OpenAI introduced ChatGPT Work in July as an agent for longer projects across connected apps and files. Its Scheduled Tasks can run once, repeat on a schedule or trigger, or monitor for changes. Users can follow progress, redirect the work, and approve important actions.

That combination matters to remodelers, builders, designers, showrooms, suppliers, distributors, and trades because much of their office work is recurring: check new bid invitations, compare yesterday's schedule with today's, flag overdue selections, assemble a weekly purchasing report, or watch for vendor-document changes. But putting a task on a schedule does not make its result dependable.

A schedule is only the trigger

A useful recurring workflow has at least three separate parts: the event that starts it, the work it performs, and the condition that proves it finished correctly. “Every weekday at 7 a.m.” defines only the first part. It does not say which project list is authoritative, how fresh the vendor data must be, what counts as an overdue selection, or who sees an incomplete run.

Before scheduling an AI task, write its work order in plain language: source systems, allowed actions, expected artifact, deadline, reviewer, and stop conditions. If a capable coordinator could not tell whether the result is complete from that definition, the agent cannot be evaluated against it either.

Silence is not a success state

Unattended work fails quietly when a login expires, an attachment changes format, a project is renamed, a required field is blank, or two systems disagree. A green “ran” status may only prove that the task started. Building businesses need separate states such as queued, running, completed, completed with warnings, waiting for approval, retrying, and failed.

Each run should leave an inspection record: when it started, which sources it read, which records it considered, what it produced, what it skipped, which instructions or model version it used, and whether a person changed or approved the result. That record turns an agent output from an unexplained answer into reviewable work.

The exception queue is the operating surface

Do not force the agent to guess through conflicts. Route unusual cases into a visible queue: a purchase order with no job number, a selection whose due date moved backward, a bid whose scope cannot be matched, or a report built from a stale export. The queue should show the source evidence, the reason for the flag, the proposed next step, and the person responsible.

This is where human attention creates the most value. The agent handles ordinary repetition; the operator resolves judgment calls and improves the rules. Over time, reviewed exceptions can become an eval set that tests whether the workflow catches the same important cases after instructions, integrations, or models change.

Retry the system problem, escalate the business problem

A temporary network error may deserve an automatic retry. Conflicting prices, missing approval, or an ambiguous client instruction do not. Classify failures before launch so technical interruptions can recover without creating duplicate messages or records, while business exceptions stop for review. Any action that sends a client message, changes a committed price, releases payment, or writes to a system of record should have an explicit approval boundary.

  • Start with one recurring task whose correct result an operator can already judge.
  • Name the authoritative sources and the maximum acceptable data age.
  • Define success, warning, approval, retry, and failure states before scheduling it.
  • Save the sources used, output, skipped records, errors, edits, and approval for every run.
  • Put ambiguous cases in an owned exception queue with evidence and a deadline.
  • Test ordinary cases plus missing files, stale data, conflicting values, expired access, and duplicate triggers.

The goal is not to wake up to more AI-generated material. It is to wake up to completed routine work, a short list of real exceptions, and enough evidence to know which is which.

The same evidence-first approach helps buyers and search systems understand the work. Google says AI Overviews and AI Mode require no special AI-only schema; helpful original content, crawlable text, accurate structured data, and visible source support remain the foundation. A concrete workflow with named inputs, states, exceptions, and approval boundaries gives people—and machines—something more useful than a generic promise of automation.

Sources Read

Next step, if this note maps to a problem on your desk: Discovery Call — a 1-on-1 leverage assessment for your business ($1,500 · 90 min).

Related Field Notes