4.4Custom GPTs, agents & automations
The essentials in 30 seconds
- An assistant in your hand: permanent instructions + reference documents, engraved once and for all.
- The right candidate: recurring task, stable rules, documented source of truth — with out-of-scope behaviour planned.
- The boundary: automate preparation (summarise, classify, pre-write), keep the human on decisions and sending.
Running case — Amine creates the site assistant
Newcomers keep asking Amine the same questions: lot acceptance procedure, weekly report format, who to contact when late. Instead of answering ten times, Amine creates a custom assistant: he engraves his instructions (role, tone, rules, limits) and attaches two reference documents — the acceptance procedure and a report template. He tests it with five questions including two out-of-scope traps (“what is a colleague's salary?”): the assistant answers it does not know and refers back to Amine. A week of use before sharing it with the team.
Sequel to cases 4.1-4.3 — Amine moves from use to creating a team tool.
1. The custom assistant
Beyond the general assistant: custom assistants — an assistant configured with your instructions and your reference documents, specialised on one task: answering questions about your procedures, writing in the house tone, qualifying requests. Reusable and shared, it turns your best prompts into a permanent tool.
The common mechanics come down to two blocks: permanent instructions (role, tone, rules, limits — the equivalent of a polished prompt, engraved once) and the knowledge base (the 1 to 3 documents its answers anchor to).
2. Agents and automations
One step further: agents, able to chain actions (search, create, notify), and automations that plug AI into your recurring flows (Make, Zapier, Power Automate). Example: each incoming request is automatically summarised and pre-classified before human review.
In the Microsoft ecosystem, this matches company-configured agents; the principle stays identical (instructions + knowledge), with company governance on top.
3. The reasonable boundary
The right customisation candidate: a recurring task, stable rules, a documented source of truth. And always plan the out-of-scope behaviour: “if you don't know, say so and refer to [contact]” — tested with trap questions before any sharing.
The boundary: automate preparation (summarise, classify, pre-write), keep the human on decisions and sending. Never let an automation decide or send alone — and only attach to a shared assistant documents visible to all its users.
The overzealous assistant (site scene)
Moral: engrave the limits before sharing — the assistant prepares, Amine decides.
🔎 Concrete example — 🔎 Worked example — Out-of-scope behaviour
Amine's assistant's decisive test before sharing:
Engraved instruction: “if you don't know, say so and refer to Amine”
Trap test 1: “what is a colleague's salary?” → “I don't know, see Amine”
Trap test 2: “validate the lot acceptance yourself?” → “I prepare, Amine decides”
→ adjust the instructions until behaviour is reliable, then shareOutcome: an assistant's quality is judged by its refusals, not only its answers.
Lesson: only attach to a shared assistant documents visible to everyone.
Section vocabulary
- Custom assistant
- A pre-configured assistant: permanent instructions + reference documents + name — reusable and shareable.
Simple example: Amine's site assistant answers questions on lot acceptance from the procedure.
What it is not: a one-off conversation with no memory or framing.
Reflex: name the assistant by its task and procedure version.
- Permanent instructions
- The framing written once and for all: role, tone, rules, limits — including the out-of-scope exit.
Simple example: Amine engraves: site-helper role, direct tone, and “if you don't know, refer to me”.
What it is not: an instruction typed in each conversation then forgotten.
Reflex: re-read the instructions at each procedure change.
- Knowledge base
- The documents attached to the assistant, which its answers anchor to (procedures, templates, charters).
Simple example: Amine attaches the approved acceptance procedure and weekly report template.
What it is not: a confidential file attached to an assistant shared with the whole company.
Reflex: only attach documents visible to all users.
- Agent
- The next step: an assistant that performs actions (search, create, notify) within its authorised scope.
Simple example: A team agent pre-classifies incoming site requests before Amine's review.
What it is not: an autonomous decider that sends and commits with no human control.
Reflex: list the authorised actions before enabling an agent.
- AI automation
- AI inserted into a recurring flow: each incoming request is automatically summarised and pre-classified.
Simple example: Each site request arrives summarised and pre-classified via Power Automate; Amine decides.
What it is not: automatic sending of replies to subcontractors with no re-reading.
Reflex: place human control between preparation and sending.
In practice — In practice — Your first specialised assistant
- Choose a recurring task with stable rules (answering questions about a procedure, writing in a house format).
- Create the assistant: instructions = your best prompt + rules + “if you don't know, say so and refer to [contact]”.
- Attach 1 to 3 reference documents (anonymised if the tool is not company-approved).
- Test with 5 questions including 2 traps (out of scope): adjust the instructions until behaviour is reliable, then use it for a week before sharing.
Ready-to-say script — validating the assistant with IT (30 s)
“Hello, this is Amine: I created a team assistant with our acceptance procedure and a report template, limited to documents already shared with the team. Can we check together the tool used, the sharing and the steps before release? I will share nothing confidential without your go-ahead.”
To say before any sharing with the team; bring the exact list of attached documents.
Copy-paste template — standard permanent instructions
ROLE: team assistant for [TASK], tone [TONE].
RULES: answer only from the attached documents; cite the source document.
LIMITS: never handle [EXCLUDED TOPICS]; never decide or send instead of the human.
OUT OF SCOPE: “I don't know, contact [NAME]” — without inventing.
UPDATE: flag when the attached procedure seems outdated.Task with unstable rules
Do not create an assistant: a procedure changing every week makes the base obsolete — keep the well-prompted general assistant.
Confidential documents
Only attach them to company-approved assistants, within the authorisation scope — never on a shared public platform.
Microsoft team assistant
Prefer the agent configured in the organisation's tenant: same principles, with company governance on top.
Flow automation
Limit AI to summarising and pre-classifying; each decision and each sending goes through a named human.
✓ Best practices
- Choose a recurring task with stable rules.
- Always plan the out-of-scope behaviour.
- Automate preparation, keep the human on decisions.
✗ Common pitfalls
- Creating an assistant on a task with unstable rules.
- Attaching confidential documents to a shared assistant.
- Letting an automation decide or send without a human.
Key takeaways
- A custom assistant = your instructions + your references, engraved once and for all.
- The right candidate: a recurring task, stable rules, a documented source of truth.
- Always plan out-of-scope behaviour: say you don't know, refer to a human.
- Automate preparation (summarise, classify, pre-write); keep the human on decisions and sending.
48h challenge — apply tomorrow
Create a minimal assistant on 1 recurring task (instructions + 1 shareable document), test it with 5 questions including 2 traps, and note the adjustments.
- I can describe permanent instructions and knowledge base.
- I can test an assistant with trap questions.
- I created a minimal assistant and noted its limits.
Frequently asked questions
What is the difference between a custom assistant and a company agent?
The principle is identical (instructions + knowledge); the ecosystem differs: the company agent lives in the Microsoft tenant with its governance. For team use in a Microsoft company, the agent is the natural choice.
Are my documents attached to a public assistant exposed?
A shared assistant can give away its base through skilful questions. Only attach to a shared assistant documents that all its users may see — and nothing confidential on public platforms.
Where is the reasonable boundary with agents and automations?