Procurement at AI Speed — A Podcast on AI in Procurement and Commercial Judgement

Imagine five professionals evaluating bids for a major contract. Before anyone records their score, they read the same AI-generated summary.

It makes a convincing case for one supplier. By the time the panel meets, their scores are remarkably similar.

That could be good news, or it could mean something else. Did five people independently reach the same decision? Or did one AI interpretation shape five opinions?

This is an imagined scenario, but the question matters anywhere AI supports important decisions.

You approved it, but did you decide? In my first episode, I argued that as AI takes over more analytical work, human value would increasingly lie in questioning assumptions and exercising judgement.

I still think that's important, but new research made me question something I had taken for granted.

What if AI influences our judgement before we even realise there is a decision to make?

On the 27th of October, Anthropic published a study involving 50 professional analysts using an AI agent.

The analysts were helping a fictional farming cooperative decide whether to expand or discontinue an advisory program for farmers. To make that decision, they first had to choose how to measure success.

Should they look at how much farmers produce from their land or how much they produce for every dollar spent?

Both approaches were reasonable, but they could lead to different conclusions. But here is what the analysts didn't know. The researchers had randomly instructed AI to favour one of the options. It was told to make the strongest case the data allowed without misrepresenting the results.

Across four analytical choices, the analysts followed the AI's preferred option 86% of the time.

If those randomly assigned preferences had no influence, you would expect agreement closer to 50%. Now, there is an important limitation.

This was a fictional task using one AI model, and the researchers deliberately strengthened its preferences.

So the 86% doesn't tell us how often this happens in everyday work.

And there was another interesting finding from this same experiment. The researchers also looked at 38 instances where analysts had already expressed their preferred option before receiving the AI's recommendation.

When the AI disagreed, they changed their minds 60% of the time. But when AI agreed, they changed their minds just 6% of the time.

This was a small number of cases, so we need to be careful about drawing broad conclusions. But it raises an important question. Were they changing their minds because the AI presented better evidence? Or because its arguments sounded more convincing?

And think about why this is hard to detect. If AI gets a formula wrong, we can check the numbers. If it misreads a contract clause, we can go back to the document.

But sometimes, the issue isn't a factual error. It's the judgement behind the recommendation. Which measure matters most, which risk deserves more weight, which supplier should make the shortlist.

Each option might be reasonable, but the way AI presents those choices can influence the final decision.

The calculations could be correct, but what if no one questioned assumptions behind them?

There is a useful example of this in procurement. When evaluating tenders, panel members can first review and score the bids independently before coming together to compare their assessments and agree on final scores.

This is also the approach outlined in South Australia's Government Procurement Guidelines and there is a good reason for it. But why does that sequence matter? Because each person brings a different perspective.

An engineer might spot a technical weakness. Finance might question a cost assumption. And procurement might identify a commercial risk in the supplier's proposal.

That's the real value of having an evaluation panel. It's not just about reaching agreement. It's about challenging each other's assumptions before making a decision. Now imagine what happens if everyone reads the same AI-generated recommendation before they have formed their own opinion.

They might all arrive at the same conclusion, but is that because they independently assessed the evidence, or because the AI influenced how they looked at it?

And here is the important distinction. The AI doesn't necessarily have to be wrong for this to become a problem.

The risk is that we lose the independent perspectives the panel was supposed to bring together, and this goes well beyond procurement.

Think about a CFO reviewing an AI-prepared investment proposal, or a risk executive assessing AI-recommended controls, or a supply chain leader deciding whether to change inventory levels based on an AI forecast.

In each case, AI can provide valuable analysis, but someone still needs to challenge the assumptions, weigh the trade-offs, and take responsibility for the decision.

Let me show you how the problem could play out. Let's imagine a manufacturer that buys 100,000 electronic sensors a year from a supplier paying $50 per sensor.

An AI sourcing assistant identifies another supplier offering compatible sensors at $47 each and recommends shifting 40% of the orders across.

On price alone that looks like a saving of $120,000 a year, with the added benefit of relying less on a single supplier.

