AI for business operations: what it handles at an ecom brand today

Plain automation follows rules you can write as if-this-then-that: tag the order, hold fulfillment, send the email. It's cheap and reliable, and it's covered in the guide to ecommerce automation services.

AI operations covers the work that rules can't describe in advance. A customer writes three paragraphs about a damaged box and a missing item. A supplier sends an invoice as a scanned PDF with a different SKU format. Someone has to read it, understand it, check it against your policy or your purchase order, and act. Language models can now do that reading and drafting, inside limits you set.

• Policy exceptions and large refunds, where a confident wrong answer costs money and trust.
• Supplier negotiations and relationships.
• Deciding which creative goes live and what the brand sounds like.
• Situations with no history in your data: a new carrier failure, a recall, a viral complaint.
• Anything physical: samples, formulas, shoot days, the warehouse floor.

The first item is where brands get hurt. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline liable after its website chatbot told a customer he could apply for a bereavement fare after travel, which contradicted the airline's own policy page. The tribunal treated the chatbot as part of the website, so the airline owned the answer.

Klarna is the other case worth knowing. In February 2024 it said its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats, with resolution time down from 11 minutes to under 2. In May 2025 its CEO told Bloomberg the company was investing in human support again and that customers should always be able to reach a person. Both statements are Klarna's own. Read together, they suggest planning for AI to take the routine volume while customers can still reach a person.

An NBER working paper followed 5,179 customer support agents through the rollout of a generative AI assistant. Issues resolved per hour rose about 14% on average. Novice and lower-skilled agents improved 34%, while the most experienced agents saw minimal change. The tool spread what the best agents did to everyone else.

MIT's NANDA initiative reported in 2025 that most enterprise generative AI pilots showed little or no measurable effect on profit and loss. It also found that more than half of budgets went to sales and marketing tools, while the biggest returns were in back-office automation.

For an ecom operator, both point the same way. Start in the back office, in queues with lots of repeat work, and write down the rules your best people already follow.

1. Pick one queue. Good candidates have steady volume, a written policy, and mistakes you can catch before they cost much: support tickets by reason, shipping claims, or vendor invoices.
2. Count it. Pull the last 90 days: how many items, sorted by reason, and roughly how many minutes each one takes. If you can't count it, fix the tagging first.
3. Write the rules your best person follows, including the exceptions they handle without thinking. This document becomes the system's instructions.
4. Set the handoff line before anything is built: the refund amount, customer tone, or unknown case that goes straight to a person.
5. Price a wrong answer. A wrong order-status reply costs a follow-up ticket. A wrong refund approval costs the refund. Start where wrong answers are cheap.
6. Run it in shadow mode. The system drafts, a person sends, and you compare its drafts against what the person actually did for two to four weeks.
7. Decide on a number you chose in advance, such as the share of drafts sent without edits, and only then let it act on its own for the easy cases.
8. Name an owner. Someone has to watch the logs, update the rules when policy changes, and test new models before they go live.

Illustrative example, not a client result: a brand picks "where is my order" tickets. It counts them, writes the rules (look up tracking, give the carrier status, escalate if the package hasn't moved in five days or the customer has written twice), and sets the handoff for anything mentioning a refund. Two weeks of drafts in shadow mode show where the rules were missing a case. After the fixes, the system answers the routine tickets and the team handles the rest.

The tools fall into a few groups. Helpdesks now sell AI agents that answer tickets, often priced per resolved conversation: Intercom lists its Fin agent at $0.99 per outcome, and Gorgias charges an outcome-based automation fee when its AI Agent resolves a ticket without a person. Workflow tools connect your systems and can call a language model at a step. Custom systems built on model APIs cover the work that crosses several systems at once, like matching a 3PL invoice against purchase orders and receipts.

Per-outcome pricing is easy to compare against the cost of a person handling the same ticket. Check how each vendor defines an outcome, though. Intercom counts a conversation where the customer doesn't ask for more help after Fin responds, and Gorgias counts one where no human gets involved within 72 hours.

To see what this does to a full team, the Zero Hire Org Chart rebuilds a $50M brand seat by seat. For outside help, see AI consulting for ecommerce and how to evaluate an AI automation agency.