Case Studies  /  Harris Farm Markets
Fresh Food Retail AI-Led Development AI Hub Program Director

Harris Farm Markets — Enterprise AI Hub rollout with Claude Cowork & Claude Code

Designing and directing the enterprise-wide rollout of Anthropic's Claude platform across a 3,600-person family-owned Australian retailer — from early adopter program through to production agents driving margin improvement.

3,600

Employees across the business

10

Integrated workstreams

3

Phased rollout gates

11

Security domains + OWASP LLM mapped

~75%

AI-led delivery compression target

A capability shift, not just a tool rollout

Harris Farm Markets is a family-owned fresh food retailer with a passionate workforce, a strong culture, and ambitions to become an industry leader in AI adoption. When the opportunity arose to deploy Anthropic's Claude — via both Cowork (for business users) and Claude Code (for developers) — HFM needed more than a technology implementation. They needed a structured, governed program that would build lasting organisational capability.

Attain AI Advisory was engaged as AI Hub Program Director to design the rollout architecture, establish governance, and direct ten integrated workstreams spanning change management, security, data architecture, use-case pipeline, and a fundamentally new IT operating model.

The engagement was distinctive for its ambition: HFM became an early adopter of Claude Cowork during Anthropic's Research Preview and Frontier Program, positioning themselves ahead of the market while building the guardrails needed for responsible enterprise AI at scale. A critical early deliverable was a comprehensive AI security baseline — assessing the current-state architecture against OWASP LLM Top 10 risks, mapping the threat profile proportionate to a fresh food retailer (not a bank), and designing a phased security maturity path from interim MCP controls through to a fully governed Azure AI Foundry ecosystem.

Unlocking AI for 3,600 people — safely, at speed, with measurable ROI

1

New AI infrastructure

HFM had no existing enterprise AI platform, no AI governance model, and no standardised approach to AI-led development. Shadow AI was emerging unchecked.

2

Board risk appetite

Data and infrastructure security were the greatest concerns. The Board needed confidence in security, privacy, and contractual protections.

3

Culture-sensitive change

A family business with strong culture. AI adoption had to feel empowering — augmentation not automation — with job security addressed early and directly with messages of scale and better ways of working.

4

Kaizen lessons to embed

Previous technology implementations had generated 12 post-implementation Kaizen recommendations. Every one needed to be structurally embedded.

5

IT capacity constraints

IT needed to shift from sole code builder to coach and quality gatekeeper — while maintaining BAU operations and the D365 platform.

6

Tangible margin improvement

The business case demanded measurable ROI: capacity uplift, SaaS rationalisation, cost avoidance, and margin improvement — not just innovation theatre.

Three-phase rollout with readiness-based gates

Phased to build confidence progressively. Gates are readiness-based — not calendar-based — embedding the Kaizen principle that you don't advance until you're genuinely ready.

Phase 1

Support Early Adopters

Feb – March 2026
  • Claude Cowork deployed on macbook devices
  • 2–3 super-users onboarded per function as AI Citizens
  • Prompt Academy launched: Seedling through Growing levels
  • Co-CEO sponsorship activated with visible leadership behaviours
  • No production data touched — learn safely in interim secure environment
  • Claude Code exploration on Mac hardware outside HFM network
  • Key comms, myth-busting FAQs, and change management plan piloted
Phase 2

Stand-up Wider Rollout

Mar – May 2026
  • Secure Azure/Fabric data layer and Vector DB governance live
  • AI runtime security deployed with 11 security domains (D01–D11)
  • Agent design labs and pre-production IT checkpoints
  • Wider device rollout and full change management activation
  • D365 Copilot Studio agents and MCP in F&O integration
  • 6+ agents designed and built in sandbox
  • Board oversight with monthly SteerCo reporting
Phase 3

Embed & Improve

From May 2026
  • Full organisation availability
  • Agents and AI-led applications promoted to production
  • Niche SaaS replacement underway
  • Large data models in use across buying, finance, and operations
  • Guardrails and architecture proven at scale
  • AI Centre of Excellence operating in BAU rhythm
  • Continuous improvement and quarterly benefit reporting to Board

Ten integrated technology + change management workstreams

A governed, safe, and measurable rollout structured across ten interdependent workstreams — each with an assigned owner, defined activities, and clear phase-gate criteria.

WS1

Change & Communications

Prosci ADKAR-aligned change management. Prompt Academy with four progression levels. Future Fridays creative collision spaces. AI Workforce Mindset Personas with tailored interventions.

