In direct materials, agentic sourcing pays off only when every agent works from one canonical model linking market prices, contract pricing mechanics, volumes and the standard cost and PPV cycle. Building and maintaining that model is harder than most manufacturers expect, which makes buying it the better economic choice for most of them.McKinsey's latest procurement piece, "Agentic AI in strategic sourcing: Hype or hero?" by Luca D'Avino, Mauro Erriquez and Pradhuman Aggarwal, published October 28, starts with an uncomfortable statistic. In a survey of 200 CPOs and senior procurement leaders, 80% plan to increase technology and AI investment in the next 12 to 18 months. Only 2% report AI embedded at scale, and 55% remain in pilot mode.The authors argue the problem lies in where companies point the technology. Automating requisitions and purchase orders acts on the smallest cost base procurement touches. Sourcing decisions act on external spend, which McKinsey puts at 30% to 60% of operating expenses. On a blended portfolio, they estimate better sourcing decisions deliver roughly 11 times the value of automating the procurement function, with 3% to 6% savings on external spend when agentic AI is deployed well.For a food, beverage or consumer goods manufacturer, the relevant figure sits closer to the top of that range, and in a more sensitive place. Raw materials and packaging land in cost of goods sold, and from there directly in gross margin. Take an illustrative manufacturer with $10 billion in revenue and $4 billion of raw material and packaging spend. One percentage point of better outcomes on that spend is $40 million of gross margin, before any effect on pricing power, working capital or continuity of supply. Indirect savings often get absorbed into budgets and are hard to trace. Direct materials savings show up in the margin bridge, and analysts question them if they fail to appear.The same sensitivity makes direct materials the hardest place to get agentic sourcing right.The most important idea in the McKinsey article gets less attention than the savings figures. The authors write that quality commercial decisions depend less on any single model than on whether every agent sees the same commercial picture and the same ontology, meaning a common definition of the entities, relationships and concepts that matter. They define the minimum context through four questions: what changed, what is affected, what the organization has learned before, and which capabilities apply now.In indirect categories, a reasonable answer to "what is affected" comes from a supplier list and a contract repository. In direct materials, the answer requires a model few companies hold in one place.Consider what an agent needs to know when PET resin prices move. It needs to know which resin grades sit in which bottle and preform specifications, and which SKUs and plants use them. It needs the pricing mechanism in each supply agreement, which is rarely a single price. A typical packaging contract prices off a published resin index with a defined lag and averaging period, adds a conversion fee, and sometimes includes volume bands, currency clauses or caps. Aluminum can contracts often combine an exchange price, a regional premium and a conversion charge, each moving on a different schedule. Agricultural inputs mix spot purchases, fixed-price forwards, basis contracts priced against futures, and index-linked agreements, often for the same commodity across different regions.Then come volumes. Committed volumes, forecast volumes from sales and operations planning, take-or-pay obligations and inventory positions all determine how much of a market move actually reaches the P&L, and when. An agent unable to tell a fixed-price forward from an index-linked contract with a three-month lag will produce confident and wrong exposure estimates.This is the canonical layer McKinsey describes, adapted to raw materials and packaging. Units of measure, grades, origins, Incoterms, currencies, index mappings, pricing formulas and contract volumes all need to resolve to one definition. Without it, each agent builds its own picture, and the orchestration McKinsey describes coordinates disagreement.Direct materials teams answer to a second system the article mentions only in passing: the standard cost and purchase price variance cycle. Most manufacturers set standard costs for raw materials and packaging once a year, often months before the year starts. Procurement performance then gets measured as PPV against those standards, and finance builds its gross margin forecast on top.The link matters for agentic sourcing in two ways. First, a sourcing decision has a different value depending on when its cost lands relative to the standard. Locking in cover below standard for the second half produces favorable PPV in that period. The same decision taken after standards reset produces none. Second, the market path over the coming months determines next year's standards, and therefore the margin plan the CFO commits to the board.A decision layer that sees market prices, contract mechanics and volumes but not standards and PPV will optimize for the wrong number. A layer that links all four lets a category manager see the consequence of a coverage decision as finance will report it, this quarter and in next year's plan. It also lets the team simulate a scenario, such as a 15% move in a key input, and show its path through spot purchases, index-linked contracts, fixed cover and PPV, by SKU and by quarter. At that point simulation becomes useful to a CFO rather than interesting to a buyer.McKinsey is candid about cost. For a medium-sized organization, the authors write, a first production workflow can run into the millions of dollars once data foundation, orchestration, controls and change management are included. They also note procurement functions often use less than 20% of the data available to them, and data quality was the most cited barrier to AI returns in their survey.For direct materials, the data problem extends well past internal systems. A useful context layer combines ERP and contract data, bills of materials, specifications, planning volumes and treasury hedge positions with external inputs, including exchange prices, weather, freight, energy and trade policy. It also needs proprietary content from price reporting agencies, market analysts and brokers, often the most valuable signal a category team has. Each of those sources arrives in its own format, on its own schedule, under its own licence terms. McKinsey flags the governance question explicitly: which supplier data a contract permits into memory. The same question applies to licensed market content.Building the layer once is achievable. Maintaining it is the harder commitment. Index methodologies change, contracts renew with new formulas, specifications get revised, agencies alter their feeds and plants change suppliers. A context layer drifting out of date by a few months produces exposure numbers nobody trusts, and the agents built on top lose credibility with the category team.The article's most interesting section concerns learning. The authors argue system value compounds only when outcomes change how the next workflow is assembled. Every decision leaves a trace, including objective, evidence, capabilities used, interventions and outcome. Evaluation then feeds back into which signals proved predictive, which models work in which market structures, and which sequences become reusable. Their example: a should-cost model holding up in engineered components may fail in capacity-constrained markets.They also note the write side needs about a year of traced decisions before it has anything to learn from. This point carries a buy-versus-build implication the authors leave open. A manufacturer building in-house starts its learning clock at zero, on its own decisions alone, and pays for the full ontology, connectors and controls before the first useful result. A specialist platform arrives with the direct materials ontology, pricing mechanisms, index mappings and source connectors already built and maintained. Its customers carry the maintenance cost collectively.The skeptic's case deserves a hearing. Buying creates dependency, and decision records hold sensitive commercial information. Any platform worth buying should leave the decision history and the customer's data under the customer's control, and should never pool one client's contract terms into another client's memory. Some very large manufacturers with deep data engineering teams and unusual category structures will reasonably choose to build. For most, the economics run the other way. The scarce resource is category judgment, and the expensive mistake is spending two years building plumbing before the first decision improves.McKinsey recommends starting with one high-value workflow, replaying it against the last 12 months of completed events, running it in shadow alongside the live team, and only then moving to human-approved live decisions. For a direct materials team, a sensible first workflow is a single volatile input with real spend, such as a packaging resin, a dairy input or cocoa. Map every pricing mechanism touching that input, link the volumes and the standards, and test whether the system's view of exposure matches what finance reported in PPV over the past year.The replay test answers a question most AI pilots skip. If the decision layer cannot reconstruct last year's PPV from last year's market moves, contracts and volumes, its forward simulations deserve no trust. If it can, the team holds something rarer than another dashboard: one model of the category in which procurement, finance and planning agree on what a market move means before the move reaches the P&L.McKinsey asks whether agentic AI in sourcing will be hype or hero. In direct materials the answer depends less on the agents than on whether the model underneath them is complete enough to deserve their speed.