Everyone wants an AI agent this year, and I understand why — I have watched a well-built one close cases and draft outreach faster than a room of people could. But I have also watched the other kind: the agent that greets a fifteen-year customer like a stranger, offers a renewal on a product they already returned, and cheerfully recommends the exact thing that caused last month’s complaint. It was not a dumb model. It was a confident one with no idea who it was talking to. And that is the quiet truth behind most disappointing agent projects: the intelligence was never the bottleneck. The context was.
An agent is a golfer. It takes the shot, it gets the applause, and it is only as good as the caddie standing next to it — the one who read the green, measured the wind, and handed over the right club. In Microsoft’s stack, that caddie is Dynamics 365 Customer Insights. Not as another app to open, but as the grounding layer underneath the agents: the thing that knows who the customer is, what they have done, and what they are in the middle of right now, so the agent’s confident swing actually lands on the green.
Why bolting an agent onto your CRM isn’t enough
The instinct is to point an agent at your CRM and call it grounded. It helps, but a CRM record is a filing cabinet: it tells the agent what was formally written down, by whoever remembered to write it. It does not, on its own, tell the agent that this customer opened your last three emails but clicked nothing, that their usage dropped 40% this month, that a support case is open right now, and that they are two days into a renewal window. Those signals live in different systems, in different shapes, and a customer is not one row — they are a scatter of rows across marketing, service, product, and web. An agent grounded in one system answers confidently from a partial picture, which is worse than answering uncertainly from the whole one. Confidence without context is just a faster way to be wrong.
What “grounding layer” actually means
Customer Insights does two jobs that turn scattered data into something an agent can safely stand on. Customer Insights – Data is the unification engine: it stitches those scattered rows into a single, continuously updated profile of each customer, resolves the duplicates, and exposes segments and measures — so “who is this, really?” has one answer instead of five. Customer Insights – Journeys is the action engine: it turns those profiles and signals into orchestrated, consent-respecting engagement across channels, at the moment the signal appears rather than in next quarter’s campaign.
Put plainly: Data gives the agent a memory, and Journeys gives it a sense of timing. An agent grounded this way stops guessing. When it drafts an email, it drafts to the person the profile describes. When it decides whether to reach out, it reads a live signal instead of a stale field. The model did not get smarter — it got informed, and informed beats clever every time a real customer is on the other end.
The shift: from system of record to system of action
This is the part worth taking to a leadership meeting. For twenty years we bought CRM to record what happened — a system of record, honest and passive, a very expensive diary. The pitch for grounding your agents in Customer Insights is that the same data stops being a diary and starts being a trigger. A disruption to a high-value account doesn’t wait for someone to notice it in a report; it becomes a proactive outreach. A high-intent buying signal doesn’t sit in a dashboard nobody opened; it surfaces to the seller while it still means something. The data crosses the line from describing the past to driving the next move. Microsoft’s own framing for this is a system of action, and the phrase earns its keep: the value is not knowing more, it is doing the right thing sooner.
The recent additions lean hard in this direction. A Journey Creation Agent that builds a campaign from a plain-language prompt in minutes, an Outreach Optimization Agent that quietly tunes timing and follow-up based on how each customer actually behaves, and a Customer Insights MCP server that lets connected agents check consent in real time before they act — these are all recent, still-maturing capabilities, and I would pilot them rather than bet the quarter on them. But the direction is unmistakable: the data layer is growing its own agents, and it is doing it with consent checking built into the plumbing rather than bolted on after the incident.
The part everyone skips: consent and governance
Grounding an agent in richer customer data is exactly as risky as it sounds if you do it carelessly, and this is where I get a little stern with clients. The whole appeal of a unified profile — the agent knows everything about the customer — is also the whole liability. An agent that can act on a signal must also be able to check whether it is allowed to, per channel, per customer, in real time. This is why “consent-aware” is not marketing garnish on the MCP server story; it is the load-bearing wall. Ground your agents in a layer that treats consent as a first-class, checkable fact, or you will eventually explain to a regulator why your very smart system did a very confident wrong thing. A grounding layer that cannot say “no, not this channel, not this customer” is not grounded. It is just well-read and reckless.
What this means for you
You do not need to boil the ocean, and you should not. If you are experimenting with agents in sales or service and they feel oddly context-blind, the fix usually is not a better prompt or a bigger model — it is asking what the agent is standing on. Start by unifying the profile for one domain that matters (your top accounts, your renewal book), prove that a grounded agent behaves noticeably better than an ungrounded one on that slice, and let the caddie earn the bag before you hand it the whole course. The agents will keep getting more capable on their own. Whether they are any good to your customers depends on what you let them know — and on your discipline about what you let them do with it.
Wrapping up
The agent takes the shot, but the caddie wins the round. Customer Insights is not the flashy part of an AI strategy — nobody demos a unified profile to applause — and that is precisely why it gets skipped, and precisely why the skipped projects underwhelm. An agent grounded in a real, consented, current picture of the customer is a professional with local knowledge. An agent without it is a confident stranger with a driver and no idea where the flag is. Give yours a caddie.
