How-to
How to wire agents into your data layers
A practical guide to connecting AI agents to your analytics, ad accounts, CRM, product events, and support - so they act on one joined view instead of siloed dashboards.
To be useful, an agent needs one joined view of your data, with read and write access scoped per agent and per surface. This guide walks through connecting the five layers that matter, in the order that gets you value fastest.
TL;DR in 60 seconds
- Connect five layers: analytics, ad accounts, CRM, product events, support.
- Join them on a single identity so every touch ties back to the deal.
- Scope access per agent and per surface - least privilege by default.
- Start with the layer nearest revenue, not the one that's easiest.
The five layers
- Analytics - traffic, sessions, funnels. The map of what happens before the deal.
- Ad accounts - spend, creative, and audience performance across Google, Meta, LinkedIn.
- CRM - the deal record: stages, owners, revenue.
- Product events - activation, usage, and the in-product moments that predict conversion.
- Support - the qualitative signal: what confuses or blocks people.
Step 1: pick a single identity key
Before connecting anything, decide how a person is identified across systems (usually an email or a stable user ID). Without this, you get five dashboards, not one view. The agent's value comes from tying an ad click to a product action to a closed deal - which is only possible on a shared key.
Step 2: connect the layer nearest revenue first
Resist starting with whatever has the simplest API. Start with the layer closest to money - usually CRM or product events - so the agent's first experiments are measured against revenue from day one.
Step 3: scope access per agent
Give each agent read/write only to what its job needs. A lifecycle agent does not need to touch ad budgets. Least privilege keeps the blast radius small and the audit trail clean.
Step 4: add a human gate before write
Reads can be broad; writes to production go through a human gate. The agent proposes and stages the change; a person ships it. See human-in-the-loop, with a guard for the full model.
What "done" looks like
You have one joined view, keyed on a single identity, with agents scoped per surface and a gate before anything ships. Now the experiment loop has something real to optimise against.