What Business Processes Should You Automate with AI? (2026)

McKinsey’s State of AI 2025 found 62% of organizations at least experimenting with AI agents and 23% scaling an agentic system in at least one function. That is not a reason to automate “everything repetitive.” It is a reason to pick the processes an agent can take a first pass on — and leave the ones that already have a correct template alone.

On June 17, 2026, Amazon Bedrock AgentCore Harness reached general availability (What’s New). Agents Classic is in maintenance for new customers after July 30, 2026. Those dates made a first production loop cheaper to host. They did not change which business processes are agent-shaped.

This is part 1 of AI Agents for Business — five executive notes for operators who have not built an agent yet. The 64-part eCommerce field guide is the library. The commercial hub is the door after this post.

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

Reproduce this — Score the process on ai-agent-readiness-checklist.md (0/1/2 per row, total /30). Below 16: do not attach write tools this quarter. Series folder: ecommerce-ai-agents-series/.

Opinionated take: automate the lookup pile first — order status, policy, a reorder risk brief — not tax, capture, or labels. Trade-off: the first agent will not “close the ticket end-to-end.” The alternative is an unsupervised write on a process that already had a correct answer.

The filter: agent-shaped vs already-solved

A process is agent-shaped when all four are true:

1. A human already does a first pass every week (tickets, a queue, a spreadsheet).
2. The facts live in named systems (OMS, WMS, help center, ERP) — not in one person’s head.
3. The first pass is messy language or incomplete evidence, not a fixed formula.
4. Someone can own a wrong answer (a named queue, not “the AI team”).

A process is already-solved when a template or state machine is correct: cancel windows, tax, payment capture, allocation, carrier labels. Agents vs workflow automation is the build note. This post is the executive cut: do not replace those paths.

FactualMinds is an AWS Select Tier Services Partner. We help commerce teams find the first workflow worth an agent — after the filter, not instead of it. There are no published AI-agent case studies on this site. Proof is the field guide and the first-party platform numbers above.

Five families, one first process

The AI Agents hub groups work into five families. Pick one process inside one family — not a fleet.

What broke — Teams that swapped a working carrier-status email template for a chat agent that “sounded helpful.” The bot invented clock-time ETAs when getShipment and the carrier page disagreed. Detection: shoppers quoting ETAs that were not in the OMS. Recovery: cite tool evidence or escalate; do not average timestamps. The support control-plane note is post 2 of the field guide.

Named substitutes when “automate it with AI” is the wrong brief

• Already a correct template → keep the template; add search if people cannot find it.
• Fixed state machine (refund window, tax, capture) → keep the workflow; agent returns a structured decision only.
• FAQ search with no tools → that is a chatbot, not an agent. Ship search.
• No named APIs, no owner, no approval queue → stop. Run the readiness assessment before a model bake-off.

If You Only Do One Thing

List the ten processes that ate last week’s calendar. Cross out every row that already has a correct template. Score the rest on the readiness checklist. Take the highest-scoring read process to the hub — not a refund tool.

What to Do This Week

1. Name one owner (ops or support) and one engineering counterpart. No owner → no agent.
2. Confirm the process has a named API or admin export. “We can screenshot Shopify” is not a tool.
3. Write the stop rule: what the agent must not do in week one.
4. Open the matching family page on the hub. If you are still choosing among families, use where to start.

What This Post Doesn’t Cover

It does not score your org /30 — that is the readiness post. It does not compare agents to chatbots in depth — that is part 3. It does not publish a client “hours saved” number. We have not run a new first-party process census for this note; the Gateway latency and ~$791/mo floor are the published benchmarks we are willing to reuse. AWS how-to stays in the field guide.

Primary next step: AI Agents hub.