How can you check an AI-drafted customer reply?

Moves the unchanged drafting pilot from the company strategy article into a dedicated guide. Preserves the actual 23 October 2026 test and its limits.

Recorded 23 October 2026 · reviewed 23 October 2026 · AI-reviewed by a separate editorial agent

Answer in this version

Use the original material, explicit source precedence and an accountable reviewer. Keep sending and transactions outside this initial trial.

What remained unresolved

Realistic cases, human usability, failure records, total review effort and separately tested permissions before external action.

Takeaway and scope

Extract and check the facts first. Draft only from the checked material, then inspect every consequential promise before staff use the reply.

One fictional run, tested on 23 October 2026 and moved here unchanged. No human usability, live transaction or business improvement measured.

Workflow in this version

1. Keep facts accurate

   Choose one customer-facing offer. Record its current price and currency, availability, delivery, terms and evidence, including country or language differences. Name the authoritative source and owner. Compare an older page or channel against it; flag conflicts instead of silently averaging them.

   People and agents can repeat the same wrong fact. An owner must resolve a conflict before it becomes a promise.

2. Bound one job

   Pick a small task such as drafting a product-enquiry reply. Record approved inputs, result, owner, permitted actions, checks and stop conditions. The downloadable template keeps these together. Start with reading and drafting only. Configure tool permissions separately; a Markdown file cannot enforce them.

   A narrow scope lets you see both useful work and the points where control should return to a person.

3. Run and inspect

   Ask for a facts/conflicts/gaps table first. Open each important source location and resolve conflicts with the owner. Then request the draft from checked facts. Check every consequential number, promise and omission; stop any unsupported external action.

   A polished reply can conceal a wrong price, unavailable stock or an invented delivery promise.

   Use only the supplied material and its stated source precedence. First return a fact table with file/section references, conflicts and missing information. Do not infer absent facts. Do not draft the customer reply until I have checked the table. Do not send, publish, order, reserve stock or change records. After my source check: draft a reply from the checked facts. Keep unresolved questions explicit and label the output for staff review.

4. Decide before expanding

   Record errors, corrections, retries and unresolved facts. Choose draft-ready, revise or stop. Record actual completion time, review/rework time and cost if measured; otherwise write unknown. Before claiming improvement, compare equivalent work with today’s process. Widen permissions only after representative runs and failure/recovery checks justify it.

   One successful draft supports that example. It does not establish business value or permission to act for customers.

# Agent workflow brief

Internal operating instructions, not an AI-search optimization file.
This document does not grant or enforce tool permissions.

## Goal and owner
- Task and acceptable result:
- Responsible owner and reviewer:
- Date / next review:

## Approved facts
- File or URL, version/date, responsible owner:
- Required currency, unit, country/language and conditions:
- Precedence when sources disagree:
- Missing facts to resolve with the owner:

## Boundaries
- Allowed: read the supplied materials and draft for review.
- Not allowed: send, publish, purchase, reserve stock, change records or access other data.
- Configure connected tools to enforce these limits. A prompt is not an access control.
- Budget / time limit / retry limit:

## Run
1. Extract relevant facts with exact file/section locations.
2. List conflicts and missing information. Do not infer absent facts.
3. Reviewer checks the table against original material.
4. Draft only from checked facts. Keep unresolved points explicit.
5. Reviewer checks every consequential claim before any external use.

## Acceptance checks — fill in before running
- Required facts and questions:
- Numbers, currencies, dates, locations and conditions:
- Unsupported promises that must not appear:
- Who resolves missing facts:

## Stop and escalate
- A central fact is missing, conflicting or outdated.
- An action exceeds the allowed scope or budget.
- A check fails; stop that action and return the issue to the owner.
- No external action without the separately authorized person and system controls.

## Record the outcome
- Actual tool/model label and date:
- Input versions and saved output:
- Errors, corrections, retries and unresolved facts:
- Acceptable draft / revise / stop, and reviewer:
- Measured completion time, review/rework time and cost, or unknown:
- Next decision and date:

Before claiming improvement, compare equivalent work with the current process.
A successful example is not permission to expand autonomy.

Result checks

• Important facts match the authoritative material, including units, currency, dates, geography and exceptions.
• Each customer request is answered or explicitly left unresolved with an owner.
• Conflicts and missing facts stay visible; the draft adds no unsupported promise.
• No action exceeded the configured scope. The reviewer recorded a decision and any rework.

Test record

Last tested 23 October 2026 · Separate Codex agent; Atlas editor agent checked the source table and final draft · Exact model version not exposed.. One complete fictional run met the predefined checks. The audit caught five conflicting old-page statements. The draft preserved USD 59, the 12-unit limit, Canada’s delivery estimate, the warranty scope, unknown charges and staff-only order placement. The editor checked the audit before requesting the draft, then checked every central conclusion. No corrective prompt or retry was needed. This was one short, synthetic text case in an agent environment with prior editorial context. The answer checklist was not supplied to the test agent; no fresh-context isolation is claimed. It tested drafting and review by AI agents, not human usability, tool-permission enforcement, a live website, multilingual work, actual sending/ordering, visibility or business impact. No comparison, time saving or cost was measured.

Sources in this version

• The lethal trifecta for AI agents: private data, untrusted content, and external communication ↗Published source; consulted 23 October 2026 · The three capabilities and indirect prompt-injection riskPublished 16 June 2025 · recorded in Atlas 23 October 2026