Transcript 0:00 Every software vendor selling you AI is selling you the same thing right now, a Formula One car for your procurement team. And very often that car is real. Some of these tools are genuinely impressive. 0:13 But an F1 car handed to a team with no pit crew, no track knowledge, and no race strategy does not win. It crashes into a wall, and we all know how expensive that is. 0:25 So today, I wanna talk to you about the pit crew, because whether agentic AI becomes a force multiplier or a very costly mistake has almost nothing to do with the car you're driving. Let's get into it. 0:37 [upbeat music] Welcome to ProcureTech Unpacked, the show for procurement professionals who want to master the technology shaping our function. 0:56 I'm Joel Conneen Demers. Quick definition first, because agentic gets thrown around quite loosely these days. Generative AI generates content. Predictive AI forecasts. Agentic AI executes with agency. 1:11 It perceives, decides, acts, and learns with minimal human intervention. 1:17 In procurement, that means functionality in your software that can evaluate quotes and award business, monitor a contract and trigger a renegotiation when a certain context exists, route a requisition through a dynamic approval flow based on live risk scoring and other factors, and this can all be quite powerful when the foundations are there. 1:39 And that's the core problem. 1:41 Agentic functionality needs five things to do any of what I just mentioned: business rules to follow, systems to access, processes to execute, feedback loops to improve, and finally, trust from its owner. 1:56 Take those things away and the agent can't generate much value for your business. You would never hire a junior buyer and say, "Hey, figure it out. 2:04 We have no documented policy, IT's locked down the ERP, and nobody can really explain how our sourcing process works." 2:12 But that is exactly what a lot of organizations are handing their AI agents and then expecting crazy returns. So let me make this more concrete. 2:23 Say you've deployed an AI agent to handle low volume, low complexity sourcing on its own. A request lands. It's for office furniture. Twenty-five thousand dollar budget, standard specs, needed in ninety days. 2:37 What should the agent do? 2:39 Send an RFQ to three pre-approved suppliers, negotiate directly with the incumbent on historical pricing, route it to a human because furniture comes with installation dependencies, or auto award to the lowest bid if the quote lands within five percent of your budget? 2:56 Well, if you can't answer w- that with a documented business rule, your agent can't answer it consistently either because agents run on one of two things: A, a hard rule that you define, or B, statistical inference from patterns in their training data. 3:13 If you haven't written the rules, then you are accepting the results of an unknown statistically driven output at scale. 3:21 And if you don't have these rules, it means that your human buyers are already making inconsistent calls based on their experience, preference, and objectives. The agent just forces you to see that. 3:33 The good news is that if you can work this furniture question, the path to readiness is shorter than it looks. 3:40 It comes down to five system design concepts, and these are the domains where autonomous procurement succeeds or fails, whether you buy a platform or build one. 3:51 Here they are as fast as I can, given this is a short minisode. Number one, policy as a rule book. 3:57 The agent's logic has to come from somewhere, and in procurement, that somewhere is your policy, your standards, your business rules, your category strategies, et cetera. 4:07 If your policies say strategic sourcing is required over fifty thousand dollars, but never defines what strategic sourcing is, the agent writes its own definition. 4:18 It will not get stucked, it will act, but you may not like that action. Number two, system ownership. Who owns what? 4:25 In general, I advise that IT keeps infrastructure security integrations with other systems, but that procurement has to do more of the work of owning the business rules, the approval routings, the category logic, and the fix when the agents get decisions wrong. 4:42 In the era of AI, everybody has to become a business analyst, and if every change to agent logic means you need to open up a ticket with IT and wait for three months, agents become a liability instead of, of an asset. 4:56 Number three, process design and exception handling. Most teams think their processes are documented. Then they go to automate it and find out that there's no standard sourcing process. 5:06 There's forty-seven variations, half of them contradict each other, they live in email chains and/or someone's memory. 5:14 Every fork in the process doesn't have to be explicitly written down, but you have to have a general sense of how you want the agent to act and react and what are the guiding principles. 5:25 Not it depends, because it depends means you're gonna rely on the model training data rather than what you want to have happen. Number four, configuration management and release discipline. 5:38 This is the procurement version of how IT and software teams ship code and system changes today. You test the change in a sandbox before it touches your live transactions. You track versions. 5:51 You ge- have the ability to roll back when something breaks and tell the humans who are interacting with agents in your production system what has changed from version to version so that you keep their trust high. 6:03 When the market moves and your, your negotiation logic from six months ago needs to change, [clears throat] 6:08 tariffs, you already have the muscle to update it safely if you have configuration management and release discipline. And then finally, number five, adoption and stakeholder trust, which, which I just alluded to. 6:22 You can buy the best system on the market, but if people in your organization do not trust it, they will route around it, agent or no agent. 6:29 Buyers override, requesters email procurement directly, finance demands the approvals back. 6:35 Adoption is not a technology or a rollout problem, it's a trust problem, and trust comes from consistent outcomes and decisions that people can actually see explained and actually agree with, not from a launch email. 6:49 And so agentic AI is only as good as the operating model behind it. 6:53 If your policy is vague, if your process logic lives in people's heads, and if procurement can't govern how the systems behave over time, autonomy and agentic AI won't give you more control. 7:05 It will scale inconsistency faster. Getting the rules, the ownership, the exception paths, and the ability to tune your systems as conditions change are the keys to making agents become a force multiplier. 7:20 Getting it wrong just automates confusion, so you're always better to start smaller and make sure you have those muscles on-- in a small context versus trying to go agentic AI end to end. 7:31 The teams winning with this aren't the ones with the flashiest tools. 7:35 They're the ones who do the boring operating model work in parallel to the technology work, and also those who partner with vendors who don't just sell them software, but help them level up their practices on these five foundations. 7:48 So before you sign an agentic AI contract with anybody or hand your team a mandate to build something, run the five questions, run the maturity assessment. 7:58 The answers will tell you more about your readiness and your vendor's readiness than any product demo can and/or proof of concept really. 8:07 I wrote the full playbook on this topic, complete with maturity model in the Pure Procurement Newsletter, and if you'd like a deeper dive into these five foundations and the things you need to master to get business results with agentic AI, the link is gonna be in the show notes. 8:22 See you next time. That's a wrap on this episode of ProcureTech Unpacked. 8:27 If this one resonated, subscribe wherever you get your podcasts and sign up for the Pure Procurement Newsletter for weekly insights between episodes. 8:35 If something we covered sparked a question or an idea, we'd love to hear from you. All the links for the reports we discussed are in the show notes. We'll see you next time. [outro jingle] Resonate.
