Automate the collection, standardization and first draft of a weekly client report. Keep calculations, interpretation and delivery under human control. A reliable workflow does not ask AI to reconstruct a week from memory: it gives the model a small, traceable data package, blocks incomplete inputs and requires a person to approve the result.
This guide gives you a six-step workflow and a copyable bilingual package you can test without a paid tool. All names, dates and figures below are fictional.
When this workflow is a good fit
It works well for independent professionals, consultancies and small teams that send recurring updates from identifiable sources such as a project board, CRM export or approved spreadsheet.
Start here when:
- the report follows roughly the same structure each week;
- someone can identify the source for every metric and completed item;
- the report is a draft until an account owner approves it;
- a stopped run is safer than a plausible but incomplete message.
Do not start with fully automated delivery. Contracts, payments, regulated decisions and unsupervised client communication need controls beyond the workflow in this article. If you are still choosing a bounded first use case, see what to automate first. Before configuring the workflow, map the business process as it actually operates so its sources, exceptions and approval boundary are explicit.
Before you automate: define the reporting contract
Write down five things before selecting a model or automation platform:
- Period: the exact start and end dates, including timezone where it matters.
- Sources: the approved location for each type of fact.
- Metrics: their names, definitions and owners.
- Reviewer: the person accountable for checking and sending the report.
- Stop conditions: the problems that must block drafting or delivery.
This contract matters because generative systems can produce confident false content. NIST describes this risk as confabulation: confidently presented but false or erroneous output. A fixed structure can constrain the shape of an answer, but it does not establish that the underlying facts are true. OpenAI’s Structured Outputs documentation makes the same distinction: schema adherence can enforce format in supported models, while factual mistakes can still occur.
Copyable input contract
Use a stable input instead of pasting an unstructured collection of messages. This fictional example is intentionally small:
client: Example Studio
period_start: 2026-09-21
period_end: 2026-09-27
completed:
- item: Landing page published
source: Project board item 142
metrics:
- name: Qualified leads
current: 18
previous: 15
source: CRM export 2026-09-27
blockers: []
decisions_needed: []
next_steps: []
Required fields should be genuinely required. An empty blockers list means “confirmed: no blockers”; a missing blockers field means “unknown.” Preserve that distinction in your validation rules.
If your team works in spreadsheets, column types and data-cleaning features can help expose incompatible values, duplicates and inconsistencies before analysis. Google Sheets documents these capabilities, but they are preparation aids—not proof that a number is correct and not a replacement for source ownership.
The six-step workflow
1. Collect
Retrieve the reporting period, completed work, confirmed metrics, blockers, decisions and next steps from the approved sources. Record a source beside every completed item and number. Do not use the model to recover missing facts from old conversations.
2. Normalize
Map the collected information to the input contract. Standardize date formats and metric names. Keep 0, an empty list and a missing value distinct: each communicates something different.
3. Validate
Run deterministic checks before AI sees the data. Confirm that:
- the period is complete and does not overlap the previous report;
- every required field exists;
- every metric and completed item has a source;
- numeric fields contain valid values in the expected unit;
- the client identifier matches the intended recipient;
- any change outside your agreed threshold is flagged for a person.
If a check fails, stop. Do not ask AI to repair the input by guessing.
4. Calculate
Calculate totals, differences and percentages with spreadsheet formulas or code whose logic you can test. Store the calculation and its inputs. Give the model the result; do not rely on free-form generation for business-critical arithmetic.
For the fictional metric above, the input may contain the confirmed current and previous values. Whether a percentage change belongs in the report is a human reporting decision, not an invitation for the model to infer one.
5. Draft
Send only the validated package and the restricted drafting rules below. Ask for a predictable report structure. If your chosen model supports schema-constrained output, it can help enforce that structure—but the draft still requires factual review.
6. Approve
The account owner compares the draft with the sources, reviews interpretations, resolves every missing marker and approves the exact message. Only then may the report enter the team’s normal delivery process. Drafting permission is not sending permission.
Reference architecture: separate data, drafting, and delivery
A weekly reporting automation is easier to test when it is not treated as one opaque block. Split it into three layers with observable outputs:
| Layer | Input | Verifiable output | Failure that must stop the run |
|---|---|---|---|
| Data | Approved exports and records | Normalized package with period, sources, and client identifier | Missing required field, wrong period, or item without a source |
| Drafting | Only the validated package | Structured draft that is not yet deliverable | Invalid output, new unsupported fact, divergent calculation, or [MISSING] marker |
| Delivery | Approved draft and confirmed recipient | Message queued or released through the authorized process | Missing approval, recipient mismatch, or version changed after review |
This separation keeps a delivery credential from turning a collection error into automatic external communication. It also lets the team change the drafting tool without rebuilding collection and approval.
