Getting started with AI in Source-to-Pay

[ronpeperkamp]

Why S2P is the right place to start with AI

Most AI adoption conversations in the enterprise stall because they lack an entry point. Source-to-Pay avoids that problem. It has three structural advantages that make it the ideal proving ground—before you expand AI into less-structured parts of the business.

Structured: data already in ServiceNow—no prep project needed. (Invoices, POs, and supplier records are clean, consistent, and high-volume—ideal AI input. No data preparation project needed.)

Defined: workflows with measurable start/end states. (Approval chains, exception handling, and supplier onboarding have clear start and end states. AI outcomes are measurable because the process has measurable steps.)

Tracked: KPIs already in place—AI maps to existing metrics. (Tracked KPIs.  Cost savings, cycle time, and touchless rate are already measured in your S2P deployment. AI impact maps directly to metrics the business already watches.)

The three-filter use case selection model

The most common reason AI initiatives stall is not a technology problem—it is a selection problem. Use this three-filter model to identify the right first use case before touching any configuration.

Filter 1 — High pain, low complexity.

Identify a process where manual effort is visible, complaints are frequent, and the data already exists in ServiceNow. Invoice exception handling and supplier onboarding are the two most common S2P candidates.

Filter 2 — Measurable outcome.

The use case must have a baseline metric today—for example, a 4-day average exception resolution time—so you can prove impact within 90 days. If you can’t measure the before, you can’t prove the after.

 Filter 3 — Contained scope.

Start with one region, one business unit, or one supplier tier. Controlled pilots generate clean data, build internal champions, and reduce risk. Broad rollouts at launch are how pilots turn into escalations.

Tip : Run the three filters in a 45-minute whiteboard session with the AP manager or procurement ops lead before any technical work starts. Output: a one-line use case brief—“We will automate invoice price-variance exceptions for UK suppliers, targeting resolution time from 4 days to under 1 day, by [date].” That brief becomes your 90-day success criterion.

Use case 1 — APO: Invoice exception handling and supplier inquiry resolution

APO delivers two distinct AI-powered capabilities that together reduce manual AP workload. They address different problems and are activated separately. Understanding the distinction is important for scoping.

Capability A: Exception handling within the invoice processing playbook

The APO invoice processing playbook guides AP specialists through a structured end-to-end workflow—from invoice capture through DocIntel, PO matching, duplicate checks, exception review, and approval. The exception review step surfaces mismatches for human resolution within the Source-to-Pay Workspace. Now Assist contributes at-a-glance summarisation within the workspace—helping AP specialists understand invoice context faster without hunting across multiple records. The human specialist reviews and acts; AI reduces the reading and lookup time.

Capability B: AI Agent — Resolve Supplier Questions (Agentic AI)

This is the flagship agentic AI capability in APO (Yokohama/May 2025 store release). It autonomously resolves high-volume, repetitive supplier inquiries—such as “what is the status of my invoice?” and “where is my payment?”—which overwhelm AP teams and cause delays.

A multi-agent system captures inquiry details via the Supplier Collaboration Portal, auto-analyses the case, drafts a resolution, and surfaces it to the AP agent for review within the same workspace. The human agent can accept, edit, or override before the response is sent. SLA clocks pause if supplier confirmation is required.

Business impact

• Radically faster response times to common invoice and payment queries
• Frees AP teams to focus on exception-based and complex work
• Elevates supplier experience through timely, consistent communication
• Reduces overall AP inquiry volume and team workload

Use case 2 — Supplier onboarding acceleration

AI agents autonomously manage supplier onboarding by validating data, checking for duplicates, generating tailored tasks, and handling outreach and reminders—all without waiting on supplier manager actions. Onboarding cycle time drops while data quality improves.

Autonomous workflow execution from intake to qualification

Dynamic task generation by supplier region and industry

Intelligent data validation, duplicate detection, flag inconsistencies

What the AI agent does

• Autonomously picks up onboarding cases and drives progress without waiting on supplier manager actions.
• Generates tailored onboarding tasks based on supplier region and industry using policy knowledge articles.
• Performs data checks, verifies documents, detects duplicates, and flags inconsistencies before human review.
• Handles outreach, reminders, and routing end-to-end.
• Escalates only when human judgment is required.

What the agent does not do (by default)

Sanctions screening (OFAC, EU, UN) is typically handled by a third-party integration or a separate risk assessment step in the onboarding playbook — not a default AI agent behaviour.

If your customer requires automated sanctions screening, confirm whether that integration is in scope and licensed separately.

Note on poor data quality

Poor supplier portal data quality is not a reason to delay this use case—it is what this use case fixes.

The AI agent generates structured, precise re-requests instead of generic rejection emails. Re-request rates drop as a direct byproduct of adoption.

Use case 3 — Spend categorization correction

The Spend Categorization Agent (part of Now Assist for SPO) uses historical patterns and AI predictions to classify procurement records—sourcing requests, purchase requisitions, and purchase orders—into the correct spend category. It flags mismatches and alerts users when the selected category differs from the AI recommendation.

How it actually works:

A procurement request (SR / PR / PO) is created or updated in the Source-to-Pay Workspace.

