Which tasks should an eCommerce brand hand to AI agents first?
Start with work you already do every week that an agent can do by reading data and drafting: daily order reports, support reply drafts, ad reports, a marketing calendar. Score each task 1 to 5 on value and effort, and push anything that spends money, places orders or talks to customers to the back. We mapped 45 agents for our two companies this way, and the first ones live only report or draft.
In October 2026 we mapped every job an AI agent could do across our two companies: Mully, a golf apparel brand, and MFS, our fulfillment business. The list came to 45 agents. Five were live, 11 were in progress and 29 were planned.
The hard part was not the list. It was the order. This guide covers the method we used to decide what goes first, and what we actually built.
Which tasks should go to AI agents first?
Tasks you already do every week, where the agent only has to read data and draft. Those are cheap to build, easy to check and low risk if the agent gets something wrong. The research points the same way. McKinsey found that four functions, customer operations, marketing and sales, software engineering and R&D, could account for about 75% of the value from generative AI (McKinsey). For a store, that means support, marketing and reporting are where to look first.
Good first candidates share three traits:
- It repeats. Daily or weekly, with the same shape each time.
- The data is reachable. Shopify, your helpdesk, your ad accounts, your analytics.
- The output is a report or a draft. A person reads it before anything happens.
How do you decide the order?
Score every task on value and effort, then sort by risk. We used five steps:
- List the repeat work. Ask everyone which tasks they do every week that feel repetitive. Those are your next agents.
- Give each one an owner, a KPI and an autonomy level. No agent gets built without a human owner and one number that says whether it is working.
- Score value and effort from 1 to 5. That puts every agent in one of four groups: quick wins (high value, low effort), big bets, fill-ins and later.
- Move risky work to the back. Agents that spend money, place orders or talk to customers come last, even when they score well.
- Prove the process by hand first. If a person cannot run the process well manually, an agent will only automate the mess.
We also put trackers before fixers. An error rate tracker comes before a quality control agent, and an "automations identifier" comes before the agent that builds automations. Each fixer then starts with real data instead of guesses.
What did we actually hand off first?
Reporting and drafting, in that order. These are from our registry, with the autonomy level each one runs at:
| Agent | What it does | Level | KPI |
|---|---|---|---|
| Fulfillment Watch | Daily fulfillment report | L3 (only reports) | Orders past SLA |
| Ad Creative | Drafts ad creatives for Meta | L1 | Creatives beating the CTR benchmark each month |
| Ad Buyer | Manages Meta campaigns | L2 (acts after approval) | Blended ROAS at or above 2.0 |
| CS Tree | Turns support tickets into a decision tree with draft replies | L1 | Share of tickets matching no branch |
| Marketing Calendar | Plans campaigns and writes briefs | L1 | Share of campaigns launching on their date |
| SEO / GEO | Tracks search and AI traffic, specs fixes | L1 | Google clicks plus AI visits |
The only agent that runs on its own is the one that only reports. The one that touches ad spend acts only after approval. Everything else drafts.
When does timing beat the score?
When waiting costs you a season. Our registry scored a retention and failed-payment agent at 5 for value and 2 for effort, higher than the SEO agent. We built SEO first anyway. Holiday gift pages need to be live well before search peaks, and our search impressions had fallen 81% after we moved the site off Shopify. A month of delay would have cost the whole holiday window. Retention went right behind it.
Use the score to make the list. Use the calendar to break ties.
How do you keep early agents from becoming expensive mistakes?
Start every agent at the lowest autonomy level and make it earn the next one. We use three:
- L1, drafts. Reports and drafts only. A person does everything else.
- L2, asks first. Can act, but only after a person approves each action.
- L3, autonomous within limits. Acts on its own inside caps, like a spend limit.
An agent moves up only by hitting its KPI. This matters because most agent projects do not survive. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear value or weak risk controls (Gartner). One owner, one KPI and a clear level answer the value and risk questions before anyone asks.
What results should you expect from the first agents?
Faster first drafts on repeat work, mostly. In a study of 5,179 support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for newer, less experienced staff (NBER). That fits what we see: the gain is in not starting from a blank page.
Do not expect the first agent to replace a role. Expect it to take one recurring chore off someone's week, with a number that shows whether it did.
Common questions
How many agents should a small brand start with?
One or two. Pick the tasks with the highest value and lowest effort that only read data and draft. Running a few well, with an owner and a KPI each, teaches you more than mapping 40 and building none. We had 45 on the map and 5 live when we started posting about it.
Should the first agent be customer support?
It can be, if it drafts replies and a person sends them. Support has repeat questions and clear data, which makes it a good first fit. Letting an agent answer customers on its own is a later step, and we have not taken it.
What is the difference between L1, L2 and L3?
They are autonomy levels we use. L1 drafts and reports only. L2 can act after a person approves. L3 acts on its own within set limits. Every agent starts at L1 and earns the next level by hitting its KPI.
Which tasks should stay human for now?
Anything that commits money or makes an exception, like ad budget changes, refunds and credits, purchase orders, pricing, and replies to upset customers. An agent can prepare all of these. A person should decide them.