During our Phoenix Spark team’s week in the US and the winners’ visit as part of the Phoenix.AI Program’s Jedi contest, one of the highlights was attending the Plug and Play Summit in Sunnyvale, one of the key events for American and global innovation. Thousands of participants attended, including a huge number of start-ups and industry leaders: investors and innovators from all over the world came together to showcase the future and work on key technological and business trends.
One of the most interesting talks was that given by Amit Patel, Partner and Head of Enterprise AI at Plug and Play, on Enterprise AI.
Amit Patel framed enterprise AI around four imperatives: organizational design, infrastructure alignment, experimentation velocity, and scale, arguing that the winners will redesign companies around AI-native capabilities rather than bolt-on tools.
1. Why smart agents still fail
Patel showed that most “smart agents” fail because they are dropped into messy enterprise environments without the right context and memory layer. A typical pipeline—request → routing → model selection/chaining → context & memory → testing/learning → response—breaks when the agent cannot reliably access CRM data, data warehouses, shared drives, or messaging channels with clear rules and provenance. The result: impressive demos that cannot safely decide, for example, which churning customers deserve discounts in production.
2. The Enterprise Context Layer: the real “brain”
A central concept was the Enterprise Context Layer, described as the company’s “central brain” for data and business logic. It aggregates structured data (databases) and unstructured data (Slack, Google Drive, docs) and uses LLMs to draft baseline context that humans then refine into business rules. This layer becomes a living, self-updating rulebook that agents query, plus a searchable collection of precedents so AI can act as if leadership were making each decision.
This context layer sits between business systems of record (Salesforce, SAP, Gong, Snowflake, SharePoint, etc.) and interfaces/agents (Claude, custom copilots, Slack bots, etc.), with specialized vendors emerging in this middle tier. For enterprises, building this layer is now as strategic as choosing the CRM itself.
3. Production AI is composite and cost-intentional
Patel emphasized that real production AI is not one model call, it is many. User requests are routed through engines such as LiteLLM or OpenRouter into a multi-model portfolio: cost-efficient models for bulk tasks, low-latency models for interactive use, and high-accuracy models for high-stakes decisions. Requirements around complexity, cost, and latency drive this routing rather than brand loyalty to a single provider.
He contrasted monolithic AI architectures -“pick the biggest model we can afford for everything” – with deliberate multi-model architectures. Drawing on examples like AT&T’s reported 8B tokens/day usage and 90% cost reductions, he argued that cost-intentional design from day one is now a board-level concern, not an optimization left for later.
4. The four imperatives
- Organizational Design
- Routine work is automated while domain experts become “AI supervisors” who define guardrails, review edge cases, and tune feedback loops.
- A formal “AI translator” role is needed between business units and technical teams to map use cases into data, models, and policies.
- Operating models must be federated: central standards and governance, but experimentation pushed out to business lines.
- Infrastructure Alignment
- The four-layer AI stack—data foundation → model portfolio → orchestration → application layer—only works if the context layer and routing are treated as first-class products.
- Data fabrics replace silos, and AI-native workflows are designed end-to-end rather than grafting chatbots onto legacy processes.
- Experimentation Velocity
- Patel advocated moving from annual planning to continuous “VC-style” funding: stage-gates, prototypes instead of PowerPoints, and rapid kill-or-scale decisions.
- Marketing, for example, shifts from quarterly pilots to weekly multivariate tests where agents and humans iterate together on campaigns and customer journeys.
- Scale and feedback loops
- In the “era of agents,” scaling means industrializing feedback: instrumenting cost per transaction, error rates, human handoff frequency, and customer outcomes for every agent.
- Winning organizations ruthlessly scale what works and shut down failing experiments quickly instead of letting pilots linger.
5. Where the real moat is
Patel closed with a strategic warning: your moat lives around the model, not inside it. Routine tasks, generic content generation, public-data-only insights, and undifferentiated workflows will all be commoditized as models converge.
Defensible moats come from proprietary data combined with compounding feedback loops, deep domain expertise, embedded process know-how, strong customer relationships, network and distribution advantages, and regulatory and compliance infrastructure. AI-native competitors already exploit the cost curve; incumbents must lean into their physical and relational assets, not just model selection.
Summary compiled by Paola del Zotto Ferrari, Resident Director – Phoenix Spark
