Can your AI integration read, recommend or act? Ask before connecting it.

Before you connect a business system to an AI assistant, ask one plain question: can this connection only read, can it recommend, or can it act? Each answer carries a different level of risk, and the setup work, review and ownership should match it.

A dated example

On November 4, 2026, Adobe announced that CX Enterprise Coworker, its agent for customer experience work, is available inside AI platforms including ChatGPT (as a plugin), Claude, Databricks Genie, Gemini for business and Microsoft Copilot, with Amazon Quick, IBM and Snowflake expected later. Adobe also announced connectors to partner tools that it says let Coworker take action across systems. Adobe announcement, November 4, 2026 (opens in a new tab)

Adobe's setup documentation, updated the same day, describes the AI-platform experience more narrowly. There, Coworker returns read-only insights, does not modify customer data and inherits the data permissions tied to your Adobe identity. It requires Coworker to be provisioned, Adobe Experience Platform plus at least one of Real-Time CDP, Journey Optimizer or Customer Journey Analytics, a supported AI platform and, for some platforms, administrator approval. Adobe says there is no extra charge for the integration, but usage draws on existing AI credits. The documentation also includes setup pages for platforms the announcement lists as coming soon, so confirm availability for your own tenant and platform. Adobe Experience League: Coworker in AI platforms (opens in a new tab)

The gap between an announcement and its documentation is the useful part. One product can read in one place and act in another. So judge each connection on its own, using three levels.

One scenario, three levels

To keep this concrete, follow one hypothetical example. A marketing lead asks an assistant why a spring webinar campaign is converting below last month's. This is an illustration, not a client result.

Read: does it fetch the right information?

At this level the assistant only retrieves what you are allowed to see. In the scenario, it reports registrations and conversion by channel. Check its answer against the dashboard your team already trusts: same date range, same metric definition, same filters. Access is not the same as access to all company data. If the assistant inherits your permissions, it may simply not see the paid-media numbers that explain the drop, and it may not say so. Ask it what it looked at.

Recommend: does it explain its reasoning?

Next, the assistant suggests a change, such as moving budget from one channel to another or tightening an audience. A good recommendation states its assumptions, its uncertainty and the data behind it. Treat it as a proposal for a person to weigh. Being able to recommend a budget shift is not the same as having the right to make it, and your process should keep those two things apart.

Act: what happens when it changes something?

Some apps and integrations can write data, send messages, change budgets or edit audiences. In the scenario, imagine the team approves the budget shift and a connected tool carries it out. This level is a general one to assess for any integration. It is not a description of Adobe's read-only AI-platform experience. Before allowing it, require:

• Explicit scope: which accounts, campaigns and limits the action may touch.
• Approval at the moment of consequence, not only at setup.
• An audit trail of who asked, what changed and when.
• Protection against duplicates when a request is retried.
• A rollback plan and a named person to escalate to when an action fails partway.

A short checklist before you connect

• Source of truth: which report or system settles a disagreement?
• Data access and admin permissions: what can it see, and who approved that?
• Actions enabled: is it read-only, or can it write, send or spend?
• Review and owner: who checks the output, and who owns the result?
• Failure and recovery: what happens after a failed or partial action?
• Costs: software, API or usage credits, and the staff time spent reviewing.

A small pilot

Write five real campaign questions your team already answers by hand. Run them in an approved test environment, without real client data and without buying anything new. Compare every answer with the dashboard, and note where the assistant was right, wrong or missing context. Decide in advance what results would justify going further. Those thresholds belong to your team. No integration guarantees a return or savings.

Start at read, earn trust before recommend, and treat act as a separate decision with its own safeguards.

Go deeper

For a closer look at assistants that operate software directly, see our guide to testing computer-use AI agents.Read: what businesses should test before adopting computer-use agents

If you would like help deciding what your own AI integrations should read, recommend or do, take a look at Build With AI.Explore Build With AI