Adapted from a talk I deliver at procurement conferences. Details on booking at the end. Also recorded as a podcast on the Public Procurement Change Agents podcast.
AI meets public procurement in three places at once. It is a product we scope, buy, and oversee. It is an enabler that compresses time-to-task across procedural work. And it is an accelerator that makes staff more productive as a teammate in complex work.
Given that, I want to lay out three current topics: foundational competencies that let you describe and communicate about AI, fluency you have to build yourself, and the next frontier of where this technology is rapidly evolving.
Foundations
Conversational AI has evolved substantially since it reached the general public in late 2022, when large language models were trained through an elaborate game of “guess the next word”. Today's systems routinely employ reasoning modules, curate data sources, run web searches, and carry real analytical capacity.
Adoption inside your own office varies much more than most leaders realize. Walk your shop, and you will find five people. One has never needed AI and does not want it. One has a specific objection and stopped. One forgets the tools exist and lacks the skill to do much beyond the basics. One uses it daily and considers it indispensable. And one is using tools far beyond what has been provisioned, and is fairly confident you cannot stop them.
That last person is not hypothetical. A PagerDuty survey published in June 2026 found two-thirds of office professionals had used AI at work despite believing it violated policy, and more than a third had entered customer data into public models. Verizon's 2026 Data Breach Investigations Report found that shadow AI detections rose fourfold in a single year. Unauthorized use drops sharply when organizations provide sanctioned tools that match what employees have already found on their own. Prohibition without provision does not produce compliance. It produces invisibility.
Most organizations that stall on AI stall for one reason: they treat it as conventional automation, where you specify a process, configure a system, and measure throughput. This technology does not behave that way. People work with it through dialogue and correction, the way they work with a colleague.
By establishing a thoughtful acceptable use policy, baseline technology provisioning, market-aligned standards for how AI itself gets procured and monitored, investment in staff training and upskilling, and a plan that encourages testing and learning - organizations build the foundation they need.
Fluency
Policies and platforms alone do not produce capable users. Practice does.
Most people move through four stages. Explorers are new and often skeptical; they need exposure and low-stakes repetition. Adopters use it regularly but struggle with consistency and cannot explain why results vary; they need technique. Amplifiers customize workflows, build reusable templates, and mentor colleagues. Builders develop and manage AI systems at the organizational level.
Most procurement offices sit heavily in the first two stages. The highest-leverage move available is pushing a meaningful number of people from Adopter to Amplifier, because Amplifiers create the templates and habits that lift everyone around them.
The evidence that personal skill matters is strong. A field experiment published in Organization Science assigned 758 consultants at a global management consulting firm to work without AI, others with GPT-4, and others with GPT-4 plus a brief prompt-engineering overview. Those using AI completed 12.2 percent more tasks, finished roughly 25 percent faster, and produced work rated more than 30 percent higher in quality. The group that received even a brief training overview performed even better.
The Next Frontier
Late 2022-2023 AI use was basically chatbots – smart chatbots, but chatbots nonetheless. A chatbot is a prompt and a completion. You ask, it answers, and you take the work.
Agentic AI does not just answer. It plans, executes, and iterates. Give it a goal, such as evaluating vendor responses for compliance and risk. It builds a plan, breaking that into compliance checking, then scoring, then risk flagging. It acts, pulling documents and comparing across responses. Then it reflects, adjusting where it found gaps, etc. Three capabilities make this work: memory that persists across steps, tool use that lets it open documents and query systems, and conditional logic so it can branch on what it finds.
The procurement translation matters. The old request was "summarize these proposals." The new request is "run a structured evaluation against our policy and scoring criteria, flag anything non-conforming, and show your work."
Orchestration is the next layer, where narrow agents each carry a role: for example, an intake agent clarifies requirements, a compliance agent validates against statute and template, an evaluation agent scores and flags risk, and then a unifying orchestration agent can compare the outputs and send individual agents off to redo tasks as an example. A primary example of this is in coding, where it is regularly used to run agents on different tasks, with an orchestration layer that unifies the outputs.
To give a real-world example from our own practice, we have built an on-demand help desk called askCivic that provides customized self-service support while simultaneously helping staff build skills in AI, and human oversight from Civic staff embedded throughout.
In standard mode, users open independent threads on a topic and work through normal prompt and completion loops to solve problems in their work. The tool cites its sources and operates within the constraints of curated knowledge bases.
In agentic mode, when a question falls outside those knowledge bases, the tool opens a ticket, routes it to a person at Civic, captures the answer, determines the best method to resolve it, and updates the knowledge base for future users. The end user never has to know to ask for any of that. This is an example of a single agent taking sequential action. For an orchestration example, we use a variety of agents in support of coding, testing, and knowledge base curation, where agents act independently, they are unified at an orchestration layer, and guided by our lead developer.
Closing
Procurement is largely knowledge work. It is reading, comparing, judging, documenting, and defending, and a large number of those tasks are ones this technology can clearly accelerate.
We have found that coming at this space from a mindset that if new tools get you to a better starting point in a fraction of the time, take the time back and reinvest it in judgment and strategy, which is the part no AI tool will ever do for you.
This article is adapted from "AI-Powered Procurement: Foundations, Fluency, and the Next Frontier," a session I present for procurement conferences, association chapters, and agency training days. The full talk includes live examples, the eight-organizational-fundamentals checklist, and a working agentic procurement demonstration. If your organization is planning a conference or professional development session, reach out.
