KPMG
GLOBAL ACCELERATOR CATALOGUE · DATA & AI

Data Platform Agents

A reusable KPMG accelerator that takes a Microsoft Fabric engagement from stakeholder evidence to governed requirements, dimensional models, built tables and reviewable pull requests.
Microsoft FabricMulti-agent deliveryDiscovery → buildAnthropic Claude Opus
WHAT IT IS

Delivery capacity, packaged—not another chatbot

Specialist agents operate a controlled engagement workflow: they extract cited requirements, maintain decisions, profile sources, design the model, generate implementation configuration, execute in Fabric and attach evidence to the pull request.

What the accelerator doesthree connected suites, one governed hand-off

01

Discover

Turns transcripts, questionnaires and emails into a cited requirements catalogue, decision log and human-action queue.

02

Model

Designs a Fabric Medallion platform, using Kimball dimensional modelling for business-ready Gold data products.

03

Build & prove

Turns mappings into Delta-Gen assets, runs Fabric, verifies tables and data quality, then opens an evidence-backed pull request.

Where it fitsadopt one stage or the complete delivery model

ANY ENGAGEMENT

Discovery

Useful wherever requirements, approvals and decisions are fragmented across people and documents—regardless of target technology.

STRONGEST FIT

Greenfield & modernisation

The full accelerator delivers most value where the Fabric architecture, modelling standards and delivery controls can be shaped from the outset.

SELECTIVE ADOPTION

Brownfield programmes

Modelling or Build can be introduced around an existing estate when mappings, repository conventions and target patterns can be aligned.

What the client keepsdeliverables now, enduring capability afterwards

CONCRETE ASSETS

A working, governed platform

Cited requirements, decisions and source profiles become implementation-ready models and mappings.

  • Delta-Gen configuration and repository template
  • Tested Fabric tables and data-quality evidence
  • Reviewable pull requests and delivery documentation
ENDURING CAPABILITY

A platform they can continue to build

Reusable standards and automation help lean client teams add data products consistently without depending on scarce specialists.

  • Repeatable delivery process
  • Explicit human approvals and ownership
  • Version-controlled, auditable evidence trail

Evidence from the reference implementationtransparent proof, including the gaps

15/15
Fabric tables verified
latest completed Kingsway build evidence · 13 July 2026
100%
traps passed
behavioural checks
133
items under management
catalogue items across engagements
100%
citation coverage
items backed by a verbatim citation
37
needs a human
open questions · approvals · high DQ

Needs a humanthe queue only a person can move

typeitemengagement
open questionOQ-001 What is the authoritative business definition of like-for-like sales for the trading pack?kingsway
open questionOQ-002 What exact revenue, margin, and discount definitions should be used so outputs match management accounts?kingsway
open questionOQ-003 What source or business rule should be used to repair the current margin-report failure when clearance lines hkingsway
open questionOQ-004 Should revenue for the pack include or exclude card fees and supplier funding, per Finance's view?kingsway
open questionOQ-005 How should like-for-like treat store closures and refits in addition to the fifty-two-week rule?kingsway
open questionOQ-006 Who owns the product hierarchy: trading or merchandising?kingsway
open questionOQ-007 Should supplier-funded markdown be netted off margin, and what is the exact treatment Finance requires?kingsway
open questionOQ-008 How should the model handle the known orphan store Croydon, which exists in StockRoom but is missing from the kingsway
+ 29 more

Catalogues being builtlive state per engagement

engagementitems draft → confirmed → approvedcitedopen Qs
Kingsway Retail Group
agent-built (run kingsway-2026-07-12-discovery-gpt54)
109
draft 109 · confirmed 0 · approved 0
100%12
Meridian Capital
reference catalogue
24
draft 5 · confirmed 15 · approved 2
100%3

Agent healthper-stage outcomes across all runs · questionnaire ≈ 84.7s · transcript ≈ 95.2s · email ≈ 96.5s

stageokfailedsuccess
catalog-maintain131
93%
normalize140
100%
extract140
100%
contradiction-find140
100%
decision-track130
100%
sync-decision-index130
100%
gap-analyse130
100%

Historical regression evidenceJuly test artefacts · not a current-release benchmark

runstage scoretrapsverdict
kingsway-2026-07-13-build
azure_openai · 2026-07-13T13:06
100%0/0star schema verified
kingsway-2026-07-12-modelling-from-discovery
azure_openai · 2026-07-12T22:34
0%0/0incomplete
kingsway-2026-07-12-modelling-golden
azure_openai · 2026-07-12T22:33
29%0/0mappings rendered
kingsway-2026-07-12-discovery-gpt54
azure_openai · 2026-07-12T22:17
0%6/6diverges
kingsway-2026-07-10-gpt54
azure_openai · 2026-07-10T12:26
0%6/6diverges