Alphabet puts AI agents to work in finance: invoice processing today, treasury next
Alphabet's finance team is now using agentic AI to process, pay, and reconcile invoices, according to CFO Anat Ashkenazi on the company's Q4 earnings call. The technology is also being rolled out in treasury. Inside engineering, AI agents already assist heavily - about half of the company's code is written with their support.
The message is simple: automation isn't a side project. It touches engineering, back office, and core finance workflows.
Why this matters for finance leaders
AI agents can take on work that doesn't fit neatly into a single system or rule. They coordinate tasks across apps, handle exceptions, and keep humans in the loop where judgment is needed.
- High-fit processes: invoice intake and three-way match, bank statement parsing, purchase order checks, account reconciliations, internal reporting, anomaly detection, and compliance monitoring.
- Analyst outlook: by 2028, about one-third of enterprise software is expected to include agentic AI, and at least 15% of day-to-day work decisions may run autonomously. The tech is early, but the direction is clear.
What to do now: a focused playbook
- Start where the volume and rules are known: accounts payable. Use agents for data extraction, vendor validation, coding, and three-way match. Keep humans for exceptions and high-value approvals.
- Add guardrails before scale: role-based access, dual controls, immutable audit logs, and confidence thresholds that route edge cases to reviewers.
- Pilot in treasury with narrow scopes: daily cash positioning, bank fee validation, FX exposure rollups, and alerting on unusual movements. Keep trading and policy decisions under human approval.
- Integrate with what you already use: ERP, procurement, banking portals, and collaboration tools. Minimize swivel-chair work by letting agents post, attach evidence, and tag GL accounts directly.
- Close the data gaps: clean vendor masters, standardize bank statement formats, and redact PII where possible. Validate external data feeds before agents act on them.
- Measure with operational metrics, not hype: cycle time per invoice, touchless rate, first-pass yield, exception rate, duplicate payment rate, and impact on DPO and cash forecasting accuracy.
- Upskill the team: prompt patterns, exception handling, and control reviews. Reassign saved time to analysis, vendor terms, and cash optimization.
A 90-day implementation outline
- Days 0-30: Map the current AP process. Label 500-1,000 invoices across vendors for training and testing. Set policies for approvals, thresholds, and audit evidence.
- Days 31-60: Deploy an agent to handle intake and coding on a subset of vendors. Keep humans in the loop for all postings. Track accuracy and exception reasons.
- Days 61-90: Expand to three-way match and automated postings under confidence thresholds. Add treasury read-only pilots for cash positioning and anomaly alerts. Formalize controls and documentation.
Risk and control checklist
- Human-in-the-loop for payments, GL postings over threshold, and any policy exceptions.
- Segregation of duties enforced across agent actions and human approvals.
- Immutable activity logs, evidence attachments, and versioned prompts/instructions for audit.
- Data protections: encryption, least-privilege access, and vendor data verification before payment.
- Periodic model and prompt reviews to prevent drift; simulate edge cases quarterly.
- Adopt a recognized framework for AI risk management, such as the NIST AI RMF.
What to watch next
As agentic AI gets embedded into more software, the boundary between "automation" and "decisioning" will blur. Expect stronger native integrations into ERP and treasury platforms, better audit trails, and clearer lines between assistance and autonomy.
Two practical tests for any new use case: can you prove the control works under audit, and can you show a measurable improvement in cycle time or accuracy within one quarter? If not, shrink the scope and try again.
Resources
- Alphabet earnings and updates: Investor Relations
- Curated tools for finance teams testing AI: AI tools for Finance
Bottom line: Alphabet's move signals where finance ops are heading. Start with invoices, ship a controlled pilot, and let the results fund the next step in treasury.
Your membership also unlocks:
