News

Pritam Debnath is the Global Head of AI & Data Platforms at Sysco, leading enterprise AI transformation at scale.

gettyThe next enterprise AI challenge will not be building more agents. It will be operating them economically at scale.

Agentic AI changes the economics of technology because one user request is no longer necessarily one transaction. A seemingly simple interaction can trigger planning, retrieval, multiple model calls, tool execution, memory access, validation and retries.

In a pilot, these inefficiencies can be invisible. Across thousands of users and millions of agentic interactions, they can become material. This changes the leadership question from “What does the model cost?” to “What does it cost to produce a successful business outcome, and is that cost sustainable as adoption grows?”

AI economics, therefore, cannot be treated simply as a cloud-billing or FinOps exercise. It must become an architectural and operating discipline built around four principles: visibility, control, ownership and engineering efficiency.

1. Make consumption explainable.You cannot optimize what you cannot explain.

A monthly AI bill provides financial visibility but little engineering insight. Enterprises need observability at the application, agent, model and transaction levels across tokens, context, model calls, retries and supporting infrastructure.

Agent-level visibility is particularly important. A single application may contain a lightweight classification agent alongside a retrieval-intensive agent and another performing complex reasoning. Application averages can hide radically different economic profiles.

Organizations should be able to distinguish whether costs increased because adoption grew or because each transaction became more expensive. They should understand whether context expanded, retries increased or a model or prompt change altered unit economics.

Ultimately, the most useful metric is not total AI spend. It is cost per successful business outcome.

2. Build guardrails that enable innovation.Enterprise governance often gravitates toward two extremes: unrestricted consumption followed by budget surprises, or restrictions that make experimentation difficult. Neither scales.

Instead, controls should reflect workload maturity and criticality. Experimental workloads can operate with defined quotas. Production workloads can have higher limits paired with anomaly detection. Individual agents can have thresholds appropriate to their expected behavior.

A lightweight utility agent should not have the same economic envelope as an advanced reasoning agent.

The objective is not to suppress AI usage. It is to create predictable boundaries within which teams can experiment confidently. Good guardrails do not slow innovation. They reduce the risk of innovation.

3. Align cost ownership with value creation.Shared AI platforms create leverage through common infrastructure, governance and capabilities. But when every workload is charged to a central technology budget, cost becomes disconnected from business value.

Enterprises should evolve toward federated cost attribution: centralized governance and transparency combined with distributed economic ownership. Business capabilities consuming AI should understand their usage, cost and outcomes, while platform teams provide common standards, controls and enterprise visibility.

This changes the discussion from “Why does the AI platform cost so much?” to “What capabilities are we consuming, what value are they producing and is that consumption efficient?”

Cost ownership should follow value creation.

4. Engineer out unnecessary inference.Not every AI engineering activity requires actual intelligence.

Load, concurrency, autoscaling, retry and resilience testing frequently validate the surrounding platform rather than model behavior. Mock-model services can simulate responses, latency, token metadata and failures without repeatedly generating real inference.

Actual models should be used when evaluating reasoning, grounding, prompt quality, safety or model performance.

That leads to a simple rule: Use real intelligence when intelligence is what you are testing.

Separating system testing from intelligence testing can eliminate unnecessary consumption without compromising engineering rigor.

5. Treat context as an economic resource.Context may be one of the most underestimated drivers of AI economics.

As agents accumulate conversation history, retrieved documents, memory, system instructions, tool definitions and execution state, interactions can progressively become larger and more expensive.

Organizations should actively manage context through summarization, relevance filtering, configurable history and selective memory retrieval.

The objective is not minimal context. Removing too much can degrade quality. Instead, optimize for minimum sufficient context: enough information to produce the desired outcome without repeatedly carrying information that no longer adds value.

Memory itself is not a cost optimization. Well-managed memory is.

6. Apply autonomy selectively.Dynamic planning is powerful, but reasoning has an economic cost. Planning can introduce additional model calls, tokens, latency, retries and execution variability.

Many enterprise processes already have known stages, business rules and dependencies. Asking an AI system to rediscover those workflows on every execution creates expense without necessarily creating additional value.

Use dynamic planning when the execution path genuinely depends on context. Use deterministic orchestration for repeatable, known or regulated processes.

Do not pay AI to reason about decisions the architecture already knows.

The strongest enterprise architectures will not make everything autonomous. They will combine deterministic engineering with AI reasoning exactly where reasoning creates incremental value.

7. Optimize beyond the model.Agentic AI economics extends well beyond tokens.

AI platforms depend on APIs, workers, databases, retrieval services, caches, orchestration components and analytics pipelines. Infrastructure should therefore be continuously rightsized using actual utilization rather than theoretical demand.

Reusable workflow definitions, configurations and execution artifacts can also be reused or cached where appropriate, reducing orchestration overhead without introducing the risks associated with caching generated answers.

Model selection should be equally deliberate. The most capable model is not automatically the right model for every task. Capability, quality, latency, risk and cost should all inform model choice.

Over time, intelligent model routing will make this optimization increasingly dynamic.

Go from AI FinOps to AI economics.The broader shift is from managing AI cost to engineering AI economics. A sustainable operating model creates a continuous loop: Observe consumption. Attribute it. Control it. Optimize it. Measure the outcome. Repeat.

The objective is not to make AI as inexpensive as possible. It is to make AI economically sustainable as adoption and business impact scale. Organizations that get this right will treat economics as a design constraint, not a finance problem discovered after deployment.

Make consumption visible. Align ownership with value. Put intelligent guardrails around growth. Engineer out waste. And invest in AI reasoning where reasoning actually creates value.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?