The Spend Categorization Agent applies historical patterns and AI predictions to classify the record into the correct spend category. The agent surfaces its recommendation inline showing the predicted category and confidence. The user can accept or override. If the selected category differs from the AI recommendation, the agent flags the mismatch and alerts the user.

In cases of limited training data, the agent uses LLM reasoning to make a prediction rather than pattern-matching alone. The agent can be configured to different confidence thresholds, giving admins control over when predictions are surfaced vs. suppressed.

What this delivers

• More accurate spend categorization at the point of procurement entry—reducing downstream miscoding
• Reduced manual audit burden—errors are caught inline, not discovered in a quarterly spend review
• Improved spend visibility for category managers and CPOs over time as category accuracy builds
• Configurable confidence thresholds—deploy conservatively, expand as model accuracy is validated

The quick-audit technique

Before deploying, run a manual spot-check: pull 20–30 recent procurement records from your highest-value category.

Ask a category manager to classify each one correctly. Compare to the existing category assignments.

In most environments, 15–30% of records will be miscoded. Show the CPO that slide. It creates immediate urgency without a data project.

Why this matters downstream

Accurate spend categorization is the data foundation for everything else: supplier consolidation, contract compliance, maverick spend identification, and sourcing prioritisation.

Fixing categorization at the point of entry is more effective than cleaning it up after the fact.

The 90-day launch model

Customers who follow a phased approach consistently outperform those who attempt broad rollouts. This three-phase model applies to all three use cases. Adjust the specific actions for the use case selected.

Days 1–30 Discover and define	Days 31–60 Enable and pilot	Days 61–90 Measure and expand
✓  Run AI Readiness Assessment	✓  Activate capability per use case	✓  Deliver impact report vs baseline
✓  Apply three-filter model	✓  Enable shadow mode	✓  Present ROI at QBR
✓  Establish baseline KPIs	✓  Train key users	✓  Identify next use case
✓  Map technical prerequisites	✓  Monitor AI accuracy vs human decisions	✓  Build expansion roadmap
✓  Identify pilot scope and team	✓  Tune routing and confidence thresholds	✓  Update mutual success plan

Prerequisites summary

Before starting any of the three use cases, confirm the following are in place. A missing prerequisite is a scoping action in the first 30 days—not a reason to delay.

Requirement	APO (invoice / inquiry)	Supplier onboarding	Spend categorization
Product	APO (Accounts Payable Ops)	SLO + Supplier Portal	SPO + Spend Category Mgmt
AI capability	Now Assist for APO 3.0.X	Now Assist for SLO plugin	Now Assist for SPO plugin
Data minimum	Active invoice processing in APO	1 active supplier tier + docs	Active SR/PR/PO volume in SPO
Human role	AP manager + AP specialists	Compliance / supplier reviewer	Category manager
External integration	Optional (D&B for enrichment)	Optional (D&B / ERP)	Not required
Shadow mode	Yes — required before go-live	Yes — parallel validation run	Yes — validate AI acceptance rate
Earliest signal	Within first sprint post-activation	Next supplier onboarding cycle	Next procurement cycle post-activation

Common objections and responses

These are the six objections that come up most consistently in S2P AI conversations.

"Our exceptions are too complex for AI to handle."

Response

AI doesn’t handle exceptions—your AP specialists do. Now Assist surfaces context and reduces the lookup time, which today takes specialists 15–20 minutes per exception. Start with high-volume, lower-complexity exception types to prove the model before expanding.

For autonomous inquiry resolution (AI Agent), the agent handles generic, repetitive supplier questions—not complex exceptions. Complex cases are escalated to a human automatically.

"We don’t trust AI to suggest financial decisions."

Response

The AI never approves or rejects anything—it summarises, predicts categories, or drafts responses. Your people make every decision. Shadow mode means you can measure AI accuracy against real decisions for 2–4 weeks before removing the human-first gate.

When customers see their actual shadow-mode agreement rates (typically high for well-scoped use cases), trust builds quickly.

"Our compliance team won’t accept AI for sanctions screening."

Response

The AI Supplier Onboarding Agent does not replace sanctions screening. Sanctions checks are handled by your existing screening service or a separate risk assessment workflow.

What the agent does is validate documents, check for duplicates, and generate tasks—freeing reviewers to focus on the risk assessment steps that need human judgement.

"We need to clean up our data first."

Response

Start with the supplier tier or procurement category where your data is cleanest—typically your top strategic suppliers or largest spend category, which represent the bulk of your volume.

For spend categorization, the AI catches miscoding at the point of entry going forward. It doesn’t require a clean historical dataset to start delivering value.

"Our taxonomy is too custom—a generic model won’t work."

Response

The Spend Categorization Agent learns from your procurement records and your category assignments. It uses historical patterns specific to your org, not a generic UNSPSC mapping.

In cases of limited training data, LLM reasoning supplements pattern-matching. The agent improves as it processes more of your records.

"This feels like a long project—we need quick wins."

Response

The Spend Categorization Agent can be activated and surfacing predictions within the first sprint—the quick win is at week 2: pull 30 recent procurement records and show the CPO the miscoding rate. That single data point creates more urgency than any ROI model.

The AI Agent for supplier inquiry resolution shows impact within the first onboarding cycle post-activation. Radically faster response times are visible immediately in SLA data.