Box chief executive Aaron Levie sees a sharp architectural break coming to enterprise software: As AI moves from feature to foundation, modern SaaS will keep a deterministic core for critical workflows while a fast-evolving layer of AI agents sits on top to guide decisions, automate tasks, and accelerate knowledge work.
That split, he argues, could ripple through pricing, security, and product design — and open a rare window for new winners.
Why Deterministic Cores Still Matter
Levie’s central point is pragmatic: Enterprises cannot entrust mission-critical processes to systems that may behave differently today than they did yesterday. The source of truth remains deterministic software: well-specified business logic, with approvals and controls in place. AI belongs in a separate but integrated layer, where it can propose, summarize, predict, and automate with the production keys still under human lock and key.
The rationale is hard-earned. High-profile mishaps have shown how generative systems can leak sensitive information or propagate unintended actions when given broad autonomy. A silly but apt example was widely cited: confidential code was pasted into a public chatbot, prompting embarrassed companies to launch roles, permission checks, and data boundaries.
Levie’s “church and state” framing mirrors how regulated industries already divide systems of record from analytics and experimentation.
Agents Become the New Superusers
Levie anticipates a future where AI agents outnumber humans by orders of magnitude — 100x, and even 1,000x. In such a world, a SaaS platform’s majority “users” are software entities working with individuals and with each other. They are analogous to today’s RPA bots and CI/CD jobs, but with natural language, retrieval, and reasoning capabilities covering enterprise content. This reality resets the stack’s design.
- Identity portrays agent personas.
- Authorization necessitates granular, context-aware scopes.
- Audit logs must record prompts, intermediate steps, and citations.
- Guardrails — policy checks, data loss prevention, and human-in-the-loop reviews — transform from bolt-ons to first-class features.
- Vendors who can elucidate and validate agent conduct will gain trust quickly.
Per-Seat to Consumption Economics
When agents are the central users, per-seat pricing fails. Users are unable to make rational decisions on a per-user basis. Consumption and volume schemes linked to tokens, documents processed, or workflow executions will become mainstream, according to Levie — more Snowflake or Twilio than traditional licenses. This change also necessitates transparent metering, cost management, and FinOps controls to allow customers to predict, cap, and optimize spending.
Present-day market data confirms the change. According to Gartner, more than 80% of companies will utilize generative AI APIs or genAI apps by 2026, up from less than 5% in 2023, and usage-based purchasing will become more conventional. Forrester pegs AI software spend at a 23.5% CAGR through 2025, nearly doubling its pre-pandemic forecast. IDC predicts that global AI spending will increase at a double-digit CAGR throughout the middle of the decade, with software and services driving the change. As utilization grows, CFOs will push for quantifiable efficiencies earned per dollar — time saved, claims settled, footage produced.
A New Opening for Startups
Levie compares this era to the mobile and cloud inflection more than ten years ago. Startups aren’t burdened by legacy roadmaps or per-seat revenue models; they can design “agent-first” from day one.
- Orchestration frameworks
- Agent policy engines
- Evaluation tooling
- Retrieval and grounding
- Secure sandboxes for testing autonomous workflows against synthetic and red-teamed scenarios
There’s a change-management gap, too. Enterprises will need opinionated playbooks to map existing processes to agent-assisted flows, plus integrations with content systems, ticketing, and ERP and data catalogs. Vendors that make rollout safe and auditable — covering identity, data residency, and industry controls like SOC 2 and ISO 27001 — will shorten sales cycles and ease procurement risk.
What Enterprises Should Do Now
Levie’s blueprint suggests several practical steps.
- Define the control plane: keep core workflow logic deterministic, expose clear APIs, and let agents operate at the edges with scoped permissions.
- Classify content and ground models in governed repositories to reduce hallucinations and leakage.
- Standardize evaluation — benchmarks for accuracy, latency and cost, and safety, with “reject” policies when confidence drops.
- Align economics to outcomes.
Dominate pilots where ROI is evident — sales proposals, contract review, customer support summaries, and knowledge retrieval are often early wins.
“McKinsey estimates that generative AI could add $2.6T to $4.4T in annual value globally, but those gains concentrate where workflows, data quality, and governance are mature,” concludes Levie. The message, then, is clear: keep the core steady, let agents accelerate the edges, and price for a world where your most active users don’t carry badges.