AI is going to change procurement. The harder question for many middle-market companies is where it will actually work better than another increasingly attractive alternative: lower-cost, highly capable sourcing support in places like the Philippines.
Artificial Intelligence versus Asian Intelligence
The comparison is particularly relevant for tactical purchasing work. Buyers spend a surprising amount of time sending purchase orders, obtaining acknowledgements, chasing promise dates, expediting late orders, confirming shipments and correcting information in the ERP. On paper, these look like ideal automation opportunities. They are repetitive, rules-based and consume time that buyers could be spending negotiating, developing suppliers, managing risk or finding savings.
In practice, they are often much harder to automate than they appear. The problem is rarely generating the purchase order. The problem is everything that happens after it leaves the building. One supplier confirms through a portal. Another replies to an email. Another sends a PDF. Another changes the delivery date without clearly identifying what changed. Another doesn’t respond at all until somebody calls them.
Large companies can often dictate the process. If a very large customer tells a supplier to use a particular portal, EDI connection or acknowledgement format, the supplier is likely to comply. A $100 million manufacturer buying $75,000 a year from that same supplier has considerably less leverage.
That matters because many procurement automation projects quietly assume suppliers will change the way they work. The technology may work perfectly, but if hundreds of suppliers have to maintain another password, learn another customer’s system and consistently enter information into it, adoption becomes the real problem. Recent studies of digital supply-chain initiatives continue to identify data quality, integration and adoption as major obstacles to getting the expected value from AI and automation.
There is also an important reason why an alternative model has become much easier to implement. COVID effectively taught corporate America how to work remotely. Companies that once believed purchasing, finance and other support functions had to sit together in an office spent several years running them through Teams, Zoom, shared systems and digital workflows. Managers learned how to train and supervise distributed employees and increasingly measure performance based on output rather than physical presence.
That groundwork also makes it much easier to integrate a sourcing support team in the Philippines than it would have been a decade ago. From the perspective of a U.S. buyer already communicating with colleagues and suppliers electronically, the person maintaining promise dates doesn’t necessarily need to sit in the next cubicle—or even in the same country.
The economics can be compelling. Even after benefits, management, technology and third-party staffing costs, Philippines-based support can cost a small fraction of a domestic buyer. That makes it possible to apply significantly more resources to work that often isn’t being done consistently today. A Philippines-based team can make sure a PO was received, obtain the acknowledgement, follow up when there is no response, clarify an ambiguous promise date, expedite a late order, confirm that a shipment actually left and update the ERP. Instead of asking every supplier to conform to a new automated process, the support team adapts to whatever process the supplier already uses.
The opportunity isn’t simply labor arbitrage. It is separating work that requires strategic judgment from work that requires persistence, attention to detail and follow-through. A domestic buyer or category manager can add considerable value negotiating a difficult price increase, developing a sourcing strategy, managing a supplier transition or resolving a quality problem. There is considerably less reason for that same person to spend twenty minutes trying to determine whether Thursday’s shipment actually left the supplier.
Done well, the result is not just lower cost. More POs are acknowledged, promise dates are more accurate, late shipments are identified earlier, supplier follow-up is more consistent and the ERP begins to contain information that buyers and planners can actually trust.
Artificial intelligence has a very different cost equation. At first glance, AI can look almost free. Current leading models charge only a few dollars per million input tokens and roughly $10–$20 per million output tokens at commonly used model tiers. For a simple task—reading an article, categorizing a transaction or summarizing an email—the direct model cost can be pennies.
But token cost is not the same thing as automation cost. An agent that autonomously monitors an inbox, reads a PO, recognizes a supplier’s response, determines whether the acknowledgement matches the order, updates the ERP, follows up on an exception and knows when to escalate requires considerably more than access to a language model. It has to be developed, integrated with email and ERP systems, tested against hundreds of real-world exceptions, secured, monitored and maintained as systems and processes change.
Agentic processes also consume more tokens than a simple question-and-answer interaction. The agent may repeatedly send context to the model, call tools, examine the result, reconsider its next step and retry when something fails. Search, computer-use and other tool calls can add additional charges as well. The cost per individual action may still be small, but across tens of thousands of POs, emails and agent steps, usage becomes something that needs to be designed and managed rather than assumed to be free. Current AI pricing reflects both token consumption and, for some capabilities, separate tool-use charges. More importantly for many middle-market companies, the development and integration cost may matter more than the tokens. Spending significant money to automate a process that could be handled inexpensively by a small offshore team may produce a poor return, particularly if supplier behavior creates a long tail of exceptions that still require human intervention.
That does not mean artificial intelligence loses the comparison. It means using it where its economics and capabilities are strongest.
One of the best examples is market intelligence. A sourcing professional cannot continuously follow hundreds of commodity indices, tariff changes, supplier announcements, regulatory developments, freight markets, currencies, geopolitical events and industry publications. AI can do that with very little incremental labor. A useful system could monitor steel, copper, resin, oil, currencies and other important inputs and tell a category manager when something has changed enough to matter. It might flag a significant commodity move, identify a new tariff affecting a sourced product, summarize a supplier announcement, warn about a disruption in an important manufacturing region or identify a material change in the economics of sourcing from one country versus another.
The goal isn’t another daily newsletter containing 50 articles. It is an alert that says: Copper has increased 9% in six weeks. It represents approximately 22% of the cost of these five categories, and $4.3 million of annual spend is potentially exposed. AI can also be particularly effective in spend classification, supplier identification, commodity mapping, contract review, price-variance analysis and first-pass market research. In these applications, the problem is primarily one of scale: reading more information, making more comparisons and identifying patterns faster than people realistically can.
That is very different from getting a supplier who ignored three emails to confirm when an order will ship.
The distinction suggests that the best operating model for many middle-market sourcing organizations isn’t choosing one version of AI over the other. It is understanding the economics and using each where it has an advantage. For a transactional process with messy supplier behavior, Philippines-based support may cost less and work better than spending heavily to develop a sophisticated agent. For monitoring thousands of news articles, commodity movements and regulatory announcements, paying a person to perform the work manually makes little sense when AI can screen vastly more information for a very low incremental cost.
Increasingly, the two can work together. AI might read an emailed acknowledgement and identify that the supplier changed the promised date. The Philippines support team can determine whether the new date is acceptable, contact the supplier if it isn’t, update the system and escalate the issue when necessary. AI can identify which orders appear to be at risk; a person can figure out what is really happening and get the supplier to do something about it.
There is another advantage to this progression. A disciplined support team can create the clean processes and reliable data that make greater automation possible later. Once acknowledgements are consistently captured, promise dates are maintained and exceptions are categorized, a company can see which activities actually follow predictable rules and which consistently require human intervention.
Over time, more of those predictable activities can migrate to automation as the technology improves and the economics make sense. The people don’t prevent automation—they can help create the conditions necessary for it to work.
For middle-market companies, the question isn’t whether people or AI win. It is where a dollar spent on each produces the greatest return. In many sourcing organizations today, the answer is likely to be both: use lower-cost human intelligence to handle the messy transactional work that still requires persistence and judgment, and artificial intelligence to process the information volumes that people could never reasonably keep up with.
