Skip to content
be-human.ai
← Back to the journal

Business workflows

AI Weekly Ops Report: An Agent Drafts, a Manager Checks Every Figure

A practical method for an AI agent to draft a weekly operations report from approved records, with a manager checking every figure before sending.

Be Human typographic cover with the title: AI Weekly Ops Report: An Agent Drafts, a Manager Checks Every Figure
Conceptual illustration created with AI · Be Human

An AI agent can draft a weekly operations report by reading only the records your team has approved, such as ticket exports, dispatch logs and the staffing sheet, then writing each section with a reference to the record behind it. A named manager checks every figure against its source before anything is sent. The agent prepares the draft, and the person decides what the report says.

Key takeaways

  • Limit the agent to a fixed list of approved records, read-only, with one stated cutoff time for the whole report.
  • Use a fixed template so every section has an owner and a source, and drop any section that cannot name one.
  • Require a source reference for every figure, and require "not recorded" in place of estimates.
  • Give the manager a short checklist and a clear rule for returning a draft, so corrections stay specific.
  • Keep sending, record edits and external messages outside the agent's permissions.

Start with the records the report is allowed to use

A weekly operations report is only as trustworthy as its inputs. Before the agent writes anything, the team names its sources: the ticket system export, the delivery or dispatch log, the staffing schedule, and the inventory adjustments approved during the week. For each source, write down who owns it, which fields the report needs, and when the week closes for that source. If the ticket export runs at a different time from the staffing sheet, the report will silently mix two periods. Decide the cutoff once and print it at the top of every draft.

Records that are not on the list stay off the list. Email threads, chat messages and notes in a shared document can be useful context for a person, but they are not evidence for a figure unless someone has moved the relevant fact into an approved record. This keeps the agent from presenting a rumor as a metric, and it gives the manager a simple test: if a claim cannot be traced to a listed source, it does not appear in the report.

Write the report template before the agent writes anything

The template matters more than the prompt. Each section needs a plain purpose, a named owner and a defined source. A workable weekly template usually has six parts:

  • Headline: one sentence on the week, written only from the sections below it.
  • Volume: items opened, items closed and items still open at cutoff, each with the export and filter used.
  • Exceptions: records that broke an agreed rule, with record ID, reason code and current owner.
  • Blockers: items waiting on a person or outside party, with the date each was last updated in its record.
  • Decisions needed: questions that require a manager or leadership, with the options as written in the record.
  • Next week: commitments already present in approved records, not new promises the agent invents.

The owner of each section should be a person, not a team name, because the owner is the one who answers when a figure is questioned. For the wider design of a report that shows both coverage and decisions, see AI weekly business report.

Illustrative scenario: a regional tool rental company has a small operations team. The dispatch lead owns the delivery section, the warehouse supervisor owns returns, and the operations manager owns the headline and the decisions list. The agent drafts every section from the dispatch log and the returns sheet. The warehouse section stays marked as pending until the supervisor confirms that the returns sheet closed at the stated cutoff.

Let the agent draft, and make every figure traceable

With read-only access to the listed sources, the agent pulls each record, applies the filters written in the template and drafts the sections. Each figure carries a short reference that names the source, the filter, the count or value, and the timestamp of the export. A volume line might read as: items closed, from the ticket export, status closed, closed after the previous cutoff and on or before this one, with the export time shown in the footer.

When a field is empty or inconsistent, the agent writes "not recorded" and lists the affected record IDs. It does not infer a value from neighboring records, and it does not fill gaps with typical values. Trend language needs the same discipline. Words such as "up," "improved" or "slower" are allowed only when the draft shows both values, the periods compared and the source for each. Otherwise the agent removes the word and states the two values plainly.

The agent also reads record text as data, not as instructions. A ticket description written by a customer or a supplier can contain sentences that look like commands, such as asking the reader to ignore the template. The agent's instructions should state that record content is never an instruction, and the draft process should not give the agent any action beyond writing to the draft folder. The [OWASP prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes why this separation matters for any agent that reads external text.