Sounds like a good deal, right? But here is what the initial recommendation might have overlooked.

The existing contract includes a price increase if the manufacturer reduces its orders below 80,000 units a year. The new supplier also has longer delivery times, which means holding more stock, add extra freight and potential quality costs, and the economics start to change. There is also a one-off cost to qualify the new supplier and make sure the sensors meet the required standards.

In this example, once we account for the price increase on the remaining orders along with additional freight, inventory and quality costs, that apparent savings disappears. Instead, the business would face around $44,000 in additional annual costs.

And that's before considering the one-off cost of qualifying the new supplier or the potential benefits of reducing supply disruption.

So does that mean the AI's recommendation was wrong? Not necessarily. Having a second supplier could help the manufacturer keep production running if the existing supplier experiences a major disruption and that protection might be worth paying for. Or perhaps moving just 20% of the orders would offer a better balance between cost and resilience without triggering the existing supplier's price increase.

But a smaller allocation might also provide less protection. And that's where the real decision lies.

How likely is a disruption? How much would a second supplier reduce its impact?

And how much should the business be willing to pay for that protection?

The human category manager might have overlooked something too. Neither the human nor the AI is automatically right. What matters is making the assumptions visible, challenging them and understanding the trade-offs before making a decision.

So how do we do that in practice? Here is one approach I would suggest organisations test when making important decisions that require judgement.

I call it View Before Review. First, before reading the AI's recommendation, record a short initial view. What would you choose? What's your main assumption? And what evidence would change your mind? Second, ask the AI for its analysis, the assumption behind its recommendation and the strongest reasonable case for a different option.

Not just a polished recommendation. Third, compare the positions. Where do they differ?

Which facts can be checked and which assumptions need a decision-maker's judgement? And fourth, record the final choice, why you made it, and how you will review the outcome.

That final step is essential. Remember the Anthropic study? Even people who had already expressed a preference sometimes changed their minds when the AI disagreed.

A first view isn't a shield. You still have to ask yourself, what changed my mind?

Was it new evidence? A better argument? Or simply the confidence of the system presenting it?

And another study explored exactly that. Does it make a difference if we form our own opinion before seeing what AI recommends?

In 2026, researchers asked trainee teachers to assess classroom situations with the help of AI.

But sometimes, the AI's recommendations were deliberately misleading. And here is what they found. When the teachers formed their own opinion first, they reached the answer considered correct by experts about 54% of the time.

But when they saw the AI's recommendation first, that dropped to 37%. Interestingly, when the AI's recommendation was correct rather than misleading, both groups performed at about the same level. The researchers couldn't find a clear difference between them. Now, these were simulated classroom situations, not real business decisions, so we can't assume the same results would apply everywhere. But it suggests something worth considering. Forming our own view first may help us recognise misleading advice while still allowing us to benefit from AI when it gets things right.

Of course, I wouldn't use this approach for every email summary, spreadsheet check, or routine decision, and I certainly wouldn't delay an urgent reversible action just to complete another form.

The point isn't to slow everything down, it's to make a space for independent thinking when the decision really matters.

But there is one challenge we need to take seriously. What if the AI is actually better than us? And sometimes it might be. Human judgement comes with its own biases. We can favour familiar suppliers, be overly confident in our experience, or hold on to forecasts we've already made. And for some decisions, like forecasting demand or predicting risks, AI may actually make more accurate predictions than we do.

So the answer isn't to defend human judgement at all costs. It's to understand when AI improves our decisions, when it doesn't, and why. For finance, that might mean checking whether an AI-assisted investment forecast proved more accurate than the project sponsor's original assumptions. For risk, it could mean reviewing why a risk was accepted, and whether that decision still makes sense as circumstances change.

And for supply chain, it might mean testing whether human adjustments to AI forecasts improved accuracy or made it worse. Different functions, but the same underlying question: are we genuinely making better decisions? Or simply making them easier to approve? Because approval is a process step. Independent challenge is a capability. And they're not the same thing.

