sits in the compliance review queue—not waiting for the system itself, but for the underlying AI model. This includes verifying that the model's training data was legally obtained. Even if the work is just converting PDF content into structured data, it cannot bypass this review gate.
— Woodson MartinMajor features per quarter jumped from 4 to 26; token spending has peaked
A year ago, OutSystems' strategy was to migrate everything to AI—major features per quarter jumped from 4 in last Q4 to 19 in Q1 this year, then 26 in Q2, with token consumption spiking accordingly. But spending peaked in June and July; since then, two things have kept it below Q3 forecast: a custom harness for his engineering team, and an LLM router that routes routine tasks to cheaper models. His conclusion: most enterprise workloads don't need cutting-edge models, and some workloads don't need models at all.
— Woodson MartinAsian bank apps and Rotterdam fuel terminals run on this system
Asked how non-technical people can build critical systems safely, Woodson gives two concrete examples: almost any Asian bank's mobile app likely runs on OutSystems; so does the system for a Rotterdam fuel terminal that ensures diesel never enters a pipeline previously carrying kerosene, preventing pollution of millions of euros' worth of product. This, he says, is the accumulated result of twenty-five years solving a single problem: letting non-specialists safely build complex systems.
— Woodson MartinA six-year legacy system overhaul was compressed to six months
Legacy systems—COBOL, AS/400, Lotus Notes—that enterprises once feared to touch are now finally being tackled, because AI can accelerate understanding the old logic, translating it to new requirements, and generating and testing each step of the new architecture. An insurance company compressed a planned six-year overhaul of a legacy case-management system into six months. Meanwhile, staff can now turn spreadsheets into dashboards talking directly to Claude, no longer queuing for requests—the new bottleneck is standardizing and integrating these self-built tools back into the backend.
— Woodson MartinAll competitors' billboards are saying the same five things
Nathan observes that AI platforms are converging into the same horizontal "do everything" product; Woodson agrees. He notes that along Highway 101 in San Francisco, billboards change logos but repeat the same pitch—because nearly all experience is now conversational, and baseline components are similar across the board. Future differentiation, he argues, will come from specialization: model tuning and distillation, cost-structure optimization for specific verticals, and deep domain knowledge accumulated over years in regulated industries—banking, insurance, healthcare, energy—where the real separation will happen.
— Woodson MartinOptimistic on entry-level talent, but enterprises haven't learned how to deploy them
Asked about career prospects for junior talent entering the field, Woodson says he is very optimistic—this cohort has been AI-native for the past five years and brings that instinct to organizations. He acknowledges, though, that most enterprises lack infrastructure for these people to truly master domain knowledge. What he is doing is transforming onboarding from a few days of training into real-time, on-demand knowledge delivered by agents.
— Woodson MartinIn their own words · checked verbatim
When AI can build all this stuff for us, the pace at which we build, the sheer volume of stuff we're managing, like, never breaks feature turns out to be even more important.
Woodson Martin8:13
the reputation that you build up over time for that kind of trust for these kind of hardened systems in regulated industries is a thing that is just very hard to get as a startup regardless of how cool your technology is.
Woodson Martin10:49
There is no finish line on security and cyber threat.
Woodson Martin25:58
I think we shipped four major features in q four last year, 19 in q one, and 26 in q two this year.
Woodson Martin27:32
Incredibly complex logic for ensuring that when you pump diesel through this pump in the terminal, you're not putting it through a pipe that last had kerosene in it.
Woodson Martin39:25
If we drive the 101 Freeway in San Francisco and you see all the billboards and they all say the exact same five words on them.
Woodson Martin1:00:16
I'm super bullish on junior talent, on people who grew up whose last five years have basically been native to all these new technologies
Woodson Martin1:03:47
Figures
| OutSystems major features launched per quarter | 4 in prior Q4; 19 in Q1; 26 in Q2 | 27:32 |
| OutSystems founded | 2001 | 4:11 |
| Woodson Martin's tenure at Salesforce | 18 years | 4:11 |
| Mentor project began | 2018 | 51:34 |
| OutSystems European customer base | approximately half | 31:32 |
Glossary
- vibe coding
- Writing code by having AI generate it directly, with developers exerting minimal control over architectural details.
- MCP (Model Context Protocol)
- A standard protocol that lets any coding agent uniformly invoke external tools and services.
- LLM router
- Middleware that distributes requests to models of different cost tiers and capability levels based on the task.
- agent foundry
- A tool that analyzes customer data in existing systems, then automatically recommends and can generate new systems suited to agent automation.
How to listen
Enterprise software decision-makers, technology leaders responsible for AI governance and compliance, and investors concerned with AI's impact on software-company cost structures.
19:48–22:43 and 35:38–38:34 are sponsor reads and can be skipped.