Explainer
TL;DR
An agency can connect an AI assistant to client Amazon Ads accounts safely when four things are settled before anyone types a prompt: each connection reaches one client's ad accounts and only the actions the job needs, client data runs on a business AI plan, the client contract allows AI use, and every change to spend passes a preview plus the agency's own sign-off.
An agency can connect an AI assistant to client Amazon Ads accounts safely when four things are settled before anyone types a prompt: each connection reaches one client's ad accounts and only the actions the job needs, client data runs on a business AI plan, the client contract allows AI use, and every change to spend passes a preview plus the agency's own sign-off. An MCP (Model Context Protocol) server is the connector that lets ChatGPT or Claude read and change an ad account; other terms are defined in the Amazon Ads MCP glossary. This FAQ answers the 20 questions agency leads, client-services directors and IT reviewers ask most, and ends with a policy and a checklist you can copy. Answers about contracts are general information, not legal advice; review them with your legal counsel.
Control of an agency's AI setup is split across four owners, and most review delays happen when a team expects one owner to cover another's layer. Use this map to route each question below to the right person.
| Layer | What it controls | Owner at the agency | Where it is set |
|---|---|---|---|
| Ad platform access | Which client ad accounts and which read or write actions a connection can use | Ad ops lead or platform admin | The ads platform, per connection |
| AI assistant | Training, retention, who can add connectors, tool approval settings | IT or the AI workspace owner | Claude or ChatGPT admin settings |
| Contract | Whether AI use is allowed, what is disclosed, which vendors process client data | Client-services lead with legal | MSA, SOW, DPA |
| Team process | Who approves which changes, audits, offboarding | Account director | Your internal policy (template below) |
Each person should create their own connection, never share one. People connect to Xnurta MCP by signing in through the browser with their own Xnurta login, or with a token that is shown only once. A connection cannot do more than its creator's own Xnurta permissions allow, so it inherits only the client ad accounts that person already has. A shared connection breaks two things agencies need: you cannot tell one person's work from another's in the change log, and you cannot remove one person's access without cutting everyone's. Freelancers and offshore staff get their own Xnurta user with only the clients they work on.
Yes. A connection can be limited to selected ad accounts when it is created, so a buyer can hold one connection for Client A and a separate one for Client B, each with its own permissions and expiry date. Selecting accounts narrows access below the person's Xnurta role; it never widens it. The working rule: one connection per client ad account, named after the client, so a prompt typed in the wrong chat has no route to another advertiser's data.
It depends on what the role changes. Xnurta MCP has read permissions (Performance data, Ads configuration, Operation logs) and write permissions (Create and edit AI-managed groups, Delete AI-managed groups, Edit campaigns, Create campaigns), each ticked separately. Juniors and analysts hold the read permissions only, and seniors hold write permissions client by client. For what each permission unlocks task by task, see read-only vs read-write MCP.
No. Anthropic says that on Team plans only Owners and Primary Owners can add custom connectors, and each user then connects individually, so Claude reaches only what that user can reach. With Xnurta MCP, that sign-in is the person's own Xnurta login, so each member sees only the client ad accounts and permissions Xnurta already gives them. Anthropic also warns that a connector using a shared credential gives everyone that credential's access. In ChatGPT, OpenAI says workspace owners, admins and authorized members can add custom apps.
Not by default on business plans. Anthropic says its commercial products are not used for training by default, and OpenAI says it does not train on Business, Enterprise or Edu workspace content, including information accessed through apps. Rated responses are the common exception, so agencies with strict clients often turn chat rating off. The plan-by-plan detail is in is it safe to connect Amazon Ads to ChatGPT or Claude.
Xnurta receives the tool requests the assistant sends and the ad account each one is for; its logs record the tool call, not the question the buyer typed. The full conversation, uploaded files such as client margin sheets, and other connectors stay in the AI assistant. If a client's procurement team needs contract terms, ask Xnurta for its data processing terms and file them with the client's paperwork.
Retention depends on the vendor and plan, and business plans give admins control. Anthropic says Claude Enterprise owners can set custom retention with a 30-day minimum, and OpenAI says ChatGPT Business and Enterprise admins control how long data is kept. If a client contract sets a deletion window, match the workspace setting to it and record the setting in the client policy.
Cross-client exposure is blocked at the ad platform and managed in the chat. A connection scoped to Client A has no route to Client B's ad accounts. Inside the assistant, the risk is human: a buyer pastes Client A's margin file into a chat about Client B. Keep one project per client, name the ad account in every prompt (for account [X]), and never mix two clients' files in one project.
