"We literally couldn't ship a new feature fast enough before the next model came out and made all of our engineering completely pointless. So we throw away finished code all the time." That's how Box CEO Aaron Levie opened his conversation with Matthew Berman at BoxWorks 2026. You assemble a team, spend three months wiring up intricate business logic, and a week before launch a new base model drops—turning your entire module into dead weight because the model now handles the job out of the box.
That raises an obvious question: how are you supposed to design software when a single model update can wipe out months of custom engineering overnight? The answer Box landed on—an announcement day that sent the company's stock to an all-time high—flips how we look at IT products: value is migrating out of application code and UI screens into structured company context, while the code and buttons themselves become disposable scaffolding.
Two Floors of Modern Software: Swappable Agents and a Foundation of Ground Truth
Two years ago, Microsoft CEO Satya Nadella predicted that all software would split into two core layers: a fluid upper layer of AI agents (Systems of Intelligence) and a foundational lower layer of Ground Truth Context. Following NVIDIA CEO Jensen Huang's estimate, which Levie uses as a benchmark, every employee will soon work alongside 10 to 100 digital assistants—amounting to 30 to 50 billion autonomous agents across the global economy.
You might think: "Great, let's just dump every company file into one big bucket and turn the bots loose!" I had that exact thought too, but in practice an untagged data dump turns into a disaster. Picture a hundred eager interns let loose in a file room where old contract drafts, current price lists, and payroll spreadsheets are piled on the floor without folders or locks. The very first intern will quote numbers from an expired draft or email everyone's salaries to the sales team.
That's why the bottom floor—modernized Systems of Record—becomes the real center of gravity. Your data layer has to enforce four non-negotiables:
- Granular access controls (RBAC and ABAC): an agent sees only the folders and files its triggering employee is cleared to view;
- Metadata and routing: crisp tagging so the model knows where the policy lives instead of brute-forcing gigabytes of archives;
- Sandboxing and audit trails: strict isolation of actions and an immutable operation log;
- Headless API-first design: just as Jeff Bezos mandated programmatic interfaces across Amazon 15 years ago, today you have to assume no human will ever look at your UI—agents talk to the system directly through fast data schemas. For a look at how teams overcome legacy inertia in practice, see our guide to enterprise AI agent adoption.
The Bitter Lesson in Product Engineering: How 30 Seconds of Agent Work Replaces Two Days at a Whiteboard
Six to nine months ago, a Box engineer ran an internal experiment: hooking a frontier language model into a workflow builder and asking it to design a complete business app. In 30 seconds, the agent built from scratch a process architecture that customers normally spent two full days sketching on a whiteboard—plus weeks stitching three to five separate services together.
Take a real-world case from an insurance company's legal department processing 5,000 reinsurance contracts a month. Automating that kind of reconciliation used to mean six months of custom parser development costing hundreds of thousands of dollars. Now, inside Box Architect (unveiled at BoxWorks 2026), a lawyer types in plain English: "Cross-check the liability limits in Appendix B against the master policy and flag discrepancies in a separate register." A frontier model (Claude Opus, Fable, or Astra) assembles the entire pipeline in half a minute, then self-optimizes the routing by analyzing failure logs. Similar patterns are emerging in customer support with Decagon, Sierra, and Fin, where a meta-agent helps operations leads build specialized bots.
Aaron Levie wryly calls this embracing "the bitter lesson"—a nod to Rich Sutton's essay The Bitter Lesson on general-purpose compute beating hand-crafted heuristics. A team can spend three months writing elaborate tool-calling scaffolding, only for a summer leap in computer use and business-logic reasoning—reflected in benchmarks like GDP Val—to render all that code obsolete before launch. The only things worth building for the long haul are durable security rails and data governance; temporary wrappers should be written off without regret as base models improve.
Unstructured Files and the NVIDIA Nemotron Index: Why the Context Owner Wins
Once code stops being a moat, a company's real edge shifts to unstructured data. Roughly 80 to 90% of critical enterprise knowledge doesn't live in SQL tables—it sits inside contracts, slide decks, CAD drawings, and financial reports.
Levie predicts that within three years, every digital event in a company will trigger a crew of 1, 3, 5, or 10 agents: in sales, updating customer dossiers and buying priorities on the fly; in cybersecurity, testing every suspicious login and code push in simulations; in logistics, parsing supplier invoices and contracts straight into ERP systems. For that to work, data has to flow freely and safely wherever work happens—into Salesforce, SAP, Oracle, ChatGPT, or Claude. You can dig deeper into connecting knowledge bases to models in our guide on choosing between scripted bots, RAG, and agents.
Yet if dozens of agents shove entire massive PDFs into their context window on every query, you'll go broke on latency and tokens. That's why Box partnered with NVIDIA to launch Box Context Retrieval, powered by Nemotron models (in Enterprise Advanced). It pre-indexes and ranks the archive, serving agents only the relevant chunks via API—complete with page-level provenance tracking to keep hallucinations in check.
- Separate access perimeters: never toss contracts, accounting files, and public templates into one folder without metadata and permission checks.
- Require page-level citations: configure retrieval so the AI backs every figure in a report with a link to the exact source page.
- Stay model-neutral: store instructions in open formats so you can swap models without rewriting integrations.
The CFO's Dilemma: $30 Seat Subscriptions vs. Unpredictable Token Bills
For all its speed, agentic automation comes with a financial catch. Headcount is always finite, while compute can scale up to the size of a data center—a pharma company can throw AI resources at molecular simulations to ship a drug faster. The trade-off is losing twenty years of predictable software budgeting.
Classic SaaS ran on flat per-seat pricing (say, $30 per employee per month). In a world where token consumption is set to grow 1,000x over three years, forecasting the cost of a task becomes nearly impossible—which sends shivers down every CFO's spine. With blind token maxing, one sloppy prompt can send agents into an infinite loop and torch a weekend budget, while relying on one or two closed labs risks vendor lock-in.
Three practical levers help keep the math in balance:
- Hybrid pricing: keeping predictable per-seat subscriptions for humans alongside capped quotas for variable agent token spend—much like marketing budgets.
- Open weights as a safety valve: just as free open-source databases grew after 2005–2006 alongside Oracle's rise to $300 billion, open and closed models push each other forward today. Open models (like
Muse Spark) trail closed flagships by just 3–6 months and handle up to 80% of routine IT tickets and standard checks for pennies. - Cascading routing and fine-tuning: letting cheap open or company-fine-tuned models (via Trajectory, Prime Intellect, Applied Compute, Fireworks, Baseten, or EnGram) handle high-volume tasks like fraud screening saves your budget for hiring expensive OpenAI and Anthropic flagships as top-level planners.
In the end, being willing to scrap outdated code and rebuild how documents are stored isn't about chasing automation metrics. When reconciling five thousand contracts shrinks from two days at a whiteboard to thirty seconds, technology stops demanding endless upkeep of spreadsheets and forms. That entire data foundation and swarm of helpers exists for one reason—to give people back the calm and time for original thinking, bold ideas, and real human conversation.