Challenging AI doesn't mean refusing to change your mind. It means understanding what changed it, and why. Now, think back to those five tender evaluators. Imagine they each assess the bids independently before seeing the AI's recommendation.

Then they compare their assessments with the AI's analysis, challenge the differences and revise their scores where the evidence supports it.

They might still recommend the same supplier, but this time, they have tested the reasoning rather than simply agreed with a recommendation.

It doesn't guarantee a better outcome, but it makes the decision easier to explain, review and learn from. So the next time someone tells you an AI-enabled process is safe because there is a human in the loop, ask yourself, did that person genuinely challenge the recommendation or simply approve it? Because approving a decision and genuinely deciding aren't always the same thing.

I'm Zina and this is Procurement at AI Speed, by KaiZina.

Some of the AI in your business didn't arrive through a sourcing decision. It arrived through a renewal.

From the first of July, Microsoft increased prices on several Microsoft 365 commercial plans, with existing customers moving to the new pricing when they renew. The update also included Copilot Chat enhancements.

Now, Copilot Chat isn't the same as the full Microsoft 365 Copilot, which requires a separate paid licence.

And Microsoft isn't alone. Last year, Google brought Gemini into its Workspace plans and discontinued the separate Gemini add-ons.

But here is the point. A business could carefully review its software renewal without ever stopping to assess the AI capabilities that come with it.

Are we choosing our AI suppliers? Or are existing software contracts increasingly shaping those choices?

That isn't necessarily a problem. Your incumbent supplier may be the right option. The platform is already integrated, your teams know it, and some security and governance arrangements may already be established.

But convenience isn't the same as a commercial assessment.

Let me give you an example. This is hypothetical. An Australian insurer renews Microsoft 365 with Copilot Chat available to its employees.

Meanwhile, the claims team is paying for a standalone AI assistant. And the customer service team's CRM supplier is proposing an upgrade with new agent capabilities.

Three reasonable decisions. But has anyone looked across them?

First, there may be capability overlap. Does the AI functionality already available meet the claims team's needs? Or does the standalone product offer something genuinely different?

If the standalone product adds real value, there may be a good reason to keep both.

But it's not just about overlapping capabilities. There is also supplier dependency. Which business processes are becoming reliant on AI from existing providers? And could different suppliers be relying on the same underlying AI models?

Understanding both puts you in a stronger negotiating position.

And before renewing, I would want two things clearly documented.

One, which AI features are included? Which ones cost extra? And how does usage affect what you pay?

The included version can become the pathway to a paid upgrade. And once teams depend on it, your negotiating position may weaken. So use the renewal to negotiate future upgrade pricing and flexibility before that dependency builds.

Two, what happens if the supplier changes the AI features, increases the price or changes how your data is handled?

You want clear notice, a say in material changes, and the option to walk away from the affected AI service if those changes don't work for your business.

For Australian financial institutions, this is already a regulatory concern.

In April, APRA, Australia's prudential regulator, called for stronger oversight of AI risk across banks, insurers and superannuation trustees.

It expects the entities it regulates to keep an inventory of AI tools and use cases, understand the AI supply chain, and have contracts that provide sufficient transparency, auditability and assurance over AI services.

APRA also observed that AI is increasingly embedded in software platforms, making some underlying dependencies harder to see.

So, an AI inventory that only captures what you bought through a separate sourcing process will miss some of it.

I don't think we need another procurement team or another layer of approvals.

What we need is to make AI part of the software renewal conversations we are already having, and make sure procurement, technology, finance and risk are all looking at the same picture.

Because the commercial decision doesn't disappear when AI arrives inside a contract you already have.

So, when your next software renewal includes AI, will you treat it as a feature or as a commercial decision?

I'm Zina, and this is Procurement at AI Speed, by KaiZina.

On the eighth of October, Reuters reported that AT&T is bringing more routine legal work in-house, using AI. Things like document review, research and drafting. Work that used to be billed by the hour at outside law firms.

The change wasn't simply a lower rate. Some of the work moved. And I think that changes the procurement question.

