The Agentic AI Value Chain Explained

Executives are told that value in AI is moving up the stack, down the stack, and to the edges, sometimes in the same briefing. This guide maps the agentic AI value chain layer by layer, shows where value actually accrues, and gives leaders a build-versus-buy rule for each layer, so that investment goes where the company can own the result.

What are the layers?

Layer	What it contains	Economics	Who leads
Compute	Chips, cloud capacity	Capital-intensive; scale	Chip makers, hyperscalers
Models	Frontier and open-weight models	Scale, falling unit prices, rapid capability change	Model providers
Platforms and gateways	Routing, agent identity, observability, evaluation tooling	Standardizing; competitive; some open source	Tooling vendors, cloud providers, internal platform teams
Tools and connectors	Access to systems and data, increasingly via MCP	Commoditizing through standards	Software vendors, integrators, internal teams
Agents	Specified, evaluated systems that do defined work	Company-specific; value depends on process fit	Enterprises with partners
Outcomes	Lower cost per task, faster cycle time, new services	Where business value is captured	Enterprises

FISTA's Model Context Protocol for the enterprise whitepaper covers the connector layer's standardization; the CIO's guide to AI and agentic AI covers the platform layer.

Where does value accrue?

At the ends. At the bottom, model providers capture scale economics: training costs are enormous, but serving costs fall with volume, and capability compounds. At the top, companies that own outcomes capture the business value agents create: an insurer whose claims agent cuts cycle time captures that value, not the model provider that supplied the inference.

The middle is standardizing. Gateways, observability, evaluation tooling, and connectors are converging on common patterns and open standards, which makes them cheaper, more interchangeable, and less of a source of advantage. For enterprises this is good news: the layers between the model and the outcome are becoming things to adopt rather than things to build.

What should enterprises build and buy?

1. Compute and models: buy. No enterprise outside the largest should train frontier models. Keep models replaceable behind a gateway.
2. Platforms and gateways: adopt standard components, build the glue. Use available gateway, observability, and evaluation tooling; build only the integration with the company's identity, data classification, and governance.
3. Tools and connectors: build company-specific ones to open standards. A governed MCP server for the company's ERP is worth building; a generic connector framework is not.
4. Agents: build, with partners. Specifications, evaluation sets, and agents for the company's own processes are the compounding assets. See AI competitive advantage explained.
5. Outcomes: own entirely. Baselines, measurement, and accountability stay with the business.

The build vs buy vs partner for AI guide applies this rule to specific decisions.

What is the strategic risk?

Being stuck in the middle: investing in generic platform infrastructure that vendors and standards will commoditize within a few years, while owning no differentiated agents or outcomes. Companies in this position have spent heavily and have nothing that compounds. The mirror risk is owning outcomes on infrastructure the company cannot govern: agents built on a vendor's closed platform with no portability, no inventory, and no evaluation the company controls. The balance is standard, governable middle layers with owned specifications, agents, and results. The LLM vendor lock-in guide addresses the governance side.

How does the chain change over time?

Models keep improving and falling in unit price, which pushes value toward the top: the same outcome becomes cheaper to produce, and the advantage shifts to whoever specifies, evaluates, and operates agents best. Standards keep spreading through the middle, which lowers switching costs. Both trends favor enterprises that own their outcomes and keep the rest replaceable. The the shift from model choice to system design discussion of model selection reflects the same logic: choose on evidence, and keep the choice reversible.

How should partners fit the chain?

A partner is useful where it accelerates the layers the company should own without taking ownership of them: building agents, specifications, and evaluation sets that the company keeps, and standing up platform layers on standards the company can govern. A partner that owns the specifications, the evaluation sets, or the connectors has moved the company's compounding assets outside the company. The test for any engagement is what stays behind when it ends.

What should executives ask?

• For each layer, are we buying, adopting, or building, and why?
• What have we built that a standard will commoditize?
• Which agents and evaluation sets do we own that a competitor could not buy?
• Could we move our agents to a different platform or model in a quarter?
• Where in the chain does our AI spend go, and where does our AI value come from?

How can FISTA Solutions help?

FISTA Solutions works in the layers enterprises should own: it builds company-specific agents, specifications, and evaluation sets through its AI agents practice, stands up standard, governable platform and connector layers through its AI enablement practice, and keeps models and vendors replaceable by design. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries.

To map your AI spend against this value chain and find where it is stuck in the middle, talk to FISTA on WhatsApp, or read build vs buy AI.