The 2027 agent stack for mid-market enterprises: what to copy from the Fortune 500, and what to skip

The Fortune 500 makes agentic AI look expensive. JPMorgan Chase built its own internal LLM platform. Walmart announced four "super agents", each with specialised agents beneath it. Most companies cannot fund either, and they do not need to.

Scale still matters. McKinsey's 2025 State of AI survey found that larger companies are more likely to have moved AI beyond pilots: nearly half of firms with over $5 billion in revenue had reached the scaling phase, against 29% of those under $100 million. But the gap is not the technology budget. It is the discipline the largest firms have learned the expensive way.

This is the final part of our 2027 agent stack series. Part 1 covered the context layer and Part 2 the control layer. This part turns both into a stack a mid-market budget can carry, building on the 2027 AI budget guide.

Questions this article answers

Can mid-market companies afford AI agents?

Yes. The cost of capable models has fallen sharply, so the model is rarely the barrier. The real costs are context, controls and integration, and those scale with discipline, not company size. See why the model is no longer the gap.

What should mid-market firms copy from Fortune 500 AI programmes?

Five patterns: one shared platform, narrow agents organised around users, a minimum context layer, deterministic controls from day one, and a model chosen for each role. See what to copy.

What should they skip?

Building their own platform or models, enterprise-wide ontologies up front, multi-agent designs where a simple workflow will do, and stacks they cannot leave. See what to skip.

1. Why the model is no longer the gap

Stanford's 2025 AI Index estimates that querying a model at roughly GPT-3.5-level performance fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a reduction of more than 280 times. The most capable models still cost far more than smaller ones, but for many routine agent tasks, model cost is no longer the deciding factor.

That shifts where the money goes. As the budget guide set out, integration, data readiness and governance usually cost more than the model. Those are exactly the context and control layers. A mid-market firm that spends well there can match outcomes the Fortune 500 paid far more to learn.

2. What to copy

One shared platform before many use cases. JPMorgan's first stage was making models available to staff through one governed platform. The mid-market version is a single AI gateway for access, logging and spend, not a different tool in every team.

Narrow agents organised around users. Walmart organises its agents around who they serve. Copy the principle, not the scale. Start with one group of users and one workflow, such as supplier queries or invoice exceptions, rather than a company-wide assistant.

A minimum context layer. You do not need a full enterprise ontology. For one workflow you need clean master data for the entities it touches, written definitions of the terms it uses, and a register of its policies and approval limits. That is Part 1's five-test check, applied to a single process.

Deterministic controls from day one. Handoff schemas, iteration limits and escalation rules are code, not headcount. They cost little to build in early and a great deal to retrofit after an incident. This is Part 2's control layer, at the size of one workflow.

A model chosen for each role. A low-cost model for triage and classification, and a stronger one only where policy or judgement is involved. Test the choice on your own cases, not on reputation.

3. What to skip

Building your own platform or models. JPMorgan can justify an in-house platform. Most firms get more from a managed platform, with their context kept in a layer they own.

An enterprise-wide ontology up front. Knowledge graphs earn their place, but build them one workflow at a time. A two-year data programme before the first agent ships is how momentum dies.

Multi-agent designs where a workflow will do. Gartner has warned that many use cases positioned as agentic today don't require agentic implementations. If the steps are known and stable, a single agent inside a fixed workflow is cheaper, faster and easier to control than a team of agents.

Stacks you cannot leave. Use proven frameworks and managed services, but keep your definitions, policies and traces exportable. That keeps your options open as prices fall and models change, as we argued in how to choose an AI vendor without marrying one.

4. Sequencing it on a sensible budget

Start with one workflow that is high in volume, has a clear owner and has a measurable outcome. Build its context and controls before widening the agent's authority. Use the execution trace to show cost per completed task, not just usage. Self-reported time savings are not P&L, as Part 1 noted of JPMorgan's own figures.

Expand to a second workflow only when the first can show a trace, a cost and a result. Each new workflow then reuses the platform, the controls and part of the context already built, which is where mid-market economics start to compound. Keep a monthly view of agent spend so it does not become the governance gap many firms are already finding.

Your agent stack on one page

If you cannot name the owner, the cost per task and the escalation route for your first agent workflow, you are not ready to scale it, whatever your budget.

That completes our 2027 Agent Stack series. For an outside view of where your own programme stands, see our AI Production Readiness & Governance Review, or request a strategy conversation.