How to Evaluate AI Agent Opportunities (2026)

Evaluating an AI agent opportunity is not a model bake-off and not a vendor ROI slide. It is whether volume, pain, data you can join, and write-risk line up — and whether you can pay the platform floor without pretending it is savings.

McKinsey’s State of AI 2025 found 62% of organizations experimenting with agents and 23% scaling in at least one function. Most of the 62% are still deciding what to fund. This note is that filter.

This is part 5 of AI Agents for Business. The longer scoring write-up in the field guide is AI Agent ROI: What to Automate First. The live tool is the eCommerce AI agent ROI calculator.

The job. Kill any opportunity whose only number is a percentage with no units. Fund one read-shaped workflow — or fund readiness.

This week. Pick at most three candidate processes. Run each through the ROI calculator with your counts. Score the org once on the readiness checklist.

A person still signs. Anything whose “win” requires refund, price, or PO in week one — if readiness is under 16 out of 30.

Skip it when the only math is a vendor slide, when there is no named owner, or when two “winners” would split evals and ownership.

Run the tools honestly — ROI calculator with your ticket or order counts. Readiness: ai-agent-readiness-checklist.md (/30; below 16 out of 30, do not let it change orders). Series folder: ecommerce-ai-agents-series/.

FactualMinds is an AWS Select Tier Services Partner. We publish no agent case-study ROI. The calculators and the field-guide priority post are the artifacts.

Our take: your first funded agent may look smaller than the vendor slide. It will have a stop rule and a cost floor you can defend.

Four questions (in this order)

1. Volume — How many times does this process run per week? If you cannot count it, you cannot evaluate it.
2. Pain — Does a human already do a messy first pass (tickets, a queue), or is this a formula that should stay a workflow?
3. Data — Are join keys and named APIs real? Readiness /30 — assessment.
4. Write-risk — Does the “win” require refund, price, or PO in week one? If yes, and the score is under 16 out of 30, do not fund.

Then add platform TCO. Model your mix on the AgentCore pricing calculator. Support-style AgentCore at 50K sessions/mo ~$791/mo platform + model is a published silhouette (decision guide) — a floor to plan against, not a round “AI will save 30%.”

How the live tools fit (and how they lie if you let them)

What broke — A deck that treated the ~$791/mo silhouette as money already saved, then attached write tools to “make ROI true.” Detection: finance asked for the store KPI; there was only a platform estimate. Recovery: label platform TCO as cost; keep week-one reads; re-rank with the ROI calculator on counted tickets. The silhouette source is the AgentCore vs Q decision guide.

Named substitutes

• Process already a correct template → do not evaluate as an agent — what to automate.
• High calculator score, no APIs → fund integration, not an agent.
• Two “winners” → where to start + decide tree.
• Need the field-guide scoring rubric → ROI priority post.
• Ready to buy the engagement, not another spreadsheet → eCommerce AI Agents on AWS.

If you only do one thing

Put three numbers on one page: weekly volume you already measure, readiness /30, and a platform-floor estimate from the AgentCore calculator. If any cell is blank, you are not evaluating — you are hoping.

For your technical lead

On June 17, 2026, AgentCore Harness reached GA (What’s New). Cheap hosting is an input to TCO. It is not a benefit line. Agents Classic is in maintenance for new customers after July 30, 2026.

First-party signals we reuse (not store KPIs) — Support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). Gateway ~180 ms → ~95 ms median tool round-trip on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Model your mix on the AgentCore pricing calculator. Treat ~$791/mo as a floor, not as savings.

What to do this week

1. Pick at most three candidate processes from last week’s calendar.
2. Run each through the ROI calculator with real counts.
3. Score the org once on the checklist.
4. Fund one read-shaped opportunity, or fund readiness. Then open the service page only if you want a scoped first agent — not a fleet quote.

What this post doesn’t cover

It does not invent a payback period or a ticket-deflection percentage. It does not replace the field-guide ROI rubric. It does not evaluate Amazon Q vs AgentCore — that compare already exists. Who should own the loop is build vs buy AI agents. We have not added a new first-party cost run for this note; ~$791/mo, 50K sessions, ~180→95 ms, 16/30, and McKinsey 62%/23% are the published figures we reuse.

Primary next steps: ROI calculator and eCommerce AI Agents.