Copyable pseudoflow
schedule: every Monday, 08:00, in the defined timezone
collect approved_sources for report_period
normalize into report_input
stop_if required_field_missing
stop_if source_reference_missing
stop_if client_id_mismatch
stop_if reporting_period_overlap
calculate metrics with tested formulas
draft from validated_input using restricted_rules
stop_if draft_invalid
stop_if draft_contains_missing_marker
stop_if draft_adds_unsupported_fact
create approval_task for account_owner
wait for explicit approval of exact_version
stop_if recipient_mismatch
stop_if approved_version_changed
release through authorized delivery process
log run_id, input_version, approval and delivery_status
This block describes responsibilities rather than ready-made commands for a particular platform. Each stop_if must become a deterministic condition or a recorded human decision. The wait must not become a fixed delay: passage depends on approval of the exact version.
Minimum evidence for each run
Keep a compact record without unnecessarily duplicating client data:
- run identifier and reporting period;
- version or timestamp of collected sources;
- result of every validation and calculation;
- draft version shown to the reviewer;
- approver, decision, and approval time;
- confirmed recipient and final delivery state;
- reason for any stopped run.
The purpose is to distinguish “did not run,” “stopped safely,” “awaiting approval,” and “delivered.” Without those states, a team may resend an uncertain run or interpret silence as success.
Copyable restricted drafting rules
Use only the data provided.
Do not create numbers, causes, deadlines or owners.
If required information is absent, write [MISSING: field].
Separate confirmed facts from interpretation.
Produce a draft for human review; do not send or publish it.
Add the requested language and headings outside this block. Avoid instructions such as “make this sound more impressive”; they reward fluency over traceability.
When not to use this workflow
Do not use this workflow when the input cannot be traced to approved sources, no accountable reviewer is available, or an incorrect report could directly trigger a payment, contract, regulated decision or irreversible action. It is also unsuitable when client data cannot be shared with the selected services under the organization’s authorization, retention, access and contractual rules. In these situations, keep the process manual or design stronger controls with the responsible legal, security and operational teams before introducing AI.
Weekly report template
# Weekly update — [client] — [period]
## Executive summary
- Confirmed outcome:
- Attention required:
## Completed work
- [item] — Source: [source]
## Confirmed metrics
| Metric | Current | Previous | Source |
| --- | ---: | ---: | --- |
| [name] | [value] | [value] | [source] |
## Risks and blockers
- [confirmed blocker or "None reported"]
## Decisions needed
- [decision, owner and real deadline if supplied]
## Next steps
- [supplied next step]
## Reviewer notes
- [interpretation or context added by the human reviewer]
“None reported” is not the same as an empty generated section. Use it only when the input explicitly confirms there are no items.
Fail closed: five conditions that stop the workflow
The system should stop and notify the responsible person when:
- an approved source is unavailable;
- the reporting period is incomplete or ambiguous;
- a number or completed item has no traceable source;
- a difference exceeds the team’s predefined review threshold;
- the generated output is empty, malformed or still contains
[MISSING: ...].
The threshold in item four is a team control, not a universal percentage. Define it per metric and client context. A stopped run is visible; a polished, silently wrong report may not be.
Troubleshoot failures: do not resend before you know the state
A missing report may have stopped during collection, validation, approval or delivery. Treating every case as “failed, try again” creates duplicates and may send a version different from the one that was approved. Identify the last proven state first.
Use a small state machine: scheduled → collecting → validated → drafted → awaiting_approval → approved → queued → delivered. A run may also end or wait in blocked_data, blocked_review, delivery_unknown, failed_confirmed or cancelled. “No error appeared” does not turn delivery_unknown into delivered.
