HPE's CFO: Making Agentic AI Work in Finance

High-level summary

• Marie Myers (CFO, HPE) describes HPE’s deployment of agentic/generative AI across a ~3,600‑person finance organization to drive speed, accuracy, and scaled operational insight.
• The approach treats AI as an iterative, enterprise transformation — not a point tool — with strong emphasis on data, governance, determinism, human oversight, change management and ROI discipline.

“Treat AI as an iterative, enterprise transformation — not a point tool.”

• Practical starting points: transactional finance (accounts payable, credit & collections, contract analysis, internal audit) and enterprise areas with clear productivity upside (software engineering, marketing, IT).
• Key platform/partners: Deloitte’s Zora platform (HPE‑branded as “Alfred”), NVIDIA (collaborated on determinism/NIM), and private‑cloud/on‑prem architecture for sensitive financial data.

Frameworks, playbooks and processes

• ROI framework

  • Separate direct (hard savings) vs indirect benefits (speed, accuracy, fraud reduction, time‑to‑insight).
  • Define an “indirect bucket” and quantify where possible; use scoring to prioritize bets and stage‑gate decisions.

• Stage‑gate / experiment-driven investment

  • Run small pilots, iterate rapidly, track outcomes, and stop projects that don’t meet ROI/scorecard thresholds.

• Workflow‑first deployment playbook

  • Redesign workflows and operating models (for example, centralize FP&A) before layering agents.
  • Standardize processes to enable scalable agent deployment.

• Data foundation playbook

  • Build and verify a single source of truth (“one time data”).
  • Reconcile and clean data before AI; treat data quality as a gating factor.

• Determinism & governance playbook

  • Require deterministic outputs for finance queries.
  • Co‑engineer or tune models (HPE worked with NVIDIA NIMs) to ensure consistent answers.
  • Prefer private cloud / on‑prem when necessary for compliance/security.

• Human‑in‑the‑loop governance

  • Maintain human oversight for judgmental, regulatory, or high‑risk decisions.
  • Define responsibilities across business owners, IT/data custodians, compliance, and audit.

• Change management & capability building

  • Sustained training and organizational learning (HPE trained thousands).
  • Focus on culture and adoption, not just technology.

Key metrics, KPIs and data points

• Organization & scale

  • Finance organization: ~3,600 people.
  • HPE trained >3,000 employees on AI capabilities.

• Data & platform scale

  • Alfred (HPE’s Zora instance) contains ~500,000 data elements.
  • Alfred was on version 65+ during development.

• Process cadence

  • “Ops call”: weekly enterprise operational review (Monday midday) automated/enhanced by Alfred.

• Suggested KPIs to track

  • Determinism / consistency rate (same answer every time for the same query)
  • Time‑to‑insight / time‑to‑decision
  • Error rate / invoice error reduction / fraud incidents
  • Throughput (invoices processed, collections actions)
  • Forecast accuracy and variance
  • Employee adoption / training completions
  • Cost of models/infrastructure vs productivity gains (ROI)
  • Number of agentic workflows deployed and iteration/version cadence

Concrete examples / case study actions (HPE)

• Platform & scope

  • Started with Deloitte Zora for transactional finance (AP) and expanded into the weekly ops call via HPE’s Alfred.

• Security model

  • Used private‑cloud / on‑prem deployment to retain control of sensitive financial data and meet compliance requirements.

• Determinism engineering

  • Co‑engineered model behavior with NVIDIA to move from probabilistic to deterministic outputs for finance queries.

• Workflow redesign

  • Centralized FP&A and standardized end‑to‑end workflows before deploying agents.

• Prioritized use cases

  • Finance (early): accounts payable, credit & collections, contract/accounting reconciliation, internal audit, controls checking.
  • Enterprise (high ROI): software engineering/code generation, marketing (creative/content campaigns), IT workflow automation.

• Human role

  • AI synthesizes and narrows areas of focus (e.g., flags top issues); humans validate, investigate, and execute corrective plans.

Actionable recommendations (priority order)

1. Build the data foundation: single source of truth; reconciled and clean datasets — non‑negotiable.
2. Redesign workflows first: centralize/restructure processes so agents can be effective and measurable.
3. Start small and iterate: pilot transactional workflows with clear metrics; apply agile development and versioning.
4. Insist on determinism for critical finance queries: engineer or configure models to return consistent answers.
5. Use private/on‑prem where compliance/security require it; otherwise evaluate hybrid models.
6. Bake in human‑in‑the‑loop controls for judgment, regulatory interpretation and final accountability.
7. Set clear stop/go criteria and monitor ROI closely; be prepared to reallocate capital.
8. Invest heavily in change management and continuous training (not one‑and‑done).
9. CFOs should lead ROI stewardship; work closely with CIOs, compliance/audit, and business owners.

Organizational and talent implications

• Role changes: reduce manual data‑gathering roles; shift finance skills toward analysis, judgment, cross‑functional partnering, and “AI + domain” fluency.
• Staffing model: fewer repetitive processing tasks; more emphasis on analytical, communication, and decision‑making capabilities.
• Learning agility: capture organizational learning from pilots (including failures) to accelerate subsequent deployments.

Governance, risk and regulatory stance

• Accountability remains human: final accountability for regulatory interpretation and finance outputs stays with people.
• Determinism & explainability: finance requires deterministic answers; explainability remains important and needs governance.
• Controls & audit: integrate compliance teams early; CIO/data custodians must enforce data controls.
• Security: private/on‑prem deployments used to control sensitive data and meet governance requirements.

Capital deployment & investment guidance

• Adopt an experimentation budget with clear tolerances for “throwaway” work; keep investments manageable and closely monitored.
• Expect models and infrastructure to evolve quickly; focus on building core skills & processes that transfer across model generations.
• Evaluate ROI not only on hard cost savings but also on speed, accuracy, risk reduction, and capacity to redeploy human capital into higher‑value work.

Advice for CFOs, CIOs and boards

• CFOs: act as stewards of AI capital — measure ROI, prioritize investments, and ensure data & governance; develop conversational-to-deeper tech acumen.
• CIOs: be central to deployment — data custodian, controls lead, governance steward; become a strategic partner to the business.
• Board members: require AI literacy; understand risks, deployment plans, and how AI can disrupt industry economics.

Risks and open challenges

• Non‑determinism and explainability challenges of generative models.
• Data quality as the primary blocker.
• Cultural resistance and the risk of “AI slop” (dropping standards).
• Rapid model lifecycle and circular financing in the AI ecosystem (chipmakers, model vendors, clouds funding one another) — judge investments by customer value and ROI.

Presenters and sources

• Marie Myers — Chief Financial Officer, HPE (primary presenter)
• Michael Krigsman — CXOTalk (interviewer/host)
• Deloitte — Zora platform partnership (referenced)
• NVIDIA — collaborated on determinism/NIM engineering (referenced)
• HPE internal platform — Alfred (HPE’s instance of Zora)
• Also referenced: Bill Briggs (Deloitte CTO, prior episode mention)