When a supplier starts using AI, it's tempting to ask: how much did you save, and how much of that should come back to us?

But you may never know their real saving. And depending on the deal, you may not be entitled to it anyway. So I'd start somewhere else.

You can't count on their saving. But you can test their price.

In a renewal, I care less about proving what the supplier saved. I care more about what the work is worth now, and how credible my alternatives really are.

Three questions help.

First: under the deal we signed, who keeps the efficiency gain? If we pay for time, less effort should mean fewer hours on the bill. If we pay a fixed price, or a price per unit, the supplier may keep the gain, unless the contract says otherwise. That distinction matters before the negotiation even starts.

Second: which parts of this service would we buy differently today? This is the question I think matters most.

Imagine a service desk with routine ticket handling and specialist escalations. AI might materially change the economics of the routine work, while the specialist work still depends on scarce human judgement. You may never split those into separate contracts. But you can price that routine work: in-house, from another provider, or under a different delivery model. And once you know that number, you negotiate the whole deal differently.

Your leverage isn't what your supplier saved. It's what you could do instead.

Third: when is our next real chance to reprice, and what needs to be ready by then? Maybe it's renewal. Maybe the contract already lets you benchmark the price, or ask for a commercial review. Or maybe you just need a credible alternative, priced and ready. That work starts before renewal, not when the supplier sends the proposal.

For longer-term contracts, I'd consider building in a price review if AI materially changes how the service is delivered. Not an automatic price cut. A reason to reopen the commercial discussion.

And there's one important caution. In August, Reuters reported that AI is already changing IT services contracts as clients push for more productivity and lower prices. But some providers are also avoiding deals built on aggressive assumptions about future productivity. That matters.

A low price that only works if unproven AI savings appear later isn't necessarily value. It may just move the commercial problem into delivery.

So at your next renewal, don't start by trying to prove how much AI saved the supplier. AI may have changed your supplier's economics. But your leverage comes from whether it has changed your alternatives.

Start with something you can actually test: what would we buy differently today?

I'm Zina, and this is Procurement at AI Speed, by KaiZina.

Your AI contract says the agent needs human approval before it makes a commitment. But what if the system still lets it make that commitment without approval?

You can write the rule into the contract. You can put it into policy. But unless something actually enforces that rule, what stops the agent?

In the last episode, I asked whether AI's promised time savings turn into commercial value. Now imagine the tool is doing more than saving time. It can act.

Picture this. Your procurement team is using an AI agent that can read contracts, communicate with suppliers and connect to your finance system.

Your internal policy says a person must approve any commitment over a hundred thousand dollars. The contract says the agent must stop for human approval at that threshold. So, on paper, the boundary looks clear.

But what if the agent's actual system permissions still allow it to send a message accepting a supplier's quote?

The contract says no. The policy says no. But can the system actually stop it?

If it doesn't, we may end up arguing afterwards about whether the agent had authority to make that commitment. That's too late. The better question is: what will stop the agent before it acts?

That's the gap procurement needs to pay attention to. A contractual restriction and a technical control are not the same thing.

This isn't just theoretical. On 28 September, NVIDIA announced technology designed to enforce boundaries outside the agent itself. I wouldn't make a buying decision from a product announcement. But it illustrates where the market is heading: telling an agent what it can do is one thing. Technically stopping it from crossing that line is another.

For procurement, I'd bring three questions into the sourcing and contract discussion.

First, what is the agent allowed to do? Reading a contract, drafting a counteroffer, sending it to a supplier and accepting a deal are four different levels of authority. Define the approved actions, data access, transaction limits and points where a person must decide.

Second, what actually enforces those limits? Don't just ask whether the control exists. Ask to see what happens when the agent tries to cross it. Give the agent a task outside its authority in a test environment. Does the system reject the action before it reaches the supplier or finance platform? What gets logged? And who is alerted?

Third, who owns that control after signature? There are three layers to this. What the contract says the agent can do. What the technology can actually block. And what controls we configure and turn on.