It depends on your contract. The ANA's model media agency contract, revised in 2023, requires agencies to obtain an advertiser's prior consent before using AI applications, as reported by MediaPost on June 6, 2023, and its July 2025 generative AI rider adds disclosure and human oversight, as reported by AdExchanger. Even if your contract predates these, ask in writing; a one-paragraph email naming the tool, the permissions and the approver usually settles it.
It may, and two places decide. The MSA or SOW may have clauses on AI use, subcontractors and confidential information. The DPA may list approved sub-processors, in which case the AI vendor and ads platform may need adding. Campaign data is rarely shopper-level, but operation logs name teammates, and margins and strategy are confidential, so check both privacy and confidentiality clauses.
Client work belongs on a business or enterprise plan the agency controls, never on staff members' personal accounts. For Claude, that is Team or Enterprise; for ChatGPT, Business or Enterprise. These plans are not trained on by default, admins control connectors and retention, and access ends when someone leaves the workspace.
The person in the chat approves it, and your policy decides whether that is enough. For budgets, bids and archiving, Xnurta MCP reads the current state, lists each object old to new, runs only after an explicit yes, then reads back the result; the server enforces that step. Where your policy needs a second approver, the buyer sends the preview to the account director and confirms only after a yes.
Yes, at three levels on the platform side. Give a person read permissions only. Tick write permissions one by one, so a buyer can edit campaigns without creating them or deleting AI-managed groups. Sub-accounts get write permissions only when the main account grants them. Generally available from 27 Oct for Sponsored Ads. Write access is turned on per account by your CSM. For one client, use read-only connections. The AI workspace can add a fourth level: Claude Team and Enterprise owners can set connector tools to Always allow, Needs approval or Blocked, and ChatGPT offers settings such as Always ask.
Archiving, new campaigns, deleting an AI-managed group and large budget or target changes should always get a second look. Pause instead of archiving, because archiving on Amazon cannot be reversed.
No longer than the work it supports. Xnurta MCP connections expire after 30, 90 or 180 days or 1 year, chosen at creation:
| Engagement | Suggested expiry | Why |
|---|---|---|
| Pitch audit | 30 days | Read-only, ends with the pitch |
| Contractor or freelancer | 30 or 90 days | Matches the statement of work |
| Ongoing retainer, write access | 90 days | Forces a quarterly review of who still needs write |
| Ongoing retainer, read-only reporting | 180 days or 1 year | Lower risk, fewer interruptions |
| Client in notice period | 30 days | Ends close to the exit date |
A Xnurta MCP token is shown only once, when it is created. If someone loses it, create a new connection and revoke the old one rather than passing a copy around.
Separation holds up when every layer points at one client. Give each client its own connection, its own Claude or ChatGPT project, and a clear naming convention:
| Layer | One per client | Example |
|---|---|---|
| MCP connection | Yes, per person | [Client] Amazon read, or [Client] Amazon edit |
| AI project | Yes | [Client] Amazon Ads |
| Files | Only that client's | [Client] margins Q4 |
| Prompt | Names the account | For account [X], compare last week with the week before. |
| Skill | Shared across clients for a consistent format | Weekly Ads Report |
Installed Skills keep report formats consistent across clients; the reporting routine is in agency client reporting with MCP.
Start with Xnurta's operation logs, which label each change by operator: a person, Xnurta's AI or an automation rule, with changes from the Amazon console, an API connection or a named teammate also identified. Each entry holds the old and new values, so you can show a client exactly what moved. Each log request covers up to 90 days, about 15 months back. A weekly read-only habit: for account [X], list every change in the last 7 days by operator, and flag any budget change over [20]%. The full routine is in who changed my Amazon campaign.
Close each layer of their access on the same day:
| Step | Where | Owner |
|---|---|---|
| Revoke their MCP connections; revocation takes effect immediately | Xnurta platform | Platform admin |
| Remove or reduce their Xnurta access | Xnurta platform | Platform admin |
| Remove their Claude or ChatGPT seat | AI admin settings | IT |
| Have the main account take over their Skills | Xnurta Skill Hub | Account director |
| Review the last 30 days of their changes | Operation logs | Account director |
Connections also stop at their expiry date, and none can do more than its creator's Xnurta permissions allow.
Xnurta MCP is for Xnurta customers, so the client ad accounts must already be managed in Xnurta under your agency's access. Each staff member needs a Xnurta login with the right client ad accounts and a Claude or ChatGPT seat on your business workspace. Connections are created directly in the Xnurta platform, and a connection can be used from Claude or ChatGPT.
Pilot on one or two clients who have already agreed to AI use, read-only, for two weeks. Check answers against the Xnurta numbers you already report, using days at least 14 days old because attributed sales keep settling. Then give one senior buyer Edit campaigns and confirm every preview by hand. For clients in AI-managed groups, the next step is steering a group's goal instead of editing campaigns one by one; see AI-managed group levers and personality. Each group has an AI on/off switch in the Xnurta platform. Widen only when the operation log shows no surprises.
