Blog

Why Your AI Employee Can't Access a Tool (and How to Fix It)

The Spinrun team

A practical approach to checking which person, account, and tool may act, using explicit sources, a reviewable output, and the controls your workspace actually offers.

Choose a job you can verify

For access review, the useful starting point is checking which person, account, and tool may act. Define that work before choosing a product or a model. A polished answer is easy to notice; an omitted record or an action on the wrong account can be much harder to find. The reviewer needs a specific output and enough context to check it.

Use the workspace role, connection scope, and intended operation. Decide what belongs in the final result and which decisions stay with a person. Keep the first attempt small enough that a teammate can compare it with the original records. That gives you a way to improve the process before it becomes recurring or widely shared.

Define the input and the finish line

The target output is an access checklist and a safe test of a known read. Include the reporting period, account, audience, and source locations in the request. If the agent cannot reach one of those sources, ask it to identify the gap instead of guessing or quietly substituting old information.

Write the output standard in terms a reviewer can observe. A result should be on time, grounded in the agreed material, and clear about the next action. Avoid judging only its tone. Confidence and completeness are different properties, and an accurate statement can still be irrelevant to the decision you need to make.

A brief you can adapt

Use this as a starting point and replace the example's sources and audience with your own. It describes a draft and preserves a human review point.

Text
Review the access needed for this workflow. Identify each connection, reader, and external write. Propose the smallest useful scope and a review point before any change. Do not widen permissions.

Read the brief as if another teammate will maintain it next month. If it relies on an account name, date range, or business rule known only to you, make that detail explicit. A task can use only the connections and permissions actually available in its workspace.

Connect the smallest useful scope

Start with a read that you can recognize. Confirm the provider account and the records returned before asking for a write. Access to an integration does not mean every teammate should use every action. Workspace roles, connection scope, and tool modes all matter.

For this workflow, pay particular attention to workspace scope, credential exposure, and sensitive writes. An instruction in a document or a provider response is source material, not permission to widen access. Keep the approval boundary in the task and use the product's permission controls for standing rules.

Review evidence, then approve actions

Check a small sample against the original source, including an ordinary record and an exception. Look for omitted rows, incorrect dates, and unsupported conclusions. Require a clear explanation when information is missing. The agent's result should help the reviewer decide, rather than conceal the decision behind fluent prose.

Review an external action separately from its draft. Confirm the recipient, account, record, and effect before accepting it. If a write may already have reached a provider, inspect the destination before repeating it. Retrying uncertain work can create duplicate messages or records even when the original response looked like a failure.

Make the working version repeatable

Once the manual process is useful, save it as a skill with its inputs, steps, exceptions, and output standard. Assign a person to maintain it. A reusable instruction should describe the team's working rule rather than the accidental details of one example.

Schedule the workflow only after a successful reviewed run. State its timezone and date-range rule. Define what it should report when a source fails and when unchanged state should stay quiet. Review the first scheduled result and revisit the instruction when the underlying process changes.

Measure the work, not the promise

Record how long preparation and review took before and after the workflow. Include corrections and maintenance in that comparison. Count whether the output arrived on time and whether the reviewer could act on it. These observations are more useful than an assumed percentage improvement.

This article describes an illustrative process, not a customer result or a benchmark. Actual usefulness depends on your input quality, available tools, permissions, and review standard. Inspect the current workspace balance and pricing when evaluating cost; a task's scope and repeated work can change its consumption.

Try the next useful step

Choose one representative input for access review. Write the expected an access checklist and a safe test of a known read before starting. Run the draft, review its evidence, and change one specific part of the instruction where the result falls short.

The product guides explain setup and permissions. Use Academy for a walkthrough, and Help if a connection or account stops the task. A process earns recurring use when the team can understand it, inspect it, and maintain it.

Give your AI real hands

Connect an app once. Every agent you use can act in it, under rules you set.

Before anything loads

We would like to switch on analytics, so we can see which pages are worth keeping. Nothing has loaded yet and nothing will until you choose. The cookies that keep you signed in and remember your language are not part of this.

Read the Cookie Policy