An AI assistant can draft a client update in seconds. The consequential moment comes one click later: choosing the recipients and sending words that may commit the company, reset expectations, expose private information, or trigger work in the field. A building business should treat that boundary as a send gate, not as a convenient final step hidden inside the prompt.

OpenAI's August 20 ChatGPT Business release notes say its Apple Messages integration can prepare or send iMessage, SMS, and RCS messages, asks users to approve the message and recipients by default, and gives workspace administrators a control to disable the integration. OpenAI Presence describes the wider production pattern: companies define what an agent may do, when approval is required, and when a person must take over. For remodelers, builders, designers, showrooms, suppliers, distributors, and trades, the lesson is simple: drafting authority and sending authority are different permissions.

Separate draft, approve and send

A safe communication workflow has three visible states. The AI may prepare a draft from approved sources. An authorized person reviews the content, evidence, tone, recipients, attachments, and timing. Only then may the message be sent. Do not collapse those states because the model produced fluent copy or because the channel feels informal. A text to a subcontractor can change tomorrow's crew plan just as surely as an email can.

Put five checks at the send gate

  • Identity: confirm the human sender, intended recipients, company, project, and communication channel.
  • Evidence: show the records supporting dates, prices, scope, selections, responsibilities, and status claims.
  • Authority: identify whether the message informs, requests, promises, approves, directs, or changes something.
  • Privacy: remove unrelated project details, personal data, access links, and attachments the recipient should not receive.
  • Record: preserve the approved draft, recipients, source references, approver, send time, delivery result, and later correction.

The review screen should make those checks easy. Show the exact message and full recipient list without truncation. Link claims to the estimate revision, schedule, selection record, contract clause, meeting note, or vendor confirmation that supports them. Highlight uncertain facts and distinguish proposed wording from information already approved by the business.

Classify messages by consequence

A reminder that tomorrow's internal meeting starts at 8:00 is not equivalent to a client message promising substantial completion, a vendor instruction authorizing shipment, or a crew text changing site conditions. Create communication tiers based on consequence, reversibility, contractual effect, privacy, and operational impact. Low-risk internal drafts may use a lighter review. Messages involving money, schedule commitments, scope, safety, personnel, disputes, or external recipients should require a named approver with the relevant authority.

Make recipient approval literal

Do not let a saved contact name stand in for recipient verification. Display the actual phone number or address, organization, project relationship, and whether the person was added by the AI. Group threads deserve special care: the wrong participant can expose pricing, homeowner details, employee information, or another project's records. Any change to recipients after approval should invalidate the approval and reopen the gate.

Design the failure path before enabling send

The system needs explicit outcomes for expired approval, duplicate send attempts, unavailable messaging services, partial delivery, conflicting source records, missing attachments, and messages that arrive after the underlying plan changes. Decide who receives the exception, whether the draft remains editable, how retries avoid duplicates, and how a correction is issued. If the workflow cannot prove who received what, it should stop rather than infer success.

Test in draft-only mode

Start with recent real communication but block all sends. Compare AI drafts with the messages authorized employees actually sent. Measure unsupported claims, wrong recipients, missing context, material edits, review time, privacy issues, and the percentage of drafts accepted without consequential correction. Include messy cases: similar contact names, stale schedules, disputed scope, changed selections, and group threads with outside participants. Enable a sending tool only for a narrow message class after the gate works reliably.

Keep the human approval meaningful

A button labeled approve is not a control if reviewers cannot see the sources or are expected to clear dozens of drafts blindly. Route only messages the reviewer is authorized to judge. Put consequential differences and unresolved facts at the top. Track review time, rejection reasons, corrections after send, and repeated exception patterns. If approval becomes a rubber stamp, reduce volume or narrow the workflow instead of weakening the gate.

Publish the operating evidence

Google says AI Overviews and AI Mode do not require special AI-only schema. Helpful, original, accessible content and accurate structured data remain the foundation. The business value an answer summary cannot replace is the operating artifact itself: your communication tiers, authority map, recipient verification, source panel, exception routes, audit record, and draft-only test results. Those details show how the company protects trust when AI moves from writing words to representing the business.

Continue with controlled agent actions

Sources Read

Next step, if this note maps to a problem on your desk: Private Training — a private working session for your team ($1,500+).

Related Field Notes