WS2

Copilot Cowork Logistics & Rollout

Claude Cowork deployed via Microsoft Copilot on existing Windows devices. M365 permissions govern data access. Private plugin marketplace curated by admins.

WS3

Secure Architecture & Technical Build

M365 tenancy architecture with Semantic Kernel orchestration, MCP plugin architecture, Foundry model deployment, and a 6-stage production promotion pathway.

WS4

Early Adopter How-To & Learning

AI Citizens ring-fenced at 20% of their time. Hands-on prototyping. Rubric-based quality framework (8+/10 minimum bar). Claude Buddies peer learning program.

WS5

AI Governance & Project Plan

Project charter, SteerCo, Board reporting, phase-gate criteria, AI risk register, independent specialist advisor, Kaizen recommendations embedded, NIST/ISO alignment.

WS6

Claude Code Rollout (IT)

Spec Driven Framework with 6-layer AI + human delivery system. Agent & Swarm architecture. AI-led SDLC stage gates. ~75% delivery compression target. Code quality gates via Snyk, CodeRabbit, and GitHub Advanced Security.

WS7

AI Data Architecture & Access

Data classification framework (Public → Restricted). Role-based access controls. PII/PHI masking. Vector DB governance. Trusted data pipelines from D365/Fabric. Model independence principles.

WS8

Security Layers & Risk Management

11 security domains (D01–D11) from AI workload runtime through to model security. Agent Control framework with tool whitelisting. AI Seatbelt runtime enforcement. Shadow AI detection via Netskope.

WS9

Use Case Pipeline & Benefit Tracking

Structured discovery workshops (3 use cases per function). Effort-impact-risk scoring. SaaS rationalisation tracker. Capacity uplift measurement. Quarterly benefit realisation reports for Board.

WS10

IT Operating Model & AI Hub

AI-First SDLC (BRD → FDD → DEV → QA → DEPLOY). Two-Speed model: AI-led development plus IT-governed production builds. IT shifts to coach, mentor, and quality gatekeeper.

Dual-path deployment model

We designed a dual-path model that maximises M365 governance for business users while preserving developer flexibility through Claude's native tooling.

Copilot Cowork via M365

For business users — the governed enterprise channel

Cloud-based within the M365 tenant, inheriting M365 security and identity. Full M365 Graph access to email, Teams, SharePoint, Calendar, OneDrive, and D365.

  • Microsoft DPA governs Anthropic as subprocessor
  • Entra ID, Purview DLP, conditional access, sensitivity labels
  • Agent 365 centralised governance and risk detection
  • Copilot Studio for custom agent builds
  • Scheduled recurring tasks — daily, weekly, on-demand automation

Claude Code Direct

For IT developers — the flexible build channel

Claude Code CLI for agent builds, code quality gates, SaaS replacement, and AI-led development. Deployed via Claude directly with Git for Windows.

  • Spec Driven Framework: 6-layer AI + human delivery system
  • MCP config per repo with tool scope whitelisting
  • claude.md files per project for agent context and behaviour
  • TDD enforcement with human approval at every key decision
  • OpenTelemetry export for real-time token and cost visibility

ADKAR-aligned governance with Board confidence

Governance was designed to give the Board confidence from day one. The structure is ADKAR-aligned (Awareness, Desire, Knowledge, Ability, Reinforcement) with clear escalation paths, decision rights documented in the Project Charter, and readiness-based phase gates — not calendar-based.

The Co-CEO serves as Executive Sponsor with Board comms accountability. A Business Sponsor drives LT collaboration and the business-view of the AI Hub. The Program Director holds strategic intent and workstream interdependencies. Every workstream owner is accountable for activities, timeframes, risk identification, and escalation.

Cadenced forums run from weekly stand-ups through to monthly Board updates, with phase-gate reviews requiring full ADKAR readiness assessment before progression.

Tier 1 — Strategic & Kaizen

Board Update (monthly), SteerCo (weekly, 45 min), Phase Gate Reviews (per gate). Chaired by Co-CEO. Escalation SLA: 24 hours to Board.

Tier 2 — Delivery & Workstream Management

Program Stand-up (weekly, 45 min), Business Alignment (fortnightly), Workstream Working Groups (weekly, 1 hour). ADKAR Health Check monthly.

Baselining the architecture — then making it secure and scalable

Before designing the future state, we conducted a comprehensive security baseline assessment. The question wasn't "what does a bank need?" — it was "what does a ~$1B fresh food retailer with 3,600 staff actually need?" The architecture had to be proportionate to the threat profile, not platinum-plated.