A client-account MCP policy should fit on one page and name an owner for every rule. Copy this template, replace the bracketed values and attach it to the client's onboarding file. General information, not legal advice: review it with counsel before adopting it.
Client-account MCP policy template
CLIENT-ACCOUNT MCP POLICY: [Client name] Owner: [Account director] Reviewed: [date] Next review: [date + 90 days] 1. Consent - Client approved AI-assisted analysis on [date] (ref: [email or contract]). - Client approved AI-assisted changes: [yes / no / read-only for now]. 2. AI workspace - Approved assistants: [Claude Team / Claude Enterprise / ChatGPT Business / ChatGPT Enterprise]. - Personal AI accounts: not permitted for this client. - Chat rating: [off / on]. Retention setting: [X] days. 3. Connections - One connection per person for this client; no shared connections. - Ad accounts in scope: [account names]. - Expiry: [30 / 90 / 180 days / 1 year]. - Named users and permissions: [Name] - Read (all) [Name] - Read + Edit campaigns [Name] - Read + Edit campaigns + Create and edit AI-managed groups - Delete AI-managed groups: [nobody / account director only]. 4. Change control - Every write: preview old to new values before confirming. - Second approver for: archiving, new campaigns, budget changes over $[500]/day or [20]%, target ACOS changes over [3] points. - Pause instead of archive. Create new campaigns paused. 5. Separation - One project per client. Every prompt names the ad account. 6. Audit - Weekly: operation logs for account [X] by operator. - Monthly: confirm the named-user list is still correct. 7. Offboarding - Revoke connections, remove Xnurta access and AI seat, hand over Skills, review the last 30 days of changes.
Run this pre-connection checklist once per client, before anyone creates a connection. Each line names who confirms it.
| # | Check | Confirmed by |
|---|---|---|
| 1 | Client agreed in writing to AI-assisted analysis, and separately to AI-assisted changes | Client-services lead |
| 2 | MSA, SOW and DPA allow the AI vendor and ads platform to process client data | Legal |
| 3 | Everyone on the account uses the agency's business AI workspace, with retention and rating set | IT |
| 4 | Each person has their own Xnurta login with only this client's ad accounts | Platform admin |
| 5 | Connections use selected ad accounts, the fewest permissions and an expiry date; juniors read-only | Ad ops lead |
| 6 | Second-approver rules and the weekly log review have named owners | Account director |
An agency should wait when consent, contracts or the client's setup are not ready. If the client has not agreed to AI-assisted changes, do read-only reporting and ask for change approval with a real example in hand. If contracts predate AI clauses, start with the clients whose paperwork is newest while legal updates the template. If a client wants AI inside the Amazon console, that is a different route from an MCP. If client accounts are already in Xnurta with AI-managed groups, use Xnurta MCP with one connection per client.
Start here. On one consenting client, create a read-only connection scoped to that client's ad account and ask: Which ad accounts can this connection see? Then ask for another client's account and confirm it refuses. That single test answers most IT reviewers' isolation question.
Xnurta MCP is for Xnurta customers. Learn more on the Xnurta MCP page, or book a demo to see per-client connections, the permission screen, a live old to new preview and the operation log by operator.
Next question: How should agencies use an Amazon Ads MCP for client reporting?
FAQ
Most agencies do not need a separate AI workspace per client. One business workspace the agency controls, with one project per client, keeps files and chats apart while admins manage retention and leavers in one place. Set up a separate workspace only when a client contract requires it, and record that in the client policy.
Start with the operation log, which shows the old value, the new value and who or what made the change. Edit the setting back, or partway back, after a preview, and tell the client what moved and why. Judge the outcome on data at least 14 days old before deciding the change failed.
Train staff on read-only access first. Have each new user run the isolation test on one consenting client, write two weekly reports with a senior checking the numbers against the console, and read real previews alongside a write holder. Grant write permissions only after a senior signs off in the client policy.
Review MCP access monthly and re-issue write connections quarterly. Each month, compare the named-user list in the client policy with the people working the account. Each quarter, let expiring connections lapse, re-create only the ones still needed and drop write permissions nobody used in the operation log.
Give everyone on that client a read-only connection with the read permissions and an expiry date, so the assistant can report and diagnose without changing anything. Hold write permissions back on that client's connections until the client approves changes in writing, and record it in the client policy.
No. Each person should create their own connection so the change log shows who did what and one person's access can be removed without cutting everyone's. A shared connection also blurs who approved a change, which matters when a client asks.
Xnurta MCP
Xnurta MCP is for Xnurta customers only. Create your MCP connection directly in the Xnurta platform. New to Xnurta? Book a demo and we'll walk through your account with you.