Intelligent procurement can be a formidable competition weapon. And it’s about so much more than efficiency and going faster. A truly intelligent system improves the quality of every decision by analyzing internal and external variables like supplier risk signals, market price benchmarks, demand forecasts, and historical category strategy recommendations. In doing so, it can provide a detailed model of how a complex business operates and, crucially, which factors drive its success. No other enterprise function can offer this level of profit margin oversight.
Enterprising procurement teams are gaining this oversight and competitive advantage by centralizing real-time monitoring and using AI agents as “chiefs of staff” to CEOs. And unlike in some enterprise areas, teams are reporting impressive ROI figures: sourcing cycle times cut by 90%; contributions to billions in savings that offset inflation and funded new investments; and savings of 15% on tail spend that materially boosted earnings per share.
Charting the wrong path, however, can cost procurement teams dearly. With such high potential gains, arguably, the biggest losses are those linked to the opportunity costs of taking too long to implement.
The pitfalls of LLMs alone
For companies that have already invested heavily in large language models (LLMs), it’s alluring to think procurement intelligence can be achieved using these alone. The vision couldn’t happen with a general purpose LLM alone.
That’s because LLMs most definitely have a role to play in procurement workflows. They can be used effectively to build discrete, point solutions for specific tasks. They provide excellent conversational interfaces that democratize complex tasks and reduce or even eliminate training costs. They are incredibly useful for reformatting unstructured RFPs and other documents.
However, discrete tools built with LLMs are typically not scalable, enterprise-grade solutions. As statistical engines designed to predict the next word, the output of LLMs varies considerably based on the input and context provided by the user.
To ensure AI makes the correct decisions, especially in situations that materially impact enterprise cost and risk, we must ensure it has the context and domain knowledge harvested from thousands of procurement transactions over many years. True intelligence reaches optimal decisions by making the right tradeoffs. It has to reason through complex requirements, identify missing information and resolve ambiguities based on category-specific sourcing history. It requires portfolio-level context of previous awards, expiring contracts, and cross-enterprise leverage.
The opportunity cost of building vs buying
Most teams seeking to gain procurement intelligence aren’t looking to rely solely on LLMs but instead use them to engineer home grown systems. Think about this decision, as the cost and impact can be significant. According to McKinsey’s 2025 State of AI Report, the average enterprise AI project takes 17 months to reach production. Gartner further found that at least 50% of generative AI projects are abandoned due to poor data quality or escalating costs.
The biggest cost of all, however, will be the opportunity cost associated with the delays in transforming. For a typical enterprise procurement organization responsible for billions in spend, every month spent building is a month of missed savings. As the figures previously mentioned suggest, these can easily run into the tens of millions, due not only to the huge efficiency gains, but the advantage of having that margin oversight to inform more profitable decisions.
Risky business
A chief reason for delays incurred when building systems with LLMs is risk. Procurement sits on the intersection between business stakeholders and suppliers, handling sensitive information and driving core business outcomes. The tools you use must be secure, reliable and effective. Simple solutions built in house often lack the required robustness to roll out across large stakeholder groups, external suppliers and handle sensitive data.
For a home-grown system to deliver appropriate governance and compliance, it would need to replicate the workflow layer that purpose-built platforms have spent years designing. For most IT teams proposing a build, that layer is not in scope. Furthermore, replicating the necessary compliance rules such as SOC 2, ISO 27001, and GDPR requires years of initial and continued investment.
A more intelligent approach: native, domain-specific AI
Evaluating any technology and approach should always start with defining the objectives and desired outcomes. Organizations requiring robust, enterprise-grade procurement intelligence to support a range of tasks are advised to seek solutions that draw on domain-specific AI models. These benefit from being pre-trained on thousands of sourcing events, deep category-knowledge and negotiation strategies.
In typical companies, where external spend levels can easily mount up to 50% of revenue, specialized models designed for procurement offer the highest returns, fastest, and with the lowest risk. In a recent report, Gartner described them as “a wish come true for unlocking AI value” that offered “up to 50% lower development costs, faster deployment and consistently higher reliability in business-critical workflows.”
Several other major studies have reached similar conclusions including one by Cambridge University that found that “specialized models had a 38% relative accuracy gain, ran at 1/80th the cost and consumed an estimated 1/200th the energy of the best-performing LLM.”
Domain-specific AI utilizes specialized agents for tasks like RFx drafting, compliance validation, and multi-turn negotiations operate within predefined enterprise constraints like spend thresholds and preferred supplier lists. Compared to the chatbot in my previous example, a domain specific AI collects supplier responses through structured portals to enforce "apples-to-apples" comparability and centralized audit trails.
Without question, LLMs play a role in procurement workflows. However, for companies striving to save materially and gain profit margin oversight from procurement intelligence, the fastest, most profitable, least risky route is to adopt domain-specific AI models.
