OutSystems Went From 4 Major Features a Quarter to 26. Token Spend Fell Anyway

Woodson Martin became CEO of OutSystems in 2025, succeeding Paulo Rosado, who had run the company for 24 years since founding it in 2001. Nathan Labenz opens with the awkwardness of a software leadership handoff in the AI era: does the incoming boss get it, and does the company? Martin starts with Rosado. Custom enterprise software at the time was years late, wildly over budget, and so rigid that changing it meant another huge project. Rosado's counter-vision was simple and, Martin says, still the core of the business: build faster, reliably every time, and evolve software at the pace of the business without breaking anything.

What got Martin excited 16 months earlier, when he joined, were the platform primitives. Relatively unskilled technical users could build complex enterprise systems, change them fast, and not break them. That was revolutionary then. It is more necessary now, because AI can generate so much more, and the volume of stuff under management makes "never breaks" even more important. Customers and users actually love the product. The remaining unsolved problems, he thought, were ones he knew how to attack after 18 years at Salesforce, spanning European marketing, the core sales platform, and a stint running HR. He wanted to turn on the afterburners.

Labenz notes that the original vision sounds almost custom-built for the AI moment, except for the never-breaks clause. Output is up tremendously, and things also break pretty often. How do you keep that promise in the vibe-coding era? Martin's answer is that OutSystems does not do AI software development the way most tools on the market do.

There is an intermediate layer: an abstraction of the application's intent. You can bring Claude Code, Codex, or OutSystems' own assistant, Mentor. Those agents manipulate the model of what you need, not the code itself. You press a button, or you ask Claude, and the platform deterministically generates an asset that already meets security and role-based access requirements, because those controls are baked in at the platform level. Every code asset created on OutSystems respects 100% of that control plane, and it automatically reuses work you already did.

The contrast is the blank-page Claude app. In a complex enterprise with hardened systems that already satisfy GDPR or HIPAA, you do not want the model reinventing those pieces and sending you through another compliance cycle. You pick up what works, add the new capability on top, and push to production knowing it works the first time. That inheritance of the enterprise control plane, Martin says, is a very different story from almost any other harness people are using to build with AI.

Labenz has watched the pattern for years: people assume AI-native startups will unseat old platforms, and more often the incumbent with well-developed primitives puts an AI layer on top and reconfigures those primitives. Martin agrees the foundations are hard to rebuild at enterprise scale. Software is only half of it. The other half is trust. A regulated European firm with hundreds of regulators, contractual obligations, and privacy exposure does not evaluate a platform as a pure technical bake-off. It evaluates a history of delivering systems like the one it needs, plus other customers it can call. That reputation is extremely hard for a startup to get, however cool the technology.

Labenz asks what enterprise-ready even means now. Anthropic and OpenAI are among the fastest-growing companies in the world and do not have conventionally great uptime, which suggests enterprises will forgive old sacraments if the product has "the spice." Martin says every enterprise is a little different, and so is every workload. Plenty of regulated firms let AI helpers do work. Very few have unleashed autonomous agents on mission-critical processes. That is partly industry maturity. SLAs and whether the model is actually there still matter. They are nowhere near the whole list.

He has customers well down the path of an agentic system they designed, tested, and got excited about, now sitting in a compliance backlog pending approval of the specific model underneath. Sometimes the job is just reading PDFs into structured data. The blocker can be whether all the training data was legally acquired. Non-functional requirements have not gone away: regulatory, contractual, and a precise account of how every bit of PII moves across a six-system workflow. Model provenance is the relatively new exam question. The rest is the same governance and security primitives a platform needed before generative AI.

Is that backlog prudence or calcification? Both, he says. Large enterprises really do need to challenge frozen process. Liability, regulation, and reputation risk are also real. The mix of caution should match the workflow. Allocating internal IT resources can take more risk. Loan origination, where a regulator can contest every decision and every step must be auditable, cannot. Labenz asks about model-company indemnification and AI insurance, including AIUC, which he has covered on the show. Martin has not seen that uptake among platform customers. It is an interesting idea, and for regulated industries it is another unproven thing that still needs to mature.

