AI-Ready Data Control Tower AGENTFORCE + DATA CLOUD · Crestwave Media Group · a candidate concept by Benny Garner
⚡ RAG & Chatbot
❝

Agents don’t fail on intelligence — they fail on evidence. Data Cloud gathers it, mastering makes it trustworthy, and the closed loop puts the agents to work defending their own grounding data. Harmonized isn’t mastered — and mastered is where trust is earned.

Benny Garner
Principal Data & AI Architect — candidate

Stage

Control Tower — presenter guide

What this is

One screen showing how a messy, five-system customer estate becomes an agent-ready grounding set: Data Cloud ingests and resolves identities, a mastering layer applies quality gates and survivorship, predictive models score the clean profiles, and Agentforce acts on what survives — with the Einstein Trust Layer auditing every answer. Data is synthetic; the architecture is the point.

Reading the flow

Red = blocked from agent use until fixed (field conflicts, dead contact points). Amber = held on purpose (missing fields, consent). Green = trusted and flowing to agents. Blue = resolution pipeline doing its job. Click any band or node — the records table filters to the accounts behind it, and the stage card names the exact Salesforce component that implements it.

The three stories to tell

Q & A you may get

Is this running on real Agentforce?
The steward agent is simulated — my own reasoning loop posting to Slack — and it’s labeled as such. The seams are exactly where the real components plug in: the detection topics become Agent Builder topics, the reasoning call becomes Atlas grounded via the Trust Layer, the Slack card is the Agentforce-in-Slack surface, and the approval callback is a custom agent action writing back through Data Cloud.
Where does the Informatica acquisition fit?
It’s the strongest possible confirmation of the thesis: Salesforce bought an MDM + data-quality + governance company because agents need mastered data. The mastering layer in this demo is exactly the capability Informatica brings first-party — survivorship, quality rules, catalog and lineage — and my career has been competing with and against that stack, so I know where it lands well and where it needs architecture around it (match tuning, survivorship politics, steward adoption).
Doesn’t Data Cloud identity resolution already solve this?
It solves matching. It doesn’t decide which system wins when merged sources disagree on credit status, doesn’t score trust, and doesn’t route borderline cases to a human. That survivorship-and-governance layer is what I’ve built for a living — expressed here natively, no third-party MDM box in the diagram.
Why do stewards approve in Slack?
Because that’s where people already are, and Slack is Salesforce. A 30-second approval in the flow of work beats a data-quality backlog review every quarter.
Why does the churn model care about mastering?
A model trained on duplicates is garbage-in twice — once at training, once at inference. Velocity Auto Mall scores 0.91 as one customer; as three fragments it never crosses a threshold.
What’s the KPI?
The Trust Score (top right): five weighted dimensions over the whole estate. It’s the number a CDO can put in a board deck — and the roadmap commits to moving it 71 → 90.

Good to know

This page: agent-force.cuppy.io · deck via ▶ Strategy deck · theme toggle ◐ persists per browser. All figures modeled on the synthetic Crestwave Media scenario; labeled accordingly. Built by Benny Garner as interview material — not an official Salesforce property.
→
Data Cloud
48,230
records ingested · resolved
→
Agent-ready
18,910
passed every quality gate
→
Agentforce
grounded actions
cited · consented · audited

Where every customer record goes

Five source systems → one trusted grounding set. Click any band or node to see the accounts behind it and the Salesforce component that implements the stage.

Blocked — conflicts & dead contact points; agents can’t act until fixed Held — enrichment & consent; right record, not yet safe Trusted — flowing to Agentforce Resolving — Data Cloud pipeline working as designed

Ask Agentforce — same question, two grounding strategies

Pick a question on the left. The right shows the same question answered by an agent grounded on raw CRM vs. the mastered Customer 360 with the Trust Layer in the loop.

⚠ Grounded on raw CRM
✓ Grounded on mastered Customer 360
Simulated agent responses over the synthetic scenario — the failure mode on the left is the one enterprises actually report.

How the loop runs — predict → reason → approve → improve

Predictive ML and generative LLM, both grounded on the mastered Customer 360 — with a human between predict and act.

31,480profiles scored ML predictsModel Builder · churn 0.91, propensity 0.87 — 2 models over mastered profiles, so they see customers, not duplicates → 100%answers cited LLM reasons & draftsAtlas via Prompt Builder · grounded on the golden record + RAG corpus, masked and audited by the Einstein Trust Layer → ~30sper card · 460/wk Human approvesSlack card in the flow of work — one 30-second decision, not a quarterly cleanup project → +9 ptstrust / 30 days Data improvesapproved fix updates the grounding set · Trust Score climbs toward 90 · both models retrain on cleaner truth

The closed loop — agents stewarding their own grounding data

The Data Steward Agent detects issues, proposes fixes with rationale, and asks a human to approve — in Slack, in the flow of work. These three cards are posted live in the demo workspace: open Slack ↗

Accounts

Click the chart to filter. Showing the named stories.

AccountSourcesStageWhyScore