What tools does an eCommerce AI agent need access to?
Only the systems that hold the facts for its one job, usually with read-only access. A support agent needs your helpdesk and Shopify orders. An ad agent needs the ad accounts and your books. An SEO agent needs Search Console and analytics. Start every connection read-only, add write access only when the agent has earned it, and check that each connection actually returns the data you think it does.
We run a set of AI agents we built for our brands, a golf apparel store and a fulfillment business. Each one does one job, and each one connects to a short list of systems. The list matters more than the model. Several of the mistakes we have caught came from an agent reading the wrong data or missing a source, not from bad writing.
This guide covers what each agent connects to, what level of access it gets, and the connection problems we have actually hit.
Which tools does each type of agent need?
The ones that hold the facts for its job, and nothing else. This is what ours connect to:
| Agent | Connects to | Why | Level |
|---|---|---|---|
| Support drafts | Helpdesk (Intercom), Shopify orders | Read the ticket, check the order | Drafts only |
| Ad reporting and buying | Meta ad account, the books | Compare what Meta claims to real orders | Acts after approval |
| Bookkeeping | Accounting system (Finaloop) | Categorize, reconcile, flag anomalies | Acts after approval |
| Marketing calendar | Klaviyo, Shopify, Meta Ads, Google Ads, PostHog | Plan against past sends, sales and traffic | Drafts only |
| SEO and AI search | Search Console, PostHog | Track clicks and AI visits | Drafts only |
| Fulfillment watch | Shopify orders | Daily report of orders past their shipping window | Reports only |
Notice what is missing. Nothing in this table sends anything to a customer on its own. The support agent writes drafts a person sends. The ad agent proposes changes a person approves.
Should an agent get read or write access?
Read, almost always, at least to start. Most platforms let you split the two. On Shopify, read_orders gives read access and write_orders gives read and write (Shopify Dev Docs). On Meta, ads_read lets an app pull ad reports, while ads_management lets it read and manage ads (Meta for Developers). Klaviyo offers a read-only private key, a full access key, or a custom key where you set no access, read or read/write per scope (Klaviyo Help Center).
We tie access to an autonomy level:
- Drafts or reports only. Read access everywhere. The output goes to a person.
- Acts after approval. Write access to one system, and every action waits for a yes.
- Acts within limits. Write access with a hard cap, like a daily spend limit. We have not given this to anything that spends money.
Our agent registry even has a planned agent whose KPI is zero unapproved actions. If an agent ever does something no one approved, it should drop a level.
What goes wrong when an agent has the wrong access?
It reports a confident number built on missing data. Three examples from our own setup:
1. The order window was too short. Shopify's order scopes only cover orders from the last 60 days unless you also request read_all_orders (Shopify Dev Docs). Our bookkeeping agent traced store credit that was counted twice across 384 orders, about $49,000 year to date. A 60-day window would have shown only the last two months of it and made the problem look smaller than it was.
2. The ad accounts were not connected to the books. Our accounting system had the store, but not the Meta and Google ad reports. So ad spend came from card payments, dated by when we paid instead of when the ads ran. The four-month totals were right, but single months could shift. The fix is to print that caveat on every ROAS report until the accounts are connected, as we do in Blended vs new-customer ROAS.
3. Analytics was connected, but to the wrong event. Our store logs a custom page view event, not the analytics tool's default one. Queries on the default event came back empty. On top of that, counting AI visits by referrer alone found 0 ChatGPT sessions in October, and counting by UTM tag as well found 65. Both were "connected." Only one was right. The details are in How do you track AI search traffic to a Shopify store?.
None of these were model errors. Each was a connection that technically worked and quietly returned the wrong slice.
How do you check an agent's access before you trust it?
Run one question you already know the answer to. Before you trust an agent's first real report, check three things:
- Can it see what it needs? Ask for a number you can verify by hand, like last month's order count or ad spend. If it is off, find out why before going further.
- Can it see more than it needs? List the granted scopes. Shopify lets an app query exactly which scopes an installation has (Shopify Dev Docs). Remove anything the job does not use.
- Does it say what it could not see? A good report names its gaps, like "ad accounts not connected" or "orders older than 60 days not included."
Also check customer data rules. Shopify withholds protected customer fields like names and addresses until an app is approved for that data, even when it has the scope (Shopify Dev Docs). A support agent that cannot see a customer's address may draft a vague shipping reply.
Which tool should you connect first?
The one behind your highest-value repeat task. For most stores that is Shopify orders plus the helpdesk, or Shopify plus the ad account. Connect it read-only, run the known-answer test, and let the agent report for a few weeks before adding a second system. For choosing that first task, see Which tasks should an eCommerce brand hand to AI agents first?
Common questions
Is it safe to give an AI agent my Shopify admin login?
No. Give it a scoped app or API token instead of a staff login. Shopify scopes let you grant order reads without product writes, and you can see exactly what was granted. A shared login gives the agent everything a person could do, with nothing to audit.
Does the agent need access to my email or Slack?
Only if that is where its work gets delivered or where the facts live. Most of our agents post a report or a draft somewhere a person will see it. Reading a whole inbox is rarely needed and pulls in a lot of data the agent has no use for.
How many tools should one agent connect to?
As few as the job needs, usually two to five. Our support agent uses two. Our marketing calendar agent read five for its holiday plan, because a plan has to check email, orders, both ad accounts and site traffic. If an agent needs ten, it is probably two agents.
What should happen when a connection breaks?
The agent should say so in its report instead of filling the gap with a guess. A missing data source should show up as a line like "ad accounts not connected, spend taken from payments" so the person reading it knows what the number can and cannot tell them.