________________________________________________________________________________________________________________________
It’ s easy to conflate model-level safety with process-level governance, but they aren’ t the same thing. Today’ s frontier models have safeguards such as behavioral guardrails and data-handling protections built in, but a perfectly well-behaved model can still be slotted into a workflow that produces no audit trail at all, simply because nobody built traceability into the process around it. The EU AI Act makes this concrete: its obligations around high-risk AI systems, such as automatic record-keeping, traceability of outputs and demonstrable human oversight, describe an architecture, not a model configuration, and the compliance burden lands on the organization deploying the system, not the model vendor. Governance of this kind needs to be designed in from the outset.
Traceability is only half of the problem, though. The other half is integration, since no model, however capable, arrives already knowing your business. It doesn’ t know your specific ERP configuration, your approval hierarchy, or years of accumulated policy exceptions unique to your organization. Building and maintaining that context is a separate undertaking altogether, and it’ s where supply chain leaders should be focusing their evaluation efforts.
Why frontier models don’ t fit routine procurement
None of that shows up on a leaderboard, which is why the industry’ s obsession with frontier benchmarks looks misplaced. Every time a new frontier AI model launches, the web is flooded with benchmark comparisons and demo reels, and a familiar debate reignites about whether machines have reached some new height of reasoning ability. Very little of it speaks to whether the model fits the way supply chain teams work day-to-day.
Frontier reasoning models have been designed with software engineering projects,
open-ended research, and multi-step planning in mind, work that requires a model to explore dead ends, backtrack, and rethink its approach across long sessions – and the benchmarks that rank them measure precisely that. Most day-to-day supply chain operations, by contrast, are largely repetitive and rule bound. Reconciling purchase orders against invoices and receipts, checking new suppliers against a fixed compliance checklist, or pushing approvals through the same delegation chain every time does not need freewheeling problem-solving. What Sourceto-Pay needs is consistency: reliable output, high throughput, and costs that don’ t swing wildly from one transaction to the next.
Applying reasoning-heavy models to routine transactional work also carries a direct cost. Claude Fable 5, for example, runs at $ 50 per million output tokens, exactly double Claude Opus 4.8’ s $ 25 standard rate. At enterprise transaction volumes, a model that reasons through a routine task
16