Case Study
12 Ad Platforms, One Answer. None of It Leaked to Big Tech.
How a UK PPC agency built an agentic unified analytics layer with an anonymisation proxy that keeps client data out of the hyperscalers
12 ad platforms. Hours of cognitive overhead just to pull one answer.
12
Fragmented ad and commerce platforms feeding into client reporting
Every report
Multi-tab manual aggregation per client, every cycle
Daily
Client interruptions to PPC managers for routine data questions
Yes
Raw client data flowing into LLM APIs without an anonymisation layer
Google Ads, Shopify, GSC, GA4, Meta, TikTok, LinkedIn, Google Merchant Centre, Microsoft Advertising, Amazon Seller Central, Apple Ads. Twelve sources, each with their own login, their own taxonomy, their own definition of a click. Client reports were assembled by hand, tab by tab, manager by manager.
The Lean AI Method
1. Discovered
Mapped the full client reporting cycle alongside the team. Where each number lived, who looked it up, how often the same answer was assembled from scratch. Found the same data being re-fetched, re-joined, and re-formatted dozens of times a week.
2. Designed
A unified data pipeline that pulls all 12 sources into one warehouse, plus an MCP layer so any AI agent (ours, Claude, in-house) can query the data without bespoke integrations. And a proxy wrapper that anonymises client data before any prompt ever leaves the platform.
3. Deployed
A full analytics platform with cross-source dashboards, on-the-fly report generation, client annotations, audit-ready backups, and a self-serve client portal that answers the routine questions without paging a PPC manager.
4. Scaled
The platform doubles as the agency's competitive moat: no other PPC shop in their space has the same data layer.
Why process comes first
Twelve platforms felt like a tooling problem. The value stream map showed it was a process problem — so we fixed the process before we built the data layer.
What we built
Cross-platform data pipeline
12 ad and commerce sources, one warehouse, one consistent schema, one source of truth.
Anonymisation proxy
Client data is stripped and tokenised before any AI prompt leaves the platform. Nothing identifiable reaches the hyperscalers.
MCP query layer
Any AI agent (Claude, in-house, future) can query the warehouse through a standard MCP interface, no bespoke connectors.
Self-serve client portal
Clients see their own data, dashboards built on the fly, with annotations from the PPC team explaining what moved and why.
On-demand report generation
Reports assembled in seconds instead of hours. No more multi-tab swivel-chair work.
Audit-ready backups
Versioned snapshots of every dataset. Client questions answered with proof, not memory.
Before vs after
| Before | After | |
|---|---|---|
| Data sources | 12 disconnected | 1 warehouse |
| Client report build | Multi-tab, hours | On-demand |
| Hyperscaler leak risk | Open | Closed by proxy |
| AI agent access | Bespoke per tool | One MCP layer |
One source of truth across all 12 platforms, queryable by any AI agent, with nothing identifiable leaking to Big Tech.
12 → 1
Fragmented platforms unified into one queryable source of truth
0
Identifiable client records reaching the hyperscalers
Seconds
To assemble a client report that used to take hours of multi-tab work
Self-serve
Clients answer routine questions without paging a PPC manager
Twelve fragmented platforms became one knowledge layer any AI agent can query. Routine reporting and client questions now answer themselves. And not one identifiable client record reaches the hyperscalers.
“Twelve platforms felt like a tooling problem. The value stream map showed it was a process problem. Once you see where the value is actually created, the technology is the easy part.”
Sound familiar?
- Are your client reports being stitched together by hand from half a dozen ad platforms every cycle?
- Is your team pasting raw client data into AI prompts without an anonymisation layer in between?
- Do clients page your account managers for routine data questions a self-serve portal could answer?
- Could a unified data layer become a line item you sell, instead of a cost you absorb?