Portfolio / Case Study 02

AI-First Finance
Operations

Deploying Microsoft Copilot, AI Builder & Copilot Studio Across a 300-Person Finance Function · Meridian Retail Group

£848K
Programme
investment
19mo
Full payback
from go-live
£571K
Year 1 benefit
delivered
£1.4M
5-year NPV
199% ROI
Background

The Organisation
& The Situation

The Organisation
Meridian Retail Group — three UK retail banners, approximately 1,200 employees and a centralised finance function of 300 staff. Core ERP: Microsoft Dynamics 365 Finance & Operations. Despite a modern platform, the finance team was operating in an almost entirely manual mode.
The Situation
Period-end close took seven working days — more than double the industry benchmark of three. AP teams processed 120+ invoice exceptions per day by hand. The IT helpdesk fielded 340+ finance queries per month, all answerable from D365 data. Staff turnover in Finance hit 25% annually — costing £480K/yr in replacement alone.
The Mandate
Deploy AI across Finance within 12 months using Meridian's existing Microsoft ecosystem
  • Cut period-end close from 7 days to 4
  • Automate 60% of AP invoice exceptions
  • Deflect 40%+ of Finance IT helpdesk queries via conversational AI
  • Reduce Finance staff attrition from 25% to below 18%
My Role — What I Personally Owned
Programme Manager · Programme Sponsor & Accountable Owner
  • Business case author — £848K board submission
  • Architecture decision owner (6 ADRs)
  • AI Centre of Excellence founder & chair
  • Vendor & SI management — fixed-price contract
  • Board reporting & C-suite stakeholder management
  • Change management programme lead
Section 1

The Challenge

Process Pain
Financial Cost
People Impact
7-day period-end close — more than double the 3-day industry benchmark. Entire team capacity consumed by data gathering and exception handling.
120+ AP invoice exceptions per day processed manually — 3 FTEs spending their entire working day on a single activity.
340+ IT helpdesk queries per month from Finance — PO status, invoice approvals, budget balances — all answerable from D365 data, none automated.
£480K/yr in Finance attrition costs — at £24K average replacement cost, 25% annual turnover across 300 staff was nearly enough to fund the entire AI programme.
FP&A variance commentary consumed 2 full days of senior analyst time per reporting cycle — zero capacity for strategic analysis.
D365 F&O held excellent data but no AI layer — every insight required manual querying; the platform investment was significantly underutilised.
Exit interviews confirmed repetitive, low-value work was the primary reason people left Finance — a retention crisis driven by process, not pay.
Senior analysts spending majority of time on data entry — analytical capability the business needed was unavailable because capacity was consumed by manual tasks.
No AI governance framework existed — deploying Copilot without one would expose Meridian to GDPR risk and audit findings on financial data handling.
CFO · Programme Kick-off

"We have brilliant people doing work that a machine should be doing. We need to fix that, and we need to do it this year."

Section 2

My Approach —
Key Decisions Made

I structured the programme around three principles: platform coherence first, governance before features, and adoption as a first-class workstream. Six Architecture Decision Records governed every subsequent delivery choice.

Decision What I Decided Why — The Reasoning
AI platform (ADR-AI-01) Microsoft-native only — Copilot for Finance, AI Builder, Copilot Studio, Azure OpenAI Evaluated Claude API, Salesforce Einstein, Llama 3. Microsoft won on D365 integration (no ETL), Azure UK South data residency, GDPR clarity, and existing Enterprise Agreement
Governance sequencing AI Centre of Excellence established in Week 2 — before any Copilot feature enabled in production Governance-first adds 3 weeks to Phase 1 but eliminates reactive remediation — which in my experience costs far more in time and organisational trust
Human-in-the-loop (ADR-AI-04) Risk-tiered: Tier 1 (financial postings — human approval required), Tier 2 (AI-drafted content — spot-check), Tier 3 (status queries — fully automated) Gave CFO and KPMG auditor confidence to approve features that would otherwise have stalled in governance review
Change management budget Ring-fenced £80K (12% of total) for adoption — defended against CFO challenge Gartner 2023: 60% of AI programme value lost through poor adoption. Committed to specific KPI: 85% of Finance staff using Copilot within 90 days
SI contract structure Fixed-price with named-resource clause — same model as ERP programme Budget certainty; prevents junior consultant substitution post-signature; milestone payments align SI delivery incentives
Adoption model AI Champions network — 8 Finance team members trained ahead of rollout as peer coaches Peer coaching outperforms formal IT-led training for adoption; champions become embedded advocates rather than helpdesk tickets
AI Governance Controls I Established (Week 2 — Before Any Feature Went Live)
Risk-Tiered Human-in-the-Loop
Three tiers: financial postings require human approval (Tier 1); AI-drafted content requires spot-check (Tier 2); status queries are fully automated (Tier 3). KPMG specifically referenced this framework at year-end audit.
Data Loss Prevention Policy
MRG-AI-Finance-DLP configured before any Copilot Studio maker access was granted — preventing finance data from reaching any unapproved connector or external endpoint.
AI Model Cards
Produced for every deployed model, reviewed quarterly. The artefact KPMG specifically requested at year-end audit — demonstrates model oversight and accountability.
CoE Dashboard
Real-time monitoring of Copilot usage, DLP policy hits, and model accuracy. Caught two DLP violations before they became incidents across the 12-month programme.
Section 3

What Was Built
& Deployed

Four AI capabilities deployed across a 12-month programme, each designed and owned by me, delivered by a Microsoft Gold SI partner under a fixed-price contract with named resources.

