GP5 / PRACTICAL GUIDES
What to prepare before an AI workflow project.
By GP5 · 23 October 2026
The fastest useful start is a defined process with someone responsible for it. Before building an AI assistant or connecting your tools, agree the problem, the knowledge it may use and how people will handle exceptions.
GP5 project preparation guide · English edition · 23 October 2026. A planning checklist, not a promise of delivery time or business results.
1. Choose one workflow.
Write a single sentence: “When this enquiry arrives, we want this person to receive this information and take this next action.” Identify the starting channel and the ending state. Avoid beginning with a replacement of the entire CRM.
2. Name the process owner.
Choose the person who approves answers, tests the workflow and reviews failures. Name a backup. Decide who can change the instructions and who can authorize the assistant to take an action rather than merely draft a suggestion.
3. Gather approved knowledge.
- Current service descriptions and scope boundaries.
- Answers to recurring questions, with the source and last review date.
- Topics the assistant must escalate.
- Booking, contact and operating-hours information.
- Approved wording for uncertainty and human takeover.
Resolve contradictory documents before connecting them. Start with a small, reviewed knowledge set. More files do not automatically produce better answers.
4. Map the current tools and access.
List the website form, CRM, calendar and team communication tools. Confirm who owns each account, which plan is in use and whether the required integration is available. Share secrets through an approved credential mechanism, never through a public form or browser JavaScript.
For WhatsApp, confirm the actual Business Platform setup, number eligibility and messaging permissions before promising a live assistant. A clickable WhatsApp link is a contact route; it is not proof of an active Cloud API integration.
5. Decide what data is needed.
Map each input to its destination and purpose. Agree access, retention, deletion, contact permissions and the handling of sensitive information. Keep unnecessary personal data out of prompts, logs and analytics. Define how an operator can correct a record.
6. Write the acceptance tests.
A good test list covers a normal enquiry, an unknown question, a duplicate submission, an unavailable integration and a request for a human. Decide what the visitor sees in each case. “It replied” is not enough if the next action never reaches the owner.
7. Record a baseline.
Measure a representative period before launch: enquiry volume, time to a useful response, manual handling steps, missed handoffs and attended qualified calls. Separate observed facts from estimates. After launch, include supervision and exception handling when assessing time saved.
8. Agree scope and ongoing care.
Confirm channels, integrations, included usage, testing, training, vendor costs, support hours and change requests in the project scope. GP5’s service packages are starting points for that conversation. Delivery depends on readiness and the agreed work; revenue and AI-search placement are not guaranteed.
Bring this to your first conversation.
- One workflow problem and an anonymized example.
- The process owner and current tools.
- Approved answers and escalation boundaries.
- Available baseline measures.
- The result that would justify a pilot.
Next: read the website-to-Monday workflow blueprint or explore GP5 AI Solutions.
More practical guides: AI assistant or workflow automation? · Measuring an AI workflow.