As the world stands by awestruck at how AI models leap forward in coding ability and reasoning depth, procurement and supply chain leaders are rightly asking themselves how quickly that power can be applied to their own department’s workload. Anthropic’s latest release is the latest example: a model posts eye-catching gains on complex, open-ended benchmarks, whetting the appetite of enterprise buyers who hope those gains will translate directly into value wherever they choose to deploy it. But the question worth asking first isn’t which model currently sits at the top of the leaderboard. It’s whether that kind of capability will be applicable to the specific area, in this case to Source-to-Pay.

Read also: How Artificial Intelligence Is Reshaping Global Supply Chains

Frontier Models Built For Different Needs

The newest generation of frontier models, including Anthropic’s latest, is explicitly optimized for days-long, complex and unsynchronized tasks such as multi-day coding projects and intricate research tasks. Source-to-Pay is a different animal almost entirely. Instead of problems that unfold over many steps without a fixed, repeatable format, the bulk of procurement activity is high-frequency and rules-governed. Three-way matching between purchase orders, receipts, and invoices; supplier onboarding against a fixed compliance checklist; approval routing that has to follow the same delegation-of-authority rules every single time. These are not the open-ended problems that Claude Fable 5 is designed to solve. Source-to-Pay problems instead tend to reward consistency, speed, and predictable cost at volume.

Cost of Tokens

Reasoning models are token-hungry as they think longer, explore more branches, and produce more output. This is necessary as they are dealing independently with multifaceted problems. When applied to a routine task such as invoice match, high token consumption becomes a liability: the model approaches every transaction like a novel puzzle, consuming high volumes of tokens for no reason as the vast majority of invoices are near-identical to thousands processed the week before. 

Claude Fable 5 costs $50 per million output tokens, roughly twice the cost of Opus 4.8.It thus routinely uses 500k to 1M tokens per task, regardless of complexity.With high enterprise transaction volumes, a fractional per-task cost increase soon balloons, despite not adding anything in terms of improvement in outcome quality, because the task didn’t need that depth of reasoning in the first place.

Governance Can’t Be an Add-On

In procurement, any decision is something a human has to be able to defend. When a system flags a supplier as compliant, recommends an approval pathway, or surfaces a contract risk, an auditor, a regulator, or an internal controls team may eventually need to trace exactly how that conclusion was reached. AI-supported decisions are clearly beneficial as they are faster to reach and higher in volume, but they must be under the supervisory domain of a human.

General-purpose frontier models increasingly include strong safety measures such as behavioral safeguards, data handling controls at the model level, but model-level safety and workflow-level governance solve different problems. A model can be exceptionally well-behaved and still sit inside a process that produces zero audit trail, because nobody built the traceability into the workflow around it.

Governance in procurement can’t be treated as a feature to bolt onto a capable model at a later date; it has to be architected in from the start. To achieve the required level of governance in Source-to-Pay processes every AI-influenced decision needs to be logged, every recommendation should be traceable to the exact policy or line of data that produced it and, most importantly, a human-in-the-loop must be positioned at the specific points in the process where judgment genuinely matters. 

The ERP Problem No Model Solves on Its Own

Integration remains a critical challenge for any model. A frontier model arrives with no inherent knowledge of a specific company’s ERP structure, approval hierarchy, or policy exceptions that have accumulated over years of operation. That context has to be built and maintained separately. This is where procurement leaders should be focusing their assessment, rather than on general reasoning or coding tests. The criteria that actually determine value in a procurement context are the ones that will ensure an AI investment is successful. 

What an Ideal Solution Would Actually Look Like

The reasoning power frontier models offer is genuinely useful for any analytical, judgment-heavy element of procurement work such as spend pattern analysis, sourcing recommendations, or risk assessment across complex supplier relationships. It is, however, important not to equate raw capability automatically with value production, especially in a high-volume transactional environment like Source-to-Pay.

An ideal solution should combine elements of frontier intelligence and procurement discipline. It should be able to apply frontier model reasoning capability selectively, at the moments where judgment genuinely adds value, while routing the high-frequency, rules-based majority of transactions through a governed, auditable, ERP-native architecture built specifically for procurement.

The models will keep getting smarter every few months. The organizations that benefit most won’t be the ones chasing whichever one currently leads the rankings, but the ones that built the procurement-native architecture capable of putting that intelligence to work responsibly, regardless of which model sits underneath it.