{"slug": "anaconda-scale-2026-lessons-on-scaling-enterprise-ai", "title": "Anaconda Scale 2026: Lessons on Scaling Enterprise AI", "summary": "Anaconda hosted its inaugural Scale event, where customers Plug&Pay, Notion and Tala described how they build and scale enterprise AI stacks on Anaconda's platform, which now incorporates the acquisitions Outerbounds, Kilo and Enkrypt AI. Notion CISO David Tsao said replacing a general-purpose model that flagged 3-4% of traffic, mostly false positives, with an Enkrypt AI guardrail layer deployed on Notion's own infrastructure cut flagged traffic to under 0.1%, a more than 30x reduction. Plug&Pay software engineer Daniel Langezaal said 80% of the engineering team onboarded to Kilo without any mandates after the company consolidated six disparate AI tools to escape credit-based pricing markups.", "body_md": "Last week, Anaconda hosted its inaugural Scale event. I was joined by my colleagues and a few of our incredible customers to share best practices and actionable steps for organizations looking to build and scale their AI stack.\n\nAt Anaconda, we are building a trusted foundation for enterprises to integrate, secure, govern, and operate AI. We’ve brought three acquisitions (Outerbounds, Kilo, and Enkrypt AI) into the fold to help power that work. I wanted to reflect on the event and share some of the insights I hope builders take with them.\n\n## The path to trillion-token scale\n\nFirst, let’s talk about what we mean by scale. Anaconda has a vision of effectively unlimited AI use, where organizations operate AI at trillion-token scale as it becomes embedded across business applications, workflows, and operations.\n\nIt’s an ambitious prediction, but a very possible one. To get there, enterprises will need to:\n\n- Secure and govern the packages, models, agents, tools, and data that comprise their AI workflows\n- Preserve choice across models, tools, infrastructure, data, and deployment\n- Move AI from development to production with repeatability, visibility, and traceability\n- Optimize costs by matching workloads to the correct models and infrastructure\n\nYou don’t have to do it all at once. Start where your greatest friction is, whether that’s development, security and governance, or production operations, and build toward a broader trusted foundation over time. The three customer stories below each start from a different place: Plug&Pay on cost and choice, Notion on security, and Tala on production.\n\n## The hidden cost of AI tool sprawl\n\nI was able to sit down with Daniel Langezaal, a software engineer at Plug&Pay, a Europe-based online platform for entrepreneurs. He shared his team’s story of building their AI workspaces, starting with the main problem: they were using six different tools that were completely disparate.\n\nBut the real risk was pricing, not tool sprawl. Tool after tool would switch from unlimited use to credit systems with seemingly random markups. They realized that this model was unsustainable if they wanted to build workflows at enterprise scale.\n\nPlug&Pay turned to Kilo, seeking a platform that provided a single place to manage the parts of their workflow that were burdened with complexity, and one that immediately understood their existing infrastructure.\n\nThe results had a broad impact across Plug&Pay’s engineering organization. Without any mandates, 80% of the engineering team onboarded to Kilo. Kilo is not their entire AI stack, but it is the piece that they trust to stay stable underneath everything else.\n\nKilo offers customers model choice and cost optimization. If you’re the person who has to explain the AI line item, that is cost optimization your team applies by default with no procurement cycle attached.\n\n## Implementing safety controls\n\nMy colleague Jane Kim, Chief Commercial Officer at Anaconda, interviewed David Tsao, CISO at Notion. Notion’s AI functionality had matured from answering simple questions to taking real action within a workspace on someone’s behalf. That level of autonomy meant that David’s team needed to enact strong controls and guardrails to ensure the agents never stepped out of policy.\n\nOriginally, the team’s first line of defense ran all prompts through a general-purpose model for judgment. It worked, but with one critical flaw: at Notion’s scale, the model flagged 3-4% of all traffic, the majority of which was false positives. This approach buried the real threats and was unable to scale at the speed they needed.\n\nNotion began using Enkrypt AI. After testing the proof of concept, they moved to a guardrail layer deployed on their own infrastructure. This migration was complicated and took significant engineering power to adapt the stack to their infrastructure, but the returns have been immense.\n\nFlagged traffic went from 3-4% to under 0.1%, a more than 30x reduction. For David’s team, this means they no longer have to sift through noise to find the most critical threats. When applied correctly, security enables velocity.\n\n## From experiment to production\n\nFinally, Todd Osborne, SVP of Global Sales at Anaconda, sat down with Will High, Head of Data Science and Machine Learning Engineering at Tala. Tala is a financial platform serving borrowers across the Philippines, Mexico, and Kenya, many of whom have been historically underserved by traditional lenders.\n\nWill’s team builds the models behind real-time credit decisions, where a slow or fragile pipeline can mean the difference between someone getting a loan when they need it and not getting one.\n\nAs Tala grew, its infrastructure was unable to keep up with the increasing demands. Kubeflow and Airflow were heavier than the team needed, retraining a model could take months, and onboarding a new hire meant manually standing up EC2 instances and Jupyter servers.\n\nTala moved onto [Anaconda AI Orchestration](https://www.anaconda.com/platform/ai-orchestration). Training, evaluation, and deployment now run without manual management, and everything lives in one shared codebase. Retraining that took months now takes days, and deployment dropped from as long as five days to about 90 minutes. Tala went from about four models a year in 2021 to 41 in 2025, with headcount staying close to flat.\n\nBy standardizing on a single, repeatable pipeline and automating the manual steps between development and production, teams can focus on the problems worth solving rather than infrastructure upkeep.\n\n## Looking Ahead\n\nThe stories from Plug&Pay, Notion, and Tala all address different challenges and opportunities, but point to the same conclusion. To scale AI, you need to build on a foundation you can trust: one that keeps costs in check, secures what your agents and models touch, and moves work from development to production without friction.\n\nAnd you don’t have to tackle it all at once. Start where your greatest friction is, and build from there.\n\nIf you weren’t able to join us live for Scale, you can [access the event on demand](https://www.anaconda.com/resources/webinar/anaconda-scale-2026).\n\nVisit our [Case Studies page](https://www.anaconda.com/resources/case-study) to see our customer stories in action.", "url": "https://wpnews.pro/news/anaconda-scale-2026-lessons-on-scaling-enterprise-ai", "canonical_source": "https://www.anaconda.com/blog/lessons-anaconda-scale-2026", "published_at": "2026-10-19 19:05:24+00:00", "updated_at": "2026-10-19 19:47:58.352784+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-safety", "ai-agents", "ai-infrastructure"], "entities": ["Anaconda", "Outerbounds", "Kilo", "Enkrypt AI", "Plug&Pay", "Notion", "Tala", "David Tsao"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/anaconda-scale-2026-lessons-on-scaling-enterprise-ai", "markdown": "https://wpnews.pro/news/anaconda-scale-2026-lessons-on-scaling-enterprise-ai.md", "text": "https://wpnews.pro/news/anaconda-scale-2026-lessons-on-scaling-enterprise-ai.txt", "jsonld": "https://wpnews.pro/news/anaconda-scale-2026-lessons-on-scaling-enterprise-ai.jsonld"}}