How to Show the Board What an AI Agent Did (2026)
Quick summary: Count sessions, handoffs, writes denied, and one wrong answer. At 50,000 sessions the published AgentCore silhouette is about $791 a month — a bill to plan, not money saved. We are not publishing a containment rate.
Key Takeaways
- At 50,000 sessions the published AgentCore silhouette is about $791 a month — a bill to plan, not money saved
- McKinsey's State of AI 2025 put 23% of organizations at scaling agents in at least one function, and 62% at least experimenting
- This is part 6 of AI Agents for Business
- FactualMinds is an AWS Select Tier Services Partner
- That figure is from our AgentCore vs Amazon Q decision guide
Table of Contents
McKinsey’s State of AI 2025 put 23% of organizations at scaling agents in at least one function, and 62% at least experimenting. A board that funded the experiment will ask what it did. A transcript and a vendor containment score do not answer that. A weekly pack does: sessions, handoffs, writes denied, one wrong answer, and what the platform actually cost.
This is part 6 of AI Agents for Business. The pre-build ranking is How to Evaluate AI Agent Opportunities. The field-guide scoring write-up is AI Agent ROI.
The job. Give the board a record of what the agent looked up, what it refused, and what it cost — not a story that it “handled” the queue.
This week. Fill one row from traces you already have. If you have no traces, you are not ready for this pack. You are ready for the ten-ticket test.
A person still signs. Anything that moves money. The pack reports that those writes were denied or handed off. It does not celebrate a silent refund.
Skip it when the loop is not live, when no one owns the handoff queue, or when leadership will only accept a tickets-down percentage. We are not publishing one.
Copy the pack — Blank rows and the figures you must not paste as your result:
agent-weekly-board-pack.md. Price your volume on the AgentCore pricing calculator.
FactualMinds is an AWS Select Tier Services Partner. We have no published AI-agent case study. The pack is a worksheet. Your counts stay yours.
Our take: report handoffs and denied writes before you report a containment rate. The board will want the rate. You will not have an honest one until the traces exist. Ship the traces.
The five lines
| Line | What it is | What it is not |
|---|---|---|
| Sessions | Turns the host ran | A chatbot vendor’s conversation count with no tool log |
| Handoffs | Stops that opened the human queue, with a session id | A low number you hope means the bot was smart |
| Writes denied | Policy blocks on refund, cancel, purchase order, quote issue, price, or credit release | Zero, when the write tool was never attached |
| One wrong answer | The worst turn: tool result versus what the model said | A green dashboard |
| Platform cost | Your invoice line for the host plus the model | ~$791/mo relabeled as savings |
Support-style AgentCore at 50,000 sessions a month is about $791 a month in platform plus model. That figure is from our AgentCore vs Amazon Q decision guide. If you are not at 50,000 sessions, do not quote it as your bill. It is a floor to plan against, the same way the ROI note already says.
Figures that are not your KPI
Gorgias, via Redo, puts where-is-my-order at about 18% of tickets. That tells you the slice is common. It does not tell the board your agent closed 18% of anything.
Gateway server-side tools cut median tool round-trip ~180 ms to ~95 ms on a B2B CRM assistant (12 tools, about 8,000 turns a day). That is a platform canary. Your order system will be slower. Do not put the canary in the board pack as store latency.
What to use instead
- Before go-live. The readiness score. Under 16 out of 30, do not discuss writes with the board at all.
- Choosing the first workflow. Evaluate opportunities. That score is not this pack.
- You need a build note. The field guide for the family you actually shipped, starting at 15 automations.
If you only do one thing
Print last week’s traces. Count handoffs and denied writes. Attach one wrong answer. Put the invoice next to them. If you cannot do those four, do not invent a containment percentage for the board.
For your technical lead
On June 17, 2026, AgentCore Harness reached general availability (What’s New). Agents Classic is in maintenance for new customers after July 30, 2026. The pack still needs a session id and a tool trace. A Classic bot with no trace cannot fill the row.
What broke — A steering deck that treated the ~$791/mo silhouette as money saved, because the unit cost of a purchase order looked large. Feasibility was low. The team still wanted the write. The ROI note already records that failure: impact is not feasibility, and the silhouette is a cost floor, not savings. Detection: the deck’s “savings” line had no invoice beside it. Recovery: report the invoice. Rank the workflow again with the veto on.
What to do this week
- Name the agent family and the human who owns handoffs.
- Export one week of session ids. If the export does not exist, stop. You are pre-launch.
- Count handoffs and policy denies. If no write tool is attached, write “no write tool” — not zero.
- Pick the worst wrong answer. Quote the tool result and the model text.
- Put your host-plus-model cost on the same page. Compare to ~$791/mo only if you are near 50,000 sessions. Otherwise use the calculator.
- Do not paste 18%, 62%, or 23% into the “our result” cell.
What this post doesn’t cover
It does not score which workflow to staff first. It does not retell the human-in-the-loop design. It does not publish a containment rate, a tickets-down percentage, or a client board deck. 23%, 62%, ~$791/mo, ~18%, and ~180→95 ms are figures we already published, reused here so the pack has a scale — not so you can claim them.
Primary next step: AI Agents.
AWS Cloud Architect & AI Expert
AWS-certified cloud architect and AI expert with deep expertise in cloud migrations, cost optimization, and generative AI on AWS.
