Capstone: a real end-to-end ChatGPT workflow

This is the central architecture fork, so resolve it first.

A Custom GPT (configured at chatgpt.com, distributed via the GPT Store or a private link) is the right home for the *internal-facing* assistant. Your two support agents open it inside ChatGPT, paste a customer email, and get a drafted reply grounded in your docs. Zero hosting, zero auth code, and it inherits Connectors and Actions natively.

The APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → (the Responses API) is the right home for anything *customer-facing* or *automated*: the widget on your website, the nightly report job. You need programmatic control, your own auth, logging, and the ability to run without a human in the loop.

So the answer is both, split by audience:

• Agent-facing drafting tool → Custom GPT.
• Website widget + scheduled report → APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition →.

Don't try to force one surface to do everything. The Custom GPT is your fast path; the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → is your durable path.

Decision 2: How docs get into the answer

The docs change weekly, so grounding strategy matters.

For the Custom GPT, use Connectors to link the live Google Drive folder where the team keeps docs. The GPT searches that source at query time, so a Monday edit is reflected Monday. This beats uploading static files to the GPT's knowledge, which would go stale and require re-uploading. See help.openai.com for the current Connectors catalog and admin controls.

For the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition →-backed widget, you own retrieval. Either run your own vector search and pass results into the prompt, or use file search as a hosted tool on the Responses APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition →. For a two-person team, hosted file search is the pragmatic choice: less infrastructure to babysit. A weekly cron job re-syncs the docs into the file store.

Decision 3: Live data via Actions vs. function calling

The assistant needs getCustomer(email) and getRecentTickets(customerId). Same capability, two implementations depending on surface.

In the Custom GPT, you expose these as GPT Actions: an OpenAPI spec pointing at your internal APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition →, with auth configured in the GPT builder. The model decides when to call them.

In the API widget, the same operations become function calling with structured outputs, so the arguments arrive as schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.View full definition →-valid JSON every time. Here is the tool definition for the Responses APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition →:

strict: True is the part worth internalizing: it guarantees the model's arguments conform to your schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.View full definition →, which means your backend code never has to defensively parse a malformed email field. Read the function calling guide for the full loop, including how you send the tool result back for the model's final answer.

The key insight: the *same OpenAPI operations* back both your GPT Actions and your function-calling tools. Write the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → once, expose it twice.

Decision 4: Escalation and the human-in-the-loop boundary

Never let a support bot guarantee a refund or close a ticket on its own. Build an explicit escalation tool, escalate_to_human(reason, urgency), and instruct the model to call it when the customer is upset, asks for money, or the docs don't cover the question.

Make escalation a *first-class tool*, not a vibe. When called, your code drops the conversation into a Zendesk queue and tells the customer a human is on it. This turns "the bot got it wrong" into "the bot knew its limits," which is the difference between a useful assistant and a liability.

Decision 5: Where the model orchestrates vs. where your code does

When you have several tools and multi-step logic (look up customer → check tickets → decide to answer or escalate), you have a choice:

• Let the model orchestrate via the Responses APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → tool loop. Simple, good for linear flows.
• Use the [Agents SDK](https://platform.openai.com/docs/guides/agents-sdk) when you want typed agents, handoffs between specialized agents (a "billing agent" vs. a "technical agent"), guardrails, and tracing. For SupportBot's escalation-and-handoff pattern, the Agents SDK earns its keep because handoffs are a built-in primitive.

If you're prototyping interactively, the ChatGPT agent (the browsing, clicking, multi-step agent inside ChatGPT) is great for one-off investigations ("go find every open ticket mentioning the new billing bug"), but it's not where you run production traffic. Keep that distinction sharp: ChatGPT agent for exploration, Agents SDK for the deployed system.

Decision 6: The weekly report

This is where Advanced Data Analysis (Code Interpreter) shines. The report needs real computation: median resolution time, week-over-week ticket deltas, a chart of top issue categories.

You have two clean options:

Option A: Scheduled task inside ChatGPT. ChatGPT supports scheduled tasks that run a prompt on a recurring basis. Create a task in your Custom GPT: "Every Monday at 8am, pull last week's tickets via the Zendesk connector, compute resolution stats, and produce a summary with a chart." Lowest effort, no code to host.

Option B: API job on your own cron. A small script queries Postgres, hands the rows to the model with the Code Interpreter tool to crunch numbers and render a chart, then emails the result. More control, version-controlled, testable. Choose this when the report feeds something downstream or the founder needs it in a specific format every time.

For a two-person team, start with Option A. Graduate to Option B when the report becomes load-bearing.

Putting the surfaces together

Here is the full picture, surface by surface:

| Job | Surface | Why |

|---|---|---|

| Agents drafting replies | Custom GPT + Connectors + Actions | Fast, no hosting, live docs |

| Website chat widget | Responses APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → + file search + function calling | Programmatic, owns auth and logging |

| Multi-agent routing & escalation | Agents SDK | Handoffs and guardrails are built in |

| Weekly report | Scheduled task (start) → APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → + Code Interpreter (scale) | Real computation on a schedule |

Notice the through-line: one backend API serves the lookups, exposed as Actions to the GPT and as function tools to the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition →. One doc source (the Drive folder) feeds both the Connector and the file-search sync. You are not building four systems; you are building two integration points (your APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition →, your docs) and pointing multiple surfaces at them.

Memory, Projects, and instructions

A few ChatGPT features quietly improve the agent-facing tool:

• Put SupportBot's tone rules, escalation policy, and "always cite the doc" rule in the Custom GPT's instructions, not in every prompt.
• Group the team's support work in a Project so conversations, files, and custom instructions stay scoped together, separate from their other ChatGPT use.
• Leave memory off for the support Custom GPT. You do not want it carrying assumptions between unrelated customers. Memory is great for a personal assistant, risky for a shared support tool.

A note on Canvas and Codex

When you're *writing* the report templates or the email copy, Canvas is the better editing surface than chat: you iterate on a document in place. And the actual implementation work (the Responses APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → script, the OpenAPI spec, the cron job) is exactly what Codex is for. Use it to scaffold the function-calling loop and the Postgres query, then review every line before it touches customer data.

Failure modes to design for

Before you ship, pre-decide what happens when things break:

• The lookup API is down. Your function should return a clear error object, and the model's instructions should say "if a lookup fails, escalate rather than guess."
• The model invents a policy. Constrain it: answer only from retrieved docs, and if retrieval returns nothing relevant, say so and escalate.
• Rate limits or cost spikes. Log tokentokenA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → usage per conversation. The website widget is your uncapped surface, so cap conversation length and add a fallback to human handoff.

Design these on paper first. They are cheaper to decide now than to debug in production.

Key Takeaways

• Split by audience, not by feature. Custom GPT for your internal team (fast, no hosting); the Responses APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → for customer-facing and automated flows (control, logging, auth).
• Build the backend once, expose it twice. The same OpenAPI operations power your GPT Actions and your function-calling tools. Same idea for docs: one source feeds both Connectors and file search.
• Make escalation a first-class tool with strict structured outputs, and instruct the model to escalate on failure or low confidence instead of guessing.
• Start on the low-effort surface, graduate when load-bearing. Use a ChatGPT scheduled task for the weekly report first; move to an APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings. + Code Interpreter cron job only when it feeds something downstream.