After the break, Labenz turns to cyber risk and the fashion for rewriting everything in Rust. Martin is investing heavily, and he is thinking about two domains at once: OutSystems' own platform and infrastructure, and the thousands of systems customers have built on it. In both, the first principle is speed of remediation. Threat velocity is rising. Patches have to keep up, for OutSystems and for customers.

A concentrated portfolio on shared primitives is simply easier to remediate than hundreds of AI-built systems, each on its own framework, stack, and grab bag of open-source components. When a vulnerability lands, the second world cannot even find where it lives, let alone fix it. Well-architected reuse means you tackle the issue once and the fix fans out. They are also raising the game: bringing the most powerful AIs in to pen-test their own systems, encouraging customers to do the same, and automating a lot of that testing into build, test, and deploy so customers do not have to think about it as much.

Labenz asks whether this is a permanent tax the frontier labs collect forever, inventing new risks only they can help you solve, or whether the industry closes the bugs in 18 months and lives happily. Martin does not believe in a finish line. Better technology also gets better at attack, and the world is still full of 60-year-old COBOL, AS/400, and Lotus Notes. Perfectly secure everything is a pipe dream. Model companies will keep burning tokens on both sides, discovering holes and plugging them. He finds it hard to see why that would stop.

Asked about token spend, Martin says the peak was June and July, and it has been tailing off. A year earlier the internal slogan was to shift everything to AI and use models for everything. They learned a lot, and delivery velocity exploded. Software engineering shipped four major features in Q4, 19 in Q1, and 26 in Q2, not minor tweaks but major new capabilities. The token burn was real and, at the time, worth it.

Then they learned they could do most of those jobs just as well, cheaper, with a better harness. They built one for their own engineering org, stuffed with their own context. They built an LLM gateway so not every job had to hit whatever Opus 4.8 was at the time. They rerouted specific work to cheaper models. They are now burning less than they had forecast for Q3. The operating lesson is that almost none of real enterprise operations, as opposed to creating a new software asset, need a frontier model. Most of that work is boring. Three-year-old models are enough. A lot of it is cheaper as deterministic code than as any model at all.

That is why agentic systems should sit on a platform that makes swapping models easy, for capability or for cost. He does not think the first-half trajectory was sustainable. Every CFO got the token bill in February and March, an oh-shit moment for the whole world. Controls followed: per-person caps, routers, pullbacks from AI-first-everywhere. Microsoft, Meta, and Uber all made public adjustments. You can still leverage AI everywhere, he says. You just have to be smart, which is the job OutSystems claims for customers.

On Chinese models he has no unique scare story. Some customers are happy to run their own versions, distill, or post-train, usually for cost. OutSystems is a global business: half the customers are in Europe, with plenty in Asia, and attitudes toward trusting US companies vary wildly. The mix of models in customer-built agentic systems is already diverse. Mentor itself is a set of MCP services any coding agent can call, with fine-tuned models underneath, so a frontier model on top is doing a lighter-weight job. The stack has many optimization layers. The platform's pitch is that OutSystems already did that hard work.

Labenz tried Mentor and asks how a platform can let relatively random users build almost anything without violating the security promise the customer's CISO expects. Martin says that balance is 25 years old. The original problem was how a person without a computer-science degree, a feel for architecture, or a deep grasp of databases and open-source libraries could still build systems that run a bank. The visual editor used to be the interface. Now the same primitives sit under a conversation with an AI, and the output is the same hardened asset.

The range is the point. If you have a bank account almost anywhere in Asia, there is a decent chance the app you use to check a balance and pay bills runs on OutSystems. If you are an oil tanker docking in Rotterdam, unloading millions of liters into terminal tanks for trucks, trains, and barges, the scheduling system is on OutSystems too. The logic that stops diesel from being pumped through a line that last held kerosene, and that manages cleaning, flushing, and quality control so you do not spoil millions of euros of product, all of that runs on the same platform. Those fundamentals took 25 years. They are why Martin thinks customers can iterate at the experience layer and still have things work together.

Are limited budgets and endless backlogs actually getting shorter? Two shifts, he says. First, the projects everyone was afraid to put on the list are finally getting their day. COBOL, AS/400, Lotus Notes: AI is accelerating every phase, including the unglamorous one of understanding what the old system even does, then translating it into modern requirements, building, testing, and rolling out. Insurers are picking up 60-year-old case management systems and planning them as six-month efforts instead of six-year ones.

