Reporting from days to seconds
A centralised analytics brain for the business — marketing, product and financial data unified into one system, with agentic automations replacing manual reporting work outright.
Where this expertise comes from: unifying fragmented data into one AI-queryable layer.
1. The challenge
- Marketing reporting lived in HubSpot
- Sales reporting was scattered across spreadsheets
- Product analytics sat in a separate BI tool
- A data lake existed, but only a handful of data engineers could query it
- Some critical data wasn't connected to anything at all
- Every question meant filing a request, then days to reports, and limited insights
2. The real work: cleaning up the foundation
Before any AI layer could be trusted, the underlying data had to be fixed. This is the part most companies underestimate, and it's the part that actually determines whether the whole thing works. Years of inconsistent field naming, duplicate records, and disconnected systems in the CRM had to be reconciled by hand. Spreadsheets living outside any official system had to be tracked down and folded in. Server-side tracking data on user journeys had to be verified against the CRM, so the two sources actually agreed with each other. None of this work is glamorous, and it rarely gets budget or credit, but it's the foundation everything else sits on. Skip it, and an AI layer just gives confident wrong answers faster. Get it right, and everything downstream becomes trustworthy.
3. The approach
| Action | Why |
|---|---|
| Revision of all fragmented data sources — sync, consolidation & centralisation | Data was scattered across HubSpot, spreadsheets, and a separate BI tool, making a single source of truth impossible |
| Legacy CRM field cleanup and consolidation, defining all fields | Years of inconsistent naming and duplicate records meant the same data point could exist in three different forms |
| MCP integration of all sources into the LLM | Connected data needed to reach the AI layer in a structured way for it to be queryable at all |
| Creating skills / gotchas / context for Claude to pull the right data | Raw access isn't enough; the AI needed guidance on which source to trust and how to interpret it correctly |
| Data validation against source systems | Prevented the AI layer from confidently returning wrong answers built on bad data |
| Defining access and governance | Opened self-serve querying to leadership while protecting sensitive fields |
| Iterating based on real usage | Refined the skills as actual questions revealed gaps the initial build hadn't covered |
On access & governance
Not everyone needs to see everything. Role-based access controls were applied so the AI layer surfaced the right level of detail to the right audience — summary metrics for leadership, granular data for the teams who needed it. Sensitive fields, like personal identifiers or commercially confidential figures, were masked or excluded from the queryable layer entirely, so opening up self-serve access didn't mean opening up everything.
4. The outcome
- Leadership went from waiting on manual reporting to getting polished, accurate answers themselves, in seconds
- Self-serve analytics through conversation — no dashboards, no waiting, no bottleneck
- Outputs included attribution modelling, pipeline snapshots and projections, financial reporting, and product adoption funnel reports
Want results like Ably's?
Book a free 30-minute call and we'll tell you honestly whether AI Brain fits your situation.
✨ Book a free AI strategy consultation