AI is last in the sequence, and that is a deliberate recommendation rather than an ordering accident.
Why last
An assistant answers from what your organisation knows. Before your records, your vocabulary and your policies exist, it has nothing of yours to draw on, so it answers generically.
Generic answers on day one are how a team decides the AI is not for them, and that impression is expensive to reverse. People try a new tool roughly twice.
So: data first, then the assistant. The wait is a week and it changes the reception entirely.
The order
1. Providers. See How it is set up.
2. Persona and guardrails. See Persona and guardrails.
3. Knowledge. See Teaching the assistant.
4. Model. See Choosing your model.
5. Agents, after everything above has settled. See AI agents.
What has to exist first
Three things, and the order is the point.
Records, so there is something to answer about.
Your vocabulary, so answers use your words rather than generic industry ones.
Somebody who will keep the knowledge current, because material that goes stale produces confidently wrong answers rather than obvious gaps.
Without the third, this becomes a setup that works for a quarter.
Providers
For most organisations there is nothing to do. The intelligence is provided, and the page shows it as active.
If you hold your own keys, this is where they go, and the connection test is the control worth using. See How it is set up.
Persona
How it sounds, and the limits it works inside.
Custom instructions are the part worth writing. Your organisation's actual rules: how you refer to your products, what you never quote without a caveat, which currency comes first. No default can guess those, and five sharp rules change every answer.
Leave the rest at default unless you have a reason. See Persona and guardrails.
Knowledge
The step that decides whether any of this was worth doing. See Teaching the assistant.
Start with vocabulary. What you call your products, stages, teams and customers. Highest value by a distance, because without it every answer reads as though written by somebody who has not worked here.
Then policies. Refunds, escalation, approval thresholds, notice periods. The things people ask about constantly and cannot find.
Then check it retrieves. Type a real question and look at what comes back. Uploading and assuming is how organisations conclude the AI is poor when it never saw the file.
Nominate an owner
Somebody has to own what the assistant knows, or the knowledge is current for one quarter and stale for every quarter after. One name, per area.
Language
If your customers read a language your team does not use internally, set the assistant's language to the customers' one before anything it writes goes outbound.
The setting is a default rather than a constraint: somebody who writes to it in another language gets an answer in that language. What it decides is the unprompted output, which is exactly the output a customer sees. See Persona and guardrails.
Agents last
An agent works on a schedule without being asked, which means a badly configured one produces work nobody wanted, on a schedule, quietly.
Start with one, and have it create tasks only. A wrong task costs somebody thirty seconds. A wrong email reaches a customer.
Use the built-in ones. Deal Reviver, Cash Watchdog and Meeting Follow-ups between them cover most of what an organisation actually needs, and they are already scoped.
Watch the first week of runs. An agent doing the wrong thing quietly for a month is the real risk, and it is entirely preventable by looking twice.
Give it a fortnight before judging
The first week of any assistant is people testing it with questions they already know the answer to. The second week is when they start asking things they actually need, and that is the week worth listening to.
What to switch off
More useful than what to switch on.
Each AI surface is separately controllable, and turning everything on at once is the most common rollout mistake. People meet the assistant, agents, AI in chat and AI on WhatsApp in the same week, form a single impression, and it is usually that the product has become noisy.
Start with the assistant only. Add the rest once people ask for it, which they will once the first one is genuinely useful.
Telling people
Worth a short message rather than a launch.
Say what it is good at, with two examples from your own business. Abstract announcements produce abstract questions and disappointing answers.
Say what it cannot do. It does not send, approve or commit, and it cannot see anything the person asking could not see. That second point answers the question everybody has and rarely asks.
Do not mandate it. Tools people are told to use get used performatively.
Cost, honestly
Worth answering before somebody asks.
Ordinary use is not expensive, and rationing questions is the wrong instinct: an organisation told to use the assistant sparingly never learns what it is for, and saves very little. What genuinely consumes is unattended work, mainly agents on a tight schedule. See Credits and what uses them.
Checking it worked
Ask it three questions only your organisation could answer. If it cannot, the knowledge is missing or not retrieving, and no other setting will help.
Ask something a person should not see, from an account that should not see it. It should decline, because it inherits permissions rather than having its own.
Check the usage page a fortnight later. Flat and low means people tried it and stopped, which is a knowledge problem rather than an adoption problem.
Next
The checklist, to confirm nothing was left half done.
Did this answer your question?
No, ask a person