Capability What I Designed & Built Outcome at 12 Months
Copilot for Finance
D365 F&O embedded 		Feature flag activation, RBAC configuration, risk-tier classification for period-end and collections use cases, KPMG briefing on human-in-the-loop controls Period-end close: 7 days → 4 days. First AI-assisted close delivered in Month 5.
AI Builder
Invoice Processor 		Pre-built invoice model deployed, Power Automate routing flow built (90% confidence threshold for auto-route, below-threshold to human queue), AP mailbox trigger configured 62% of invoices auto-routed from Month 1. AP team FTE redeployed to supplier relationship management.
Copilot Studio
Finance Agent 		Agent built in Copilot Studio, system prompt authored (3 iterations), D365 OData connector via Power Automate (service principal auth), published to Teams. 12 topics: PO status, budget queries, invoice approvals, open items 340 monthly helpdesk queries reduced to under 80. Agent handles 76% of inbound finance queries without escalation.
M365 Copilot
FP&A & Reporting 		60 Copilot licences deployed to Finance and FP&A. Prompt library developed for variance commentary, board pack drafting, budget narrative generation. Power BI + Copilot integration for AI-enriched dashboards FP&A variance commentary: 2 days → 4 hours. Senior analyst capacity freed for strategic analysis.
Section 4

Outcomes &
Business Impact

7d → 4d
Period-end close
76%
Helpdesk query deflection
91%
Copilot adoption in 90 days
17%
Staff turnover (down from 25%)

At the 12-month benefits review, the programme delivered against all five primary KPIs and exceeded three of them.

KPI Baseline Target Achieved
Period-end close cycle 7 days 4 days 4 days ✓
AP exceptions auto-routed 0% 60% 62% ↑
Finance helpdesk deflection 0% 40% 76% ↑
Staff turnover (Finance) 25% 18% 17% ↑
Copilot adoption — Finance staff 0% 85% in 90 days 91% ↑
Annual benefit delivered (Year 1) £0 £552K £571K ↑
Finance Director · Meridian Retail Group · 12-Month Review

"The Finance team are working completely differently. The change is visible. People are doing analysis again, not data entry."

Section 5

Lessons
Learned

What Worked
AI Champions network delivered 91% adoption — peer coaching outperformed formal training alone by a wide margin. Now being replicated for the Procurement rollout.
Governance-first setup meant zero reactive remediation in 12 months — the CoE dashboard caught two DLP violations before they became incidents.
Fixed-price SI contract with named resources held — no scope creep, no substitution of junior consultants post-signature.
Risk-tiered human-in-the-loop gave the CFO and external auditor confidence to approve features that would otherwise have stalled in governance review.
What I Would Do Differently
Data quality sprint — 6 weeks not 4. We hit data issues in Month 3 that slowed AI model accuracy and required a mid-programme remediation sprint.
Involve Finance team in prompt design from Day 1, not as reviewers at the end. The Copilot Studio system prompt needed three iterations before agent tone was right.
Finance Lake (Synapse Link) should have been scoped into Phase 1. The OData API approach hit rate limits at higher query volumes than modelled.
Add 10% licence contingency for Microsoft AI deployments. Power BI Copilot integration required additional Premium capacity not in the original licence budget.
Section 6

Supporting
Artefacts

The following documents were produced during this programme and are available as part of this portfolio.

MRG-IT-BC-AI-001
Business Case — AI-First Finance Operations
Board-ready submission: £848K investment, £552K annual benefit, 19-month payback, 5-year NPV £1.4M (199% ROI). Approved by CEO and CFO.
Available on request
MRG-IT-ADR-AI-001–006
Architecture Decision Record (6 ADRs)
Covering: AI platform selection, LLM provider evaluation, agent framework, human-in-the-loop policy, data strategy, and rollout sequence.
Available on request
MRG-Phase2-Workbook
Hands-On Configuration Workbook
18-step technical workbook: D365 Copilot setup, AI Builder model creation, Copilot Studio agent build, and AI CoE governance configuration.
Available on request
MRG-D365-Screenshots
D365 Configuration Evidence
Screenshots of D365 v10.0.46 Copilot features, AI Builder model testing, Power Automate flows, and Copilot Studio agent in Teams.
Available on request
MRG-AI-Gov-001
AI Governance Policy
Responsible AI policy including risk-tier classification, model card template, DLP policy definition, and AI CoE charter. Referenced by KPMG at year-end audit.
Available on request
Note on Confidentiality: Company name, financial figures, and personal details have been anonymised for portfolio use. The architectural approach and delivery model reflect proven practices from previous implementations. The financial figures have been scaled to suit the scenario and are intended to be directionally representative rather than exact client data.