LOW
CTI Risk Score

Measured against peers in the Infotrust Cyber Threat Intelligence assessment (Feb 2026). Seven MCP servers connected with no runtime security controls.

OWASP LLM Top 10 — Prioritised for HFM
CRIT
Prompt Injection (LLM01)

Direct attack vector on AI agents accessing live financial and operational data via MCP

CRIT
Excessive Agency (LLM06)

MCP agents with broad database permissions manipulated into data extraction as agentic use expands

HIGH
Sensitive Info Disclosure

PII, pricing, margins, and HR data requires DLP inspection on AI traffic

HIGH
Secure the MCP Design

Cross-server privilege escalation across 7 active MCP servers — one compromised server could invoke all others

HIGH
Supply Chain Vulnerabilities

External skills and unvetted plugins could embed malicious instructions — bypassing guardrails

HIGH
Shadow AI & Uncontrolled Tools

Employees using AI without IT knowledge — browser extensions and unsanctioned tools creating uncontrolled data leakage

From loosely managed interim to governed enterprise ecosystem

Must-Have foundations for any AI deployment
MCP / Agent Governance

Control what tools agents can call. Rate limiting. Human-in-the-loop for sensitive actions. Read-only data governance as baseline with selected overrides.

Data Loss Prevention

DLP inspection on prompts before they leave the organisation. Block PII, BSB, ABN, TFN patterns. Shadow AI detection and blocking.

Identity & Access Control

SSO + MFA for all AI access. Role-based access control — buyers see pricing, support doesn't. Conditional access policies enforced.

Prompt Security

Runtime prompt injection detection. Content safety filtering on inputs and outputs. Covers OWASP LLM01 — the highest priority AI attack vector.

Audit & Visibility

Full audit trail — who prompted what, when, with what data. Tied to identity. Essential for Privacy Act compliance and cyber insurance.

Data Residency / ZDR

Zero Data Retention agreements. Data processed in-region where available. No model training on organisational data. DPA with residency clauses.

Phased Security Maturity — Cost-vs-Risk at Every Stage
Phase 1: Now — Interim

Lock in interim essentials

  • Entra SSO + MFA already in place
  • SQL-MCP read-only governance
  • AI policy (informative) deployed
  • Activate DLP inspection on AI traffic
  • Deploy prompt injection firewall
  • Establish audit logging in interim
  • Negotiate Zero Data Retention with LLM provider
Phase 2: Build — Core Security

AI Foundry + governed ecosystem

  • Azure AI Foundry deployment — all data stays within Azure
  • AI governance framework with agent guardrails and rules
  • MCP Gateway + Agentic Broker for tool control
  • Purview classification and sensitivity labelling
  • Full audit trail + monitoring dashboards
  • Skills management and version governance
  • CI/CD pipeline scanning for AI supply chain
Phase 3: Mature — Advanced Controls

Scale with confidence

  • AI attack path mapping and exposure management
  • Red team / eval gates in orchestration pipeline
  • Multi-model routing by cost, latency, and task
  • Advanced agent-to-agent orchestration
  • Rate limiting for non-human AI traffic
  • Automated adversarial testing in CI/CD
  • Continuous posture management at enterprise scale

Tangible outcomes designed into the architecture

Margin improvement pipeline

Structured use-case discovery workshops per function with effort-impact-risk scoring. Every use case tied to capacity uplift, cost avoidance, or revenue growth — tracked quarterly for Board reporting.

SaaS rationalisation

Systematic identification of niche SaaS subscriptions that Claude-built agents can replace. SaaS rationalisation tracker embedded in WS9 with cost-savings register and Board visibility.

Scale via AI, not FTE growth

The operating principle: grow capability through AI-augmented workflows rather than headcount growth. Capacity uplift measured in hours saved per function per month.

Enterprise security architecture

11 security domains (D01–D11) designed from scratch covering AI runtime, SaaS posture, shadow AI detection, prompt injection defence, code security, and model integrity.

~75% delivery compression

Claude Code's Spec Driven Framework with 6-layer AI + human delivery system targeting approximately 75% delivery compression on agent and application builds compared to traditional development.

Organisational AI capability

Prompt Academy with four maturity levels. AI Citizens across every function. A structured progression from Seedling to Mastering — building lasting capability, not tool dependency.

"Claude is not just a tool — it is a capability shift. Business users build solutions; IT is no longer the bottleneck."

— Harris Farm Markets CEO

Planning your own AI rollout?

Whether you're an early adopter or scaling enterprise-wide, we can help you design the governance, architecture, and change program to make it stick.

Start a Conversation → ← All Case Studies