AI Agents in Finance with HPE's Chief Financial Officer (CFO)

High-level summary

• Marie Meyers (CFO, HPE) describes HPE’s multi-year, enterprise AI/agent rollout for finance: an internal platform called “Alfred.”
• Focus areas: proving ROI, managing risk, maintaining determinism for financial data, and scaling agentic AI across ~3,600 finance employees.
• Core thesis:

  Treat AI investments as agile, iterative transformation programs — not one-off tooling purchases. Success requires (1) data quality and a single source of truth, (2) workflow and operating-model redesign before automation, (3) cross‑functional governance, and (4) human-in-the-loop controls where accountability remains with people.

Frameworks, playbooks and processes

• Direct vs. Indirect ROI framework
  • Direct: measurable cost/time savings, headcount/productivity effects.
  • Indirect: speed, accuracy, reduced errors/fraud, improved decision timeliness — defined up front so projects can be scored and compared.
• Stage-gate / iterative experimentation playbook
  • Small, manageable investments → rapid iterations → measure outcomes → continue, pivot, or kill the project.
  • Expect throwaway work; treat pilots as learning vehicles.
• Workflow-first approach
  • Redesign operating models before automation; standardize and centralize processes (example: centralizing FP&A) so agents can be applied effectively.
• Human-in-the-loop governance model
  • AI synthesizes and highlights issues; humans validate, apply judgment, and remain accountable.
• Determinism engineering (finance-grade AI)
  • Technical work and vendor co-engineering (e.g., with Nvidia) to make outputs deterministic and repeatable for financial reporting.
• Cross-functional “team sport” governance
  • Involve business leaders, IT/CIO, controls/compliance, audit, and CFOs for ROI stewardship and risk control.

Key metrics, KPIs and program indicators

• Scale and scope
  • Target rollout: agents across a ~3,600-person finance organization.
  • Alfred contains ~500,000 data elements used for deterministic queries.
  • Alfred development cadence: 65+ versions (iterative product maturation).
• People & training

  • 3,000 employees trained on AI capabilities.

• Operational cadence
  • Example: a weekly “ops call” replaced an old PowerPoint-based process; Alfred provides pre-analyses for human review.
• Measurement approach
  • Direct ROI metrics: time saved per process, errors reduced, AP/collections throughput.
  • Indirect metrics: speed of decision-making, reduction in error/fraud incidents, time-to-insight.
  • Use stage-gate milestones to determine continuation vs. termination of projects.

Primary use cases and examples

• Finance-specific
  • Transactional automation: accounts payable, credit, collections (where RPA had success).
  • Contract accounting/analysis: contract interpretation and mapping to accounting standards.
  • Controls and internal audit: automated controls checking and reconciliation.
  • FP&A: analytical view generation, Excel automation agents, forecasting and scenario analysis enabled by workflow change.
  • Operational meeting support: Alfred analyzes weekly performance and surfaces targeted issues for human investigation.
• Enterprise-high ROI areas
  • Software engineering: code productivity gains via generative AI.
  • Marketing: campaign and image generation workflows.
  • IT workflows: internal IT productivity and customer service support.
• Implementation example
  • Platform strategy: used a partner platform (“Zora”) and built Alfred internally; opted for private-cloud / on‑prem architecture for compliance/security.
  • Determinism: co-engineered with Nvidia on a NIM to convert probabilistic LLM outputs into deterministic, repeatable financial answers.

Risk, controls and governance

• Determinism is non‑negotiable for finance: outputs must return the same verified number irrespective of prompt/region.
• Data is foundational: cleaned, reconciled, single source of truth required before meaningful agent deployment.
• Regulatory and judgmental work: AI accelerates data gathering and initial analysis, but regulatory interpretation and final decisions remain human-controlled; legal/regulatory accountability stays with named humans.
• Architecture choice: private cloud/on‑prem used to maintain data security and compliance.
• Cross-functional teams: include IT/CIO (data custodians), business process owners, compliance, audit, and the CFO for ROI/governance.

Organizational and culture recommendations

• Redesign processes and operating models first (centralize FP&A as an example) before layering agents.
• Prioritize change management: human adoption and mindset change are often the hardest parts; poor change design causes pilot failures.
• Capture lessons and codify patterns so subsequent projects accelerate (organizational learning agility).
• Invest in large-scale, ongoing training — HPE trained thousands.
• Expect role evolution: finance professionals will move from data assembly to judgment/analysis; soft skills and analytical judgment increase in value.

Investment strategy and capital allocation

• Adopt an experimental/agile posture: small bets, frequent iterations, measure outcomes, and be willing to stop unsuccessful pilots.
• Recognize benefits beyond hard savings: speed and decision-quality improvements are important returns.
• Use stage gates to reallocate capital if ROI is not materializing.
• Develop internal skills and capabilities to preserve institutional learning as technologies evolve.

Actionable recommendations (practical takeaways)

1. Define direct and indirect ROI before funding projects; create a scoring approach and stage gates.
2. Clean and reconcile core data; implement a single source of truth before deploying agents.
3. Redesign and standardize workflows (centralize FP&A as a template) so agents operate on consistent processes.
4. Require human-in-the-loop for judgmental, regulatory, and final-decision workflows; retain accountability with people.
5. Engineer determinism for finance use cases (partner with vendors if necessary).
6. Use private-cloud/on‑prem for sensitive financial data when compliance and determinism are required.
7. Start small, iterate rapidly (expect many versions), measure frequently, and be prepared to kill non‑performing efforts.
8. Invest heavily in training and organizational change management from day one.

Limitations and market observations

• Marie notes industry interdependence (chipmakers, model providers, cloud providers funding each other) and cautions that the technology and economic landscape change quickly.
• Emphasis on learning agility rather than making massive up-front bets, since models and economics can become obsolete.

Presenters and referenced technologies

• Marie Meyers — Chief Financial Officer, HPE (speaker)
• Michael — host, CXOTalk (interviewer)
• Referenced partners/technologies: Nvidia (co-engineered determinism), Zora platform (partner platform), HPE internal agent “Alfred”