For the people who took the workshop, and the ones about to. Tools change every month; this guide is kept current between sessions.
Last updated: September 2026
Five questions, in this order. Answer them for your workflow before you open a tool.
List the steps. The time usually hides between them: copying, reformatting, waiting, redoing.
A repetitive step follows a rule you could write down. Judgment is where you weigh something only you know.
Reading, writing, summarizing, sorting, pulling fields out, changing formats, drafting. Not being right without a source.
Decide before anything runs: what a person checks, and when.
Minutes per occurrence and times per week, before and after. Note what the figure leaves out.
The examples are the ones shown in the current session. Names change; the kinds don’t.
Copy this table into a document or a sheet, one row per step. The example is an invoice, from the inbox to payment.
| # | Step | Who | Time per week | Repetitive or judgment | Could AI take it? | Who approves |
|---|---|---|---|---|---|---|
| 1 | Open the invoice email and save the PDF | Accounting | 40 min | Repetitive | Yes: automation | Nobody |
| 2 | Type the supplier, date and amounts into the sheet | Accounting | 2 h | Repetitive | Yes: a desktop agent | Spot-check one in ten |
| 3 | Match it against the purchase order | Accounting | 1 h | Mostly repetitive | Partly | Every mismatch |
| 4 | Approve the payment | Manager | 20 min | Judgment | No | Manager |
Frequent, rule-based, easy to check, cheap if wrong. In the example: step 2, two hours a week.
Most workflows people bring look like one of these. Start from the closest.
Six parts. Save the ones that work in a project or as custom instructions, so the context is given once.
Context: I manage customer service for a 60-person distributor. Replies go out in French. Task: Sort the emails below by type (order, delivery, billing, other) and urgency (today, this week). Input: the 20 emails pasted below. Format: a table with the sender, the type, the urgency and a one-line summary. Then a reply draft for each email marked today, under 100 words. Limits: don’t promise dates or refunds. If an email is unclear, say so instead of guessing. Check: at the end, list the emails you were not sure how to sort.
Before you change anything, time the step for a week: minutes per occurrence, times per week. Do the same after. Note the rework and the misses, and what the figure leaves out: a first response is not a resolution, and your own week is not a study.
Some workflows are more than a workshop: several systems, sensitive data, permissions to respect, runs without anyone watching, high volume. That is a build, and it is what Koeo does.
Discuss your challengeThe words you’ll hear, in both languages. In Québec, teams often use the English.
| Term | In French | Meaning |
|---|---|---|
| agent | agent | An AI that takes steps on its own toward a goal, using tools. |
| assistant | assistant | A chat tool that answers and drafts, like ChatGPT or Claude. |
| automation | automatisation | A chain of steps between apps that runs on a trigger. |
| trigger | déclencheur | The event that starts an automation: a new email, a form, a file. |
| connector (MCP) | connecteur (MCP) | What lets an assistant reach one of your tools. MCP is the common standard. |
| browser agent | agent de navigateur | An agent that clicks and types in web pages for you. |
| context | contexte | Everything the model sees: your instructions, the documents, the conversation. |
| context window | fenêtre de contexte | How much the model can read at once. |
| language model (LLM) | modèle de langage (LLM) | The engine behind the assistants, trained to predict text. |
| hallucination | hallucination | A confident answer that isn’t true. |
| prompt | prompt | The instructions you give the model. |
| custom instructions | instructions personnalisées | Context saved once and applied to every conversation. |
| project | projet | A space that keeps files and instructions for one kind of work. |
| token | jeton (token) | The unit a model reads and bills by, about three quarters of a word. |
| retrieval (RAG) | retrieval (RAG) | Looking up your documents before answering, so the answer can cite them. |
| extraction | extraction | Pulling fields out of a document into a table. |
| template | gabarit | A reusable format the result has to follow. |
| human review | révision humaine | A person checks before the result is used or sent. |
| permissions | permissions | What a tool may see and do, usually the same as the user. |
| prompt injection | injection de prompt | Hidden instructions in a page or a document that try to steer the AI. |
| API | API | How two systems talk to each other without anyone clicking. |
| workflow | workflow | A chain of steps that produces a result. |
| personal information | renseignements personnels | Anything that identifies a person. In Québec, Loi 25 governs it. |
| business plan | forfait entreprise | The paid tier meant for work data, with admin controls. |