Give AI agents a job with boundaries

A weekly marketing report needs a source for each number, an explanation of what changed, and someone responsible for the next action. Giving an AI agent access to a spreadsheet doesn’t settle any of those questions. You still have to define the job.

For a founder with a small marketing team, start with one recurring task you can inspect. Write down its inputs, output, permissions, and completion check before connecting tools. An agent that drafts a report and an agent that edits your campaigns need different boundaries.

Which work needs an agent?

Use an agent when the task needs flexible decisions about what to do next. If the steps are already known, a fixed workflow may be easier to inspect and maintain.

Anthropic’s distinction between workflows and agents is useful here. Workflows follow predefined paths; agents choose their process and tool use dynamically. Its engineering guidance recommends starting with the simplest approach that serves the task.

For a marketing report, fetching an approved export and calculating changes can follow fixed rules. Explaining a gap may require interpretation. Start by separating those jobs, so a plausible explanation can’t quietly change the underlying data.

The marketing engineer role connects this work to brand, content, and demand. Decide which part of the report needs those skills before handing the whole thing to a system. The operator-record exercise helps you document which decisions the person retains.

What belongs in the job brief?

Define the output and the conditions for accepting it. A job title such as “reporting agent” leaves too much unanswered. Specify which data it can read and where its work should go.

Illustrative task brief. Prepare a draft marketing report from approved exports. This is a planning exercise, not a client implementation or a result.

Part of the job	What to write down
Input	The approved export, date range, metric definitions, and comparison period
Output	A draft with sourced observations, data gaps, and proposed next actions
Access	Read the approved files and save a draft in the agreed folder
Fact check	Each reported value matches its source, label, and period
Decision owner	A named marketer checks explanations and chooses actions
Stop condition	Missing definitions, mismatched periods, or inaccessible source files

The brief should fit on a page. If you can’t describe what a completed draft contains, resolve that before adding another tool.

Where should the agent stop?

Stop when it lacks the evidence or permission needed for the next action. A reporting task can flag a missing export. It shouldn’t fill the gap with a guessed number or rewrite campaign settings to make the chart look better.

OpenAI’s guide to human intervention recommends escalation when failure thresholds are exceeded or actions carry higher stakes. Apply that to the reporting job. Set a retry limit for a failed read, identify who receives the error, and keep publishing or campaign changes outside this draft-only brief.

Permissions belong in the connected tools as well as the prompt. Give the reporting task access to its source folder and draft destination, rather than every account your marketer can access. Record which credentials it depends on without putting credential values in the brief.

Keep the handoff visible. The marketer should receive the draft, its source references, and any unresolved gap. A completed-looking report with an invisible failed data fetch is harder to use than an unfinished draft that explains what’s missing.

How can you test the brief?

Run the same brief against inputs that exercise the decisions you care about. For the report exercise, include a complete export, a missing comparison period, and an undefined metric. Keep the expected response beside each test.

A missing comparison period should produce a gap note, not a growth percentage. An undefined metric should produce a question for its owner. A complete export should produce values you can trace back to the file, with interpretation kept separate from observation.

NIST’s 2024 generative AI risk profile identifies confidently stated false content as a risk. For this task, test for it directly by leaving a required input absent and inspecting whether the draft acknowledges that absence.

You can start with this instruction:

Prepare a marketing report draft from the approved export and definitions.
List each observation with its metric, period, value, and source reference.
Keep proposed explanations separate from the observed data.
If a definition or comparison period is missing, name the gap and the
person who needs to resolve it. Don't calculate an unsupported comparison.
Save the draft and gap list in the agreed folder. Don't publish the report
or change a campaign. Return the saved location and any failed steps.

Inspect the saved output. A success message alone doesn’t show that the expected file or source reference exists. For a page built from your terminal, that check includes the resulting page and its links.

What would make you keep the workflow?

Keep it when the checked output is useful enough to justify the work of running it. Record setup, checking, corrections, and maintenance alongside tool costs. Compare that with how you currently prepare the same report.

If the draft repeatedly mislabels periods, fix the input contract and rerun the affected tests. If its explanations need extensive rewriting, narrow the job to arranging verified observations. A smaller useful task is worth keeping.

Choose one report, write its brief, and test a missing-input case before connecting a live account. The AI true-cost worksheet helps you record the cash and hands-on work. If you need the connection and checked reporting output built as a defined project, Custom AI Builds is the relevant service to discuss. Recurring monitoring and maintenance are separately scoped through Marketing Automation.