Despite the structural limits outlined in our first article How P2P platforms are actually changing in the era of AI, artificial intelligence has delivered real and defensible improvements inside procure-to-pay (P2P) platforms. These gains are visible in production environments, measurable in operational metrics and widely adopted because they address specific sources of friction without challenging existing control models.
AI did not enter P2P platforms as a strategic redesign. It entered as a pragmatic response to operational friction, targeting narrow, well-defined problems where outcomes could be measured objectively: accuracy, speed and reduction of manual effort. In these areas, AI has delivered real, defensible improvements.
The most visible gains have been in document capture and data extraction
Invoice digitization is the clearest example. Traditional OCR struggled with layout variation, low-quality scans and unstructured content. Machine learning models trained on large invoice datasets significantly improved header and line-level extraction accuracy, reduced manual correction rates and enabled higher straight-through processing. In practice, this translated into fewer keystrokes, faster cycle times and lower cost per invoice.
Similar improvements appeared in classification and coding. Instead of relying entirely on static rules or templates, ML models learned from historical postings to suggest GL accounts, cost centers and tax codes. These suggestions were not perfect, but they reduced the cognitive load on AP clerks and improved consistency over time. Importantly, they operated within existing controls, making them easier to adopt.
Routing and prioritization showed similar gains
AI-assisted routing models helped determine the most likely approver or exception handler based on past behavior, organizational structure and workload. In environments with high document volume and frequent reassignments, this reduced idle time and manual forwarding. In approval scenarios, reminders and nudges driven by simple predictive models shortened approval cycles without changing governance.
User interaction improved incrementally
Natural language search and conversational interfaces made it easier for users to find invoices, check status and retrieve basic information without navigating complex menus. For casual users, this lowered the barrier to entry and improved perceived usability. For power users, it reduced time spent on routine inquiries.
Analytics advanced from static reporting to early prediction
Some platforms began applying predictive models to identify late-payment risk, duplicate invoice probability or potential overpayments. These models worked best when paired with clear thresholds and human review, acting as early warning systems rather than autonomous decision-makers. In finance teams, this shifted effort from reactive cleanup to proactive monitoring.
These improvements share important characteristics:
- They operate within the existing document–workflow–rule paradigm.
- They optimize steps rather than redesign processes.
- They are assistive, not authoritative.
- They improve efficiency without altering accountability.
This is precisely why they succeeded initially.
They did not challenge governance models, audit structures or user roles. They made existing processes faster, cheaper and slightly more reliable. For most organizations, this was the right trade-off. However, these gains also reveal a ceiling.
As transaction complexity increases, supplier ecosystems diversify and policies vary across regions and business units, the same AI techniques begin to show diminishing returns. Accuracy improves, but exception volumes remain high. Routing becomes faster, but approvals still stall. Predictions exist, but actions remain manual.
At this point, adding more AI does not change outcomes. It just accelerates the point at which humans must intervene. This is not because the models are weak. It is because the underlying system was never designed to absorb intelligence beyond optimization.
In the next article, we will examine this plateau in detail: where AI additions stop changing outcomes and why the problem is structural rather than technological.
Read more on this topic on our dedicated ‘AI in Procurement’ page, and feel free to reach out to discuss how you can leverage our analysts’ knowledge.