| Symptom | First hypothesis to test | Minimum evidence | Safe state | Next action |
|---|---|---|---|---|
| The report did not appear | Trigger did not run, or time/timezone is wrong | run_id, scheduled time, timezone and trigger history |
delivery_unknown |
Inspect the trigger; do not start another delivery |
| The latest data is missing | Processing window, period or filter is wrong | Source timestamp, extracted range and filter version | blocked_data |
Wait for the defined window or rerun collection only |
| Figures differ from the source | Stale snapshot, unit or formula mismatch | Source version, inputs and calculation version | blocked_data |
Reconcile one metric end to end before drafting |
| The draft adds an unsupported cause | Drafting-rule breach | Validated input, applied rules and output version | blocked_review |
Reject the draft; do not complete the story by inference |
| Delivery is not confirmed | Queue, provider, recipient or permission problem | Delivery ID, provider state, approved recipient and version | delivery_unknown |
Reconcile provider state before retrying |
| Two copies were created | Non-idempotent retry or version mismatch | run_id, delivery ID, period and approved version |
failed_confirmed |
Stop retries and designate the authoritative copy |
| The creator lost access | Invalid owner account or revoked authorization | Owner and credential audit | blocked_data |
Transfer or recreate under an active accountable owner |
| The link or attachment cannot be opened | Permission, licence or unsupported export | Recipient-view test, export result and access rule | blocked_review |
Choose an authorized distribution path; do not bypass access |
Evidence packet before any retry
Keep one incident record containing the run identifier; period and timezone; source versions and timestamps; validation results; calculation version; hash or exact version of the approved draft; minimized recipient identifier; provider delivery ID and state; and the last confirmed transition. This packet lets you retry only the failed stage without recalculating or rewriting approved work.
Retry only after failure is confirmed and the destination or adapter actually enforces deduplication. Reuse the same stable idempotency key for the same delivery and create a distinct attempt_id for each attempt. A run_id provides traceability; it prevents duplication only when the destination explicitly applies it as the idempotency key. Never retry while delivery is delivery_unknown; first establish whether the provider accepted the message. If only delivery failed and the approved version exists, reuse that version—do not generate a new draft. Do not switch recipients as a troubleshooting shortcut.
Native reporting services have their own delays and states. GA4 documentation says events may take up to 72 hours to process and that reports sent near the start of a month may omit the final three days of the previous month; it also distinguishes states such as pending, expired and invalid creator. Power BI troubleshooting covers stopped delivery, refresh-dependent runs, inaccessible attachments or links, and permission requirements. These documented behaviors form hypotheses; the actual state still has to be confirmed in the service you use.
Test six failures before the pilot
Using fictional data, simulate: late source data; wrong period or filter; creator access loss; a timeout after queueing delivery; a duplicate retry; and a draft changed after approval. The test passes only when every scenario reaches a distinguishable state, preserves the approved version and prevents blind resending.
Privacy gate before using real client data
Keep the first test fictional or redacted. Before sending real client information to any service, confirm with the responsible person in your organization:
- authorization and a defined purpose for the use;
- data minimization: only indispensable fields are included;
- retention policy and whether submitted data may be used for training;
- processing and storage region where relevant;
- which people, integrations and providers can access the data;
- the contract and data processing agreement required by your organization or jurisdiction.
The European Commission lists purpose limitation, data minimization and storage limitation among the GDPR principles. These are useful design checks, but this article is not legal advice. Review the current official policy of every provider you use. When authorization or terms are unclear, keep real data out of the workflow and use fictional or redacted inputs.
Test it for four reporting cycles
Do not declare success after one polished draft. Run the same scorecard for four cycles:
| Measure | What to record each cycle |
|---|---|
| Manual baseline | Minutes needed to produce the report with the existing process, measured when a comparable manual run is available |
| Review time | Minutes from draft receipt to approval, excluding unrelated waiting |
| Factual corrections | Count of changes to names, dates, numbers, statuses or source-backed statements |
| Silent failures | Count of missing, duplicated or incorrect items that the workflow did not flag |
Also note stopped runs and why they stopped. A useful workflow may stop on bad input; that is different from silently passing it.
After four cycles, compare the record. Continue only if factual corrections and silent failures are acceptable to the accountable reviewer. If review remains difficult, improve the source data and validation rules before rewriting the prompt. This scorecard does not prove ROI or guarantee time savings; it gives your team evidence about its own process.
Human review checklist
Before approving a report, verify:
- client and reporting period;
- each completed item against its cited source;
- every metric, unit and calculation;
- the distinction between fact and interpretation;
- blockers, decisions, owners and supplied deadlines;
- removal or resolution of all missing markers;
- tone, confidentiality and intended recipient;
- the exact final version that will be sent.
Frequently asked questions
Which tool should automate a weekly report?