The supplier might provide the controls, but we might be responsible for configuring them and controlling access to our own systems. So the contract needs to be clear about who sets the limits, who tests them when things change, what evidence we receive, and who acts if something goes wrong.

Not every contractual obligation can be turned into a technical control. Incident notification, cooperation and remedies still matter. But if an agent isn't supposed to take an action, we shouldn't rely only on what happens after it does. We should also ask what will stop it in the first place.

Procurement doesn't need to design the security architecture. But it does need to make sure the commercial authority we negotiate matches what the agent is actually permitted to do.

So before the next agentic AI contract is signed, ask: which boundaries exist only in the contract, and which ones will the technology actually enforce?

I'm Zina, and this is Procurement at AI Speed, by KaiZina.

Imagine your procurement team tells you AI saved 10,000 hours. Sounds impressive. But my next question would be: what changed because of those 10,000 hours?

Did we get to important renewals earlier? Did we reduce the backlog? Did we have more time to investigate supplier risks, prepare for negotiations, or get to work the team previously couldn't? Or did everyone remain just as busy?

Last episode, I asked where procurement creates value when AI does more of the work. For me, this is where that question becomes practical.

Saving time is valuable. But freeing up time and turning it into a better business outcome are not necessarily the same thing.

BCG gives an interesting example from technology procurement. An automotive manufacturer saw productivity gains from early AI pilots, but struggled to turn that freed-up time into commercial results. The broader benefits came after the procurement lifecycle was redesigned. It's one example, not a rule for every AI implementation. But I think it makes the point.

Imagine AI cuts the time needed to prepare a supplier comparison from eight hours to one. Great. But the recommendation still waits two weeks for a decision. Requirements keep changing. And the category manager still doesn't have enough time to prepare properly for the negotiation.

We've made one task much faster. We may not have made the commercial decision any faster.

And even those seven hours saved are worth checking. Did we actually free up seven hours? Or did some of that time go into checking the output and correcting mistakes?

That's where I think procurement needs to follow the work beyond the task AI performs. Where does that freed-up time actually go?

Maybe we get to renewals earlier. Maybe we can cover more strategic sourcing activity with the same team. Maybe we improve risk oversight. Maybe we stop producing an approval or report that no longer adds value. And sometimes the benefit is simply a more manageable working day. That matters too.

But we should be clear about what we're claiming. Hours saved can represent capacity value. But capacity value isn't automatically financial ROI. If we're claiming ROI, we need to compare the benefits actually realised with the full cost of implementing and operating the technology.

So for every significant AI use case, I'd add one simple sentence: the time we free up will go towards this outcome, owned by this person, and we'll know it happened because of this evidence.

Then come back later and ask two questions: did AI actually reduce the work? And: did the outcome we wanted actually improve?

Because if an AI initiative reports thousands of hours saved, somebody should eventually be able to answer a very simple question: what did those hours become?

So the next time an AI initiative reports time saved, I'd ask one practical question: who owns turning that time into an outcome we can actually see?

And there's another commercial question worth coming back to: when AI makes our suppliers more productive, who gets the value?

I'm Zina, and this is Procurement at AI Speed, by KaiZina.

If AI does more of the work, where does procurement create value? If AI can increasingly do the work, what actually are we paying procurement people to do? I don't think the answer is to defend the work we've always done.

I think the better question is, where do we want our people creating value? Think about how much of procurement has traditionally involved finding information, analysing data, comparing suppliers, reviewing contracts, preparing documents, and coordinating processes. AI is getting better at all of it.

And increasingly, it can move beyond analysis and actually execute parts of the process. And I think that's a good thing. BCG recently estimated that redesigning procurement around AI could free up as much as 60% of buyer capacity.

But capacity on its own isn't value. What matters is what we do with it. Our role can't simply shift from producing the work to checking what AI produces. That would be a pretty limited view of the opportunity.

If AI gives us capacity back, I would rather see procurement spend more of that time on the decisions where we can actually change the commercial outcome.

This is an excerpt. Listen to the full episode on Apple Podcasts, Spotify or Amazon Music.