Second, a lot of demand went self-serve. The classic dashboard request used to sit on a backlog for months. Now someone tells Claude to turn a spreadsheet into a dashboard and has it. Companies then discover they have lost standardization and integration, and they have cleanup to do. The consolation is that employees already iterated the experience until it felt right, so the remaining job is wiring backends to common data definitions and role-based access. The backlog is no longer "wait for a developer." It is "go back and integrate."

Why is this not showing up in GDP yet? Martin keeps it customer by customer. Kevin Hearn, who runs software engineering at San Diego digital bank Axos Bank, is seeing huge productivity gains and using the extra speed to compete for deposits and loans. YESCO, the company behind the "Welcome to Fabulous Las Vegas" sign, can take on more new business with the same resources. Productivity still takes time to hit a P&L, and a lot of current experiments will not pay off. Discrete manufacturing's costs sit in physical materials, so tariffs may move margin more than AI. Consulting is mostly information, so AI can take on much more of the work, including demand and margin structure. Fast movers and wait-and-see firms will split inside the same industry on both margin and share.

Labenz wonders how to tell early, painful experimentation from work that the next two model generations will make easy. Mentor itself has been a long lift. The project started in 2018, before Martin arrived, as machine learning across the software-development lifecycle. There was no generative AI. It is now a set of MCP services that help you build software that still meets non-functional and compliance requirements, and the models underneath change constantly as they optimize for cost. Capability is better, people can do more faster, and the hardest question has barely moved: what should I actually build?

OutSystems learned that the hard way across thousands of customer applications. They built an AI that looks at telemetry, structure, and data flow and tells the customer which agentic system to create on top of assets already on the platform. A new "agent foundry" is in the hands of customer-success teams: it forecasts time-savings ROI, suggests turning a manual job into an automated one, and, if you press the button, lets the AI build a v1 that still goes through normal development, testing, and validation. Solving "what to do," Martin says, is a huge part of agentic transformation.

Mentor looks like the familiar AI sidebar wrapping a legacy app: drag-and-drop became tool use, and both still exist. Labenz asks where the puck is going, including an MCP world where you never even see the tools. Martin still expects more than 100 apps on his phone a year from now. Not everything wants to be a conversation. Google Maps is the example he uses every day, for pizza and for navigating a foreign city. Voice can give turn-by-turn directions. He still wants to know what the world looks like around him and what he will pass. Privacy, compliance, and preference will keep interaction paradigms diverse. Big portals with 80 features may get widgetized into the flow of other work. That is not the same as everything becoming chat.

At the June user conference they showed a bank demo: a customer talking to Claude about a home remodel is told they are preapproved for a $30,000 loan, then slides into the bank's mobile app with the same agent behind both surfaces. Because the bank already has the customer, it does not ask for a bank statement. Agents handle or recommend each backend step, with humans in the loop where needed. Some software asset, here a mobile app, still has to exist for compliance and privacy. The question that interests Martin is not how everything becomes the same. It is what creators will imagine once AI and platforms like OutSystems remove the old constraints.

Competition is the ugly twin of that sameness. Drive Highway 101 in San Francisco, Martin says, and the billboards all have the same five words with a different logo. Conversational UI flattened the surface. Behind it you still need systems, data, and how they interact, which quickly becomes an agent platform, an AI builder, and a pile of integrations, the basic blocks every vendor is rushing to fill. Differentiation, he argues, will be specialization on top: tuning, distillation, cost structure, and industry workflow. OutSystems is betting on regulated industries, banking, insurance, government, healthcare, transportation logistics, energy and utilities, which are already the core of the installed base. Beyond the commodity list they have 37 extra features that help a customer actually get the first system built and deployed, 25 years of residue. Right now the industry is confusing buyers by insisting everyone has everything.

On talent he is "super bullish" on juniors whose last five years were native to these tools, AI-pilled people who can accelerate innovation if they also learn the industry. Organizations have not built the infrastructure to indoctrinate that specialized knowledge yet. OutSystems is leaning into full-time employees who help customers decide where money will matter, treating every new piece of work as a question of human, agent, or blended team, and replacing two-day onboarding dumps with just-in-time knowledge at the moment it can be used. He closes by pointing listeners to the OutSystems world tour in Las Vegas, then Labenz lets him get back to work.