There is no universal tool. Choose the architecture by locating where the manual work actually occurs:
| Current situation | Proportional starting point | What remains human |
|---|---|---|
| Data already lives in one simple spreadsheet | Spreadsheet + scheduled Apps Script trigger or a visual Make scenario | Check sources, exceptions and the final message |
| The report is mainly a metrics dashboard | A scheduled GA4 or Power BI report, when plan and permissions allow it | Interpret changes and decide actions |
| Data comes from several tools and must become a narrative | Automation for collection and normalization + AI only for the first draft | Validate figures, context, recipient and delivery |
Start with the option that adds the fewest components. Apps Script supports time-driven triggers, but they run under the creator’s account, require authorization and are subject to quotas. GA4 can schedule PDF or CSV reports; Power BI supports email subscriptions, but permissions, licensing and external-recipient rules can constrain the design. None of these features replaces the data contract and approval gate in this guide.
Five-question decision matrix
Answer each row before selecting a tool:
| Question | If yes | If no |
|---|---|---|
| Do the metrics already exist in one trusted source? | Avoid another integration layer | Plan collection and reconciliation |
| Does the recipient need numbers only, without narrative? | Prefer a native report subscription | Use a structured draft |
| Are clients or other external recipients involved? | Validate licensing, sharing and privacy | Internal distribution is usually simpler |
| Could a delivery mistake trigger a wrong decision? | Require human approval before sending | Still log the failure and owner |
| Can the team maintain scripts and credentials? | Apps Script or code may be proportional | Prefer a visual automation or native feature |
The result does not select a vendor automatically. It removes architectures the team cannot safely operate and shows what the pilot must prove.
Do I need a paid AI or automation tool?
No. You can test the contract manually with a spreadsheet, a text file and any model your organization has approved. Tool choice comes after the workflow is clear.
Can I use a spreadsheet as the source?
Yes, if ownership, definitions and validation are clear. Column types and cleanup features can expose some inconsistencies; they do not verify the business meaning of a value.
Can the system send the report automatically?
Not in the first version of this workflow. Keep delivery behind human approval until your own evidence supports a different control design. Even then, sensitive or consequential communication may still require approval.
Why is the latest data missing?
Check the period, timezone, filters and source refresh time first. Some platforms process data with a delay; GA4 events, for example, may take up to 72 hours. Put the run in blocked_data, record the gap and rerun only collection after the defined window—never invent the missing days.
The email did not arrive. Can I resend it?
Only after confirming that the first delivery failed. If the provider accepted the message but the final outcome is unknown, use delivery_unknown and investigate by delivery ID. Retrying in that state can create a duplicate. Once failure is confirmed, reuse the exact approved version and the same destination-enforced idempotency key; give each attempt its own attempt_id.
Who should own the schedule?
An active institutional account with a named owner and transfer process. Avoid depending on one person’s personal account: triggers and subscriptions can stop when the creator loses access, authorization or a licence. Include ownership verification in the workflow’s periodic review.
What about confidential or personal data?
Use fictional or redacted data until your organization has confirmed purpose, authorization, minimum fields, retention, access, processing location and contractual requirements.
Which process should I automate first?
Choose a repeated task with identifiable inputs, visible errors and a reversible outcome. Weekly reporting can fit when a person remains accountable for the final message. For a broader selection method, read what to automate first.
After four measured cycles, use the automation ROI calculator to compare observed time and cost with the original scenario without treating released capacity as cash automatically.
Before any report is sent, define the human approval boundary. Keep the workflow’s stop_if conditions connected to an automation failure contingency plan so a failed or uncertain run moves to a controlled path instead of being resent blindly.
Sources
- National Institute of Standards and Technology (NIST), Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. Supports the description of confabulation risk. Consulted September 27, 2026. https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=958388
- OpenAI, Structured model outputs. Supports statements about schema adherence, the supported subset of JSON Schema and the continuing possibility of factual mistakes. Rechecked October 13, 2026. https://developers.openai.com/api/docs/guides/structured-outputs
- Google Docs Editors Help, Use Sheets Smart Cleanup to prepare your data for analysis. Supports statements about detecting duplicates, inconsistencies and anomalies during preparation. Consulted September 27, 2026. https://support.google.com/docs/answer/10098582?hl=en
- Google Docs Editors Help, Use tables in Google Sheets. Supports statements about column types and warnings for incompatible data. Consulted September 27, 2026. https://support.google.com/docs/answer/14239833?hl=en
- European Commission, Principles of the GDPR. Supports statements about purpose limitation, data minimization and storage limitation. Consulted September 27, 2026. https://commission.europa.eu/law/law-topic/data-protection/information-business-and-organisations/principles-gdpr_en
- Google Developers, Installable Triggers. Supports statements about time-driven triggers, creator-account execution, authorization and quotas. Rechecked October 13, 2026. https://developers.google.com/apps-script/guides/triggers/installable
- Google Analytics Help, Schedule reports. Supports statements about scheduling, processing, states and report troubleshooting. Rechecked October 19, 2026. https://support.google.com/analytics/answer/13722168?hl=en
- Microsoft Learn, Email subscriptions for reports and dashboards in the Power BI service. Supports statements about subscriptions, permissions, licensing and external-recipient constraints. Rechecked October 13, 2026. https://learn.microsoft.com/en-us/power-bi/collaborate-share/end-user-subscribe
- Microsoft Learn, Troubleshoot Power BI subscriptions. Supports documented hypotheses for stopped delivery, refresh, attachments, links and permissions. Consulted October 19, 2026. https://learn.microsoft.com/en-us/power-bi/collaborate-share/service-troubleshoot-subscribe
Material-claim traceability appendix
The publication date is marked “not displayed” when the cited page does not expose one clearly. The original sources were checked on September 27, 2026; OpenAI and Apps Script were rechecked on October 13, 2026; GA4 and Power BI were rechecked on October 19, 2026. The remaining sources retain the historical date shown in each row.
