An editor approved an advertisement saying “Pickup until 11:00”. The assistant replaced it with “Delivery until 12:00” and sent it for publication with the old approval flag. The fulfilment method and time changed, while confirmation still belonged to another text. The error appeared between drafting and taking action.
For a marketer, this assistant starts with a description of the work: which sources it reads, what it prepares and which operations it may perform. The conditions and amounts in this lesson's examples are illustrative. The team must connect facts, output versions and tool permissions so the client receives the work they approved.
Choose an outcome before the technology
A GPT in ChatGPT is a configured assistant with instructions, reference material and selected capabilities. Creation and publication depend on the account and workspace permissions; review the available conditions in OpenAI's GPTs help. Configuring this assistant within ChatGPT and embedding a chat in a separate website through an API are different tasks.
OpenAI is winding GPTs down: it plans to retire them in affected Enterprise workspaces on 11 December 2026, expects other plans to follow the same timeline and recommends moving workflows to Plugins. Personal Free, Go, Plus and Pro accounts can no longer create new GPTs (OpenAI help, as of 15.10.2026). Everything below about sources, permissions and versions applies to any assistant: a GPT, a plugin or an agent built on the API.
In this lesson, an agent is a system that chooses steps and uses tools to complete a task. Fixed-sequence automation follows an order established in advance. Both approaches can work with external services. Listing available actions is more useful to a client than debating the system's name.
A first launch needs a process with a clear end. Consider a weekly review of competitors' public offers: selected pages, comparison with an earlier snapshot and an internal draft with links. Completion means sources were processed, inaccessible pages flagged and each conclusion linked to evidence. Looking at a few sites does not establish that the whole market was researched.
This task needs permission to read selected pages and write to a designated folder. The customer database, sending emails and changing ad budgets do not contribute to that output. Exclude those operations from the connection before testing writing quality.
Give reference material an owner
Product terms, a brand guide and advertisement examples answer different questions. The guide helps choose tone but does not establish the current price. An old advertisement shows an earlier presentation; the current offer needs a current source. Factual documents benefit from a version, effective date and owner.
On a conflict, the assistant should show both sources and refer the question to the fact owner. If one document offers pickup and another delivery, the model must not select a convenient promise because it reads better. It flags the discrepancy and prepares the parts of the work that do not depend on it.
Tests include similar product names, withdrawn offers and absent answers. The assistant should identify missing information with references to the material reviewed. Substituting an unsourced number hides the gap from the editor and may carry it into advertising.
For competitor reviews, retain the address and observation date. Mark an unavailable page as unread: access may have failed. Record the observed result and leave the check open while the cause is unknown.
Support restrictions with tool permissions
“Drafts only” must match the available operations. If a connection can only save a file in a working folder, a writing error cannot immediately become a public advertisement. This is a useful first-release process for advertising material: draft, human review and separate approval to publish.
| Action | Rule in our process | Record retained |
|---|---|---|
| Read a public source | Selected list only | Address, date and reading result |
| Save an internal review | Working folder only | Version, sources and open questions |
| Publish material in an expanded process | Final text and destination approved | Who approved and what was sent |
| Change an advertising budget | Outside the task | Refusal and reason |
Reading authorised in advance proceeds within its scope without asking again at every step. For consequential operations, the system compares permission with action parameters. A general “user confirmed” flag is insufficient if the text or recipient changes afterwards.
In the example, pickup until 11:00 was version 1; delivery until 12:00 became version 2. New conditions go back to the editor while the earlier approval stays tied to the earlier text. Publication receives the version the person saw during review.
Resolve a lost response without repeating the action
A service can execute an operation while a connection fault prevents the assistant from receiving the response. If it assumes nothing happened, another request may create a duplicate. That could be a second email or another record instead of continuing the first.
The integration should retain an operation identifier and determine its state. If the service accepts an idempotency key, a retry with the same key will not create a second email. Whether it does is in the service's documentation.
Logs distinguish preparation, a request sent, confirmed execution and an unknown outcome. For an unknown state without duplicate protection, establish what happened first. “Done” should follow from the operation's result, not merely an attempt to send a request.
Treat another page as source data
A competitor's page might contain a demand to ignore rules and send someone else's database. In this workflow, that is part of the text being examined. It does not extend the authority the team gave the assistant.
OpenAI's agent workflow safety guidance addresses prompt injection, attempts by external content to influence system instructions. Limited tools, data checks and testing are needed together. Access to send a customer database has no place in a public-offer review.
In a test, the assistant should retain the useful meaning of the page and flag the suspicious passage. Assess actual tool calls: a correct explanation after an unauthorised transmission does not undo it. Add this case to the regular checks before expanding access.
Include review work in the calculation
Suppose a manual review took 4 hours and the employee's internal planning rate is 5,000 ₸ per hour. That gives 4 × 5,000 = 20,000 ₸ per edition. After configuring the assistant, checking takes 2 hours and corrections and formatting another 1 hour. Labour costs 3 × 5,000 = 15,000 ₸.
Add 2,000 ₸ per edition for computation and tools. Total cost is 17,000 ₸, or 3,000 ₸ less than the manual edition. If initial configuration cost the team 30,000 ₸, its calculated payback takes 30,000 / 3,000 = 10 editions at the same scope and quality.
Freed hours save money only if they go to other work. Show hours, assigned cost and validated review quality in the decision. If editing time rises, update the estimate instead of retaining the promised saving.
Exercise: define tests for your assistant
Choose a recurring project task and describe its outcome, required sources and permitted operations. Prepare cases involving conflicting old and new prices, text changed after approval, a lost service response and instructions within another page. Record the expected result for each.
Conflicting facts should remain visible with their sources. A changed promise returns for approval, an unknown send requires state verification, and external text remains data. In the first release, public material passes through a draft and a human before publication.
The time calculation complements these tests: an assistant requiring more review may not recover its setup cost in the chosen process. The client needs reproducible work with clear costs and action boundaries. Expand authority after verifying that outcome.