What to collect.
Recent emails, chats, calls or documents, covering the normal cases and the awkward ones.
FAQs, price lists, terms, preparation instructions, tone of voice: the only sources the agent may answer from.
Which inbox, calendar, CRM or ERP, with the least access that does the job.
Complaints, refunds, legal or medical topics, VIPs: the list that always goes to a person.
The person who signs off answers and decides when the agent is wrong.
Response time, hours saved or documents processed: a number to judge it by.
Are you ready? A quick test.
| Question | Ready if… | Not ready if… |
|---|---|---|
| Can we find 50 real examples in a day? | They are in one inbox or system | They are spread over personal phones |
| Is our price list current? | One source, updated this quarter | Three versions in three places |
| Do we know who approves answers? | A named person with time | "Everyone" or nobody |
| Does the software have an API on our plan? | Yes, or we have checked | Unknown |
| Do we know what personal data is needed? | A short list of fields | "Everything, just in case" |
Data the agent should not get.
Less data makes the build faster and the privacy questions easier.
Record the baseline now.
Write down today's numbers before the build: how long a reply takes, how many items per week, how many hours. The NIST AI Risk Management Framework calls this measuring against intended purpose; in plain terms, without a baseline nobody can say whether the agent helped.
Not legal advice.
This guide is general information, checked against the sources listed, and not legal advice. Rules change; check the current text or ask a lawyer before you rely on it. Checked November 2026.
Common questions.
Do we need a data warehouse first?
No. Most first agents connect directly to an inbox, calendar or CRM. A warehouse helps for reporting across many systems, not for a first agent.
Do you train a model on our data?
Usually not. The agent uses a general model and looks up your approved content at the moment it answers, which keeps your data out of model training.
What if our data is messy?
Then the first step is fixing the few sources the agent needs, not all of them. The Data Readiness service exists for exactly this.
How many examples are enough?
Enough to cover the common cases and the known edge cases. For a front desk, a few weeks of messages usually is.
Want this in your business?
Tell us how you work today. We'll tell you honestly what AI can take off your team's plate, what it costs, and whether we are the right fit.