| Material claim | Entity and primary source | Publication date | Consulted | Passage or location to verify | Verification owner | Status |
|---|---|---|---|---|---|---|
| Generative AI can produce confidently presented false or erroneous content, described as confabulation. | NIST — https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=958388 | July 2024 | September 27, 2026 | Section 2.2, “Confabulation” | Independent Marketing Validation Specialist — Espadas | Independently verified on September 27, 2026 |
| Structured Outputs can enforce schema adherence in supported models, uses a supported subset of JSON Schema, and does not eliminate factual mistakes. | OpenAI — https://developers.openai.com/api/docs/guides/structured-outputs | Not displayed on cited page | October 13, 2026 | “Structured Outputs” guide; schema-support and limitations passages | Independent Marketing Validation Specialist — Espadas | Independently revalidated on October 13, 2026 |
| Sheets Smart Cleanup can help identify duplicates, inconsistencies and anomalies before analysis. | Google Docs Editors Help — https://support.google.com/docs/answer/10098582?hl=en | Not displayed on cited page | September 27, 2026 | Smart Cleanup overview and cleanup suggestions | Independent Marketing Validation Specialist — Espadas | Independently verified on September 27, 2026 |
| Google Sheets tables support column types and warnings for values that do not match the selected type. | Google Docs Editors Help — https://support.google.com/docs/answer/14239833?hl=en | Not displayed on cited page | September 27, 2026 | “Use column types” and data-warning passages | Independent Marketing Validation Specialist — Espadas | Independently verified on September 27, 2026 |
| GDPR principles include purpose limitation, data minimization and storage limitation. | European Commission — https://commission.europa.eu/law/law-topic/data-protection/information-business-and-organisations/principles-gdpr_en | Not displayed on cited page | September 27, 2026 | “What data can we process and under which conditions?” principles list | Independent Marketing Validation Specialist — Espadas | Independently verified on September 27, 2026 |
| Apps Script provides time-driven triggers that run under the creator account, require authorization and are subject to quotas. | Google Developers — https://developers.google.com/apps-script/guides/triggers/installable | September 3, 2026 | October 13, 2026 | “Restrictions” and “Time-driven triggers” sections | Independent Marketing Validation Specialist — Espadas | Independently revalidated on October 13, 2026 |
| GA4 supports scheduled PDF or CSV reports; events may take up to 72 hours to process; reports near the start of a month may omit the previous month’s final three days; and the service distinguishes states such as pending, expired and invalid creator. | Google Analytics Help — https://support.google.com/analytics/answer/13722168?hl=en | Not displayed on cited page | October 19, 2026 | “Choose date” and “Report format” fields; “View and manage scheduled reports” > “Statuses” | Independent Marketing Validation Specialist — Espadas | Independently revalidated on October 19, 2026 |
| Power BI supports email subscriptions with permission and licensing requirements, and documents troubleshooting for stopped delivery, refresh-dependent runs, attachments, links and access. | Microsoft Learn — https://learn.microsoft.com/en-us/power-bi/collaborate-share/end-user-subscribe and https://learn.microsoft.com/en-us/power-bi/collaborate-share/service-troubleshoot-subscribe | Not displayed on cited pages | October 19, 2026 | Subscription requirements and troubleshooting sections for delivery, refresh, attachments, links and permissions | Independent Marketing Validation Specialist — Espadas | Independently revalidated on October 19, 2026 |