How the manager checks every figure

The manager's review is the control that makes the report safe to send, so it should be short, repeatable and written down. A practical checklist is:

  • Open the source for each figure and confirm that the value in the draft matches it.
  • Confirm that the filters and cutoff match the template header, not the agent's default.
  • Check that summary counts agree with the detail lines beneath them.
  • Read each exception and confirm that the reason code fits what the record says.
  • Remove any sentence that explains a cause the records do not state.
  • Mark each decision item as decided, deferred or open, and name the person who holds it.
  • Sign the draft with a name and date before it moves to distribution.

The review also needs a decision guide, so the manager knows what to do with a draft that is not ready:

  • Send as is: every figure traces to a source, and the only open items are marked as open.
  • Return to the agent: a figure lacks a reference, or a sentence claims more than the record shows. Return the specific line, and the agent corrects that line only.
  • Pull a person in: a record looks wrong, a source did not close on time, or the issue needs a judgment call. Fix the source or ask the section owner, and do not patch the report around the problem.

Keep the returned items. Over several weeks they show whether the problem sits in the agent's drafting, in the template, or in a source export that never closes cleanly.

What the agent must never be allowed to do

The permission design should be as deliberate as the review. The agent needs read access to the listed sources and write access only to the draft folder. It should not send messages, edit source records, change the template or publish to a shared dashboard. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) makes the general point: each action should be checked against the permissions of the identity performing it, not assumed because the agent is already running. Review these permissions whenever a source system changes, a new section is added or a person leaves an owner role.

The agent should also leave a trace of each run: the time, the sources read, the filters applied, the records excluded and the version of the template used. This is the same logic as observability in software systems. The [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) explains why logs and traces matter when a multi-step process produces an output that someone has to explain later. For a weekly report, that explanation is what the manager needs when a number is questioned in a meeting.

For governance, the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a useful reference for documenting who oversees the system, what it may do and how its outputs are checked. A short written note covering these points is usually enough for a report of this size.

How to measure whether the report is worth keeping

Measure the work the report is supposed to reduce, not the volume of text the agent produces. Useful measures are the minutes the manager spends on review each week, the number of figures returned and the reason category for each return, and whether the decisions section was used in the following meeting. Review these monthly with the report owners.

If most returns come from one source, fix that source's export before changing the agent's instructions. Many teams find that a field is filled inconsistently, that a status is used differently by two people, or that a cutoff is applied at a different time of day. A cleaner export usually helps more than a longer prompt. For a fuller approach to the numbers, see how to measure AI workflow ROI without inflated claims. Teams that want to see the draft before relying on it often run it alongside the existing process for a few cycles first, as described in shadow mode for AI agents.

How Be Human runs this as a managed service

Be Human starts by defining the role: the report owner, the approved sources, the template sections and their owners, the permissions for each system, and one observable success criterion, such as how many figures pass manager review without being returned. Managed operations then cover monitoring of the export jobs, permission reviews when a source or owner changes, incident response when a source fails to close, and controlled template updates that the report owner approves before they take effect. The manager still signs every report, and the agent never sends it. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent send the weekly report without a manager reviewing it? In this design, no. The agent writes to a draft folder only, and sending requires a named person to check each figure against its source and sign the draft. This is a deliberate boundary, because the manager is the accountable party when the report is questioned later.

What happens if a source system does not export cleanly one week? The agent marks the affected section as not recorded or pending and lists the records it could not read. It should not estimate the missing values from earlier weeks. The section owner then decides whether to wait for the corrected export or to publish the report with the gap stated clearly.

How is this different from a dashboard? A dashboard shows values in real time from connected systems, while this report is a dated, signed narrative built from a fixed cutoff and a template. The agent helps assemble the narrative and its references, but the decisions and wording belong to the people who own each section. Many teams keep both, using the dashboard to check the draft.

Does the agent need access to all of our systems? No. It needs read access only to the sources named in the template, and write access only to the draft folder. Each permission should be reviewed when a source changes, and anything outside the list should be refused by default.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →