Enterprise AI spend is growing up fast, and some of the growing pains are starting to hurt.

In a recent conversation with 15 IT procurement leaders from Fortune 100 companies, NPI asked what they’re seeing as AI moves deeper into enterprise operations. The discussion covered ROI, consumption commitments, hardware costs, new commercial models and the skills procurement teams need as AI buying accelerates.

What we heard was less about where AI might be headed and more about what’s happening right now. Commitments are being burned through months earlier than expected. Discounts are disappearing when usage exceeds forecasts. Vendors can’t always explain what drives the bill. Hardware orders placed ahead of price increases are being repriced anyway.

At the same time, procurement teams are finding some approaches that work. Here are seven takeaways that stood out.

1. AI ROI is Still Pretty Fuzzy

Everyone wants to talk about AI ROI. Measuring it is a different beast. About 53% of the procurement leaders in our discussion said their organizations are tracking AI usage, while 40% are beginning to connect usage with output.

The biggest obstacle? Nobody defined the desired outcome at the beginning. When asked about their biggest ROI measurement challenge, 40% pointed to undefined outcomes, 27% lacked a baseline and 20% said they couldn’t isolate AI’s impact.

There’s also a cost problem hiding in the ROI equation. Token and consumption spend get most of the attention, but they’re only part of the bill. AI adoption can bring additional infrastructure, governance, security and control costs with it. As one procurement leader pointed out during the discussion, that adjacent spend is often missing from the ROI math.

For procurement, that puts more weight on the questions asked before the purchase:

• What outcome are we trying to achieve?
• What does the process cost today?
• How will we measure improvement?
• What other costs will this AI investment create?

If those answers don’t exist going into the deal, proving ROI later gets a lot harder.

2. The Best AI ROI Stories Are Surprisingly Small

One of the clearest ROI examples shared during the discussion didn’t involve a sweeping AI transformation.

An accountant built an agent using Claude and Copilot to replace a manual month-end accrual consolidation process. The task had taken roughly 10 hours each month. Now AI generates the output and the employee validates it, and the approach has been rolled out across the accounting team.

The math works because it’s simple. There was a defined task, a known amount of labor and a measurable change after AI was introduced.

The idea also came from an internal contest for the best AI use case, which points to something procurement leaders may want to consider. Instead of chasing an enterprise-wide answer to “What’s our AI ROI?”, look for repeatable workflows where the before-and-after comparison is obvious.

Those smaller wins can give organizations the evidence they need to make smarter decisions about where AI investment should go next.

3. A Two-Year AI commitment Could Be a Five-Month Problem

Companies are learning the hard way that overages need as much negotiation as anything else in the AI contract.

One company signed a two-year AI consumption commitment in mid-March. By August, it was gone. Once the committed volume was exhausted, the negotiated discount went with it. The company found itself back at the negotiating table, except this time the vendor knew exactly how much the customer was consuming.

Another participant described dealing with a similar issue.

This is where AI commitments can get dangerous. The discount may look attractive at signing, but if the consumption model is wrong, the customer can end up renegotiating with established usage, dependent users and very little time.

One procurement leader described another twist: after the company exceeded its commitment, the vendor treated the higher consumption level as the new baseline. There was no additional discount for the added volume because that growth was now considered expected.

The takeaway for procurement is important: negotiate the overage as carefully as the commitment.

Before signing, know what happens when consumption reaches 100%, 120% or 150% of forecast:

• What rate applies?
• Does the original discount survive?
• Are there predefined growth bands?
• What triggers a renegotiation?

With AI, the scenario you thought was an edge case can arrive surprisingly fast.

4. If Your Vendor Can’t Explain the Meter, Don’t Bet Big on the Forecast

AI pricing models are introducing a growing collection of tokens, credits, agents, hosts and consumption units. Understanding what actually generates a charge can be harder than it sounds.

One procurement leader spent more than two hours with a vendor trying to determine which agent classes carried a cost. The vendor couldn’t provide a satisfactory answer. Its own portal AI also calculated a host count roughly twice the customer’s actual engineering population.

The eventual solution was refreshingly practical: run it live for a month and see what gets billed.

That may be a useful model for other AI buyers. If the billable unit is unclear, a three-year forecast built on that unit isn’t going to become more reliable just because the spreadsheet is detailed.

Push for a pilot, measurement period or short-term arrangement that lets you observe real consumption before making a major commitment. Better usage data can be worth far more than a bigger discount against the wrong forecast.

5. Buying Hardware Early May Not Protect Price

You can’t talk about AI without talking about hardware.

One procurement leader described placing hardware orders early specifically to get ahead of announced price increases. The strategy made sense. The problem was that the orders weren’t fulfilled. The customer was eventually told the equipment would not be delivered unless the orders were repriced and resubmitted.

That raises an important question for any procurement team trying to buy ahead of hardware inflation: Does your price protection survive until fulfillment?

An early PO doesn’t necessarily protect the budget if the supplier retains the ability to reprice before delivery. With AI infrastructure demand putting pressure on parts of the hardware market, buyers need to look closely at how long quoted prices are valid, what happens when fulfillment is delayed and which contractual protections actually lock the price.

AI spend is much bigger than the AI invoice. Hardware, infrastructure and other adjacent costs belong in the conversation too.

6. Traditional Volume Discounts Are Harder to Apply

AI also creates an awkward question for procurement: what do you negotiate when volume itself is difficult to predict?

Some companies are looking beyond traditional volume discounts. In our discussion, 60% of participants said they’re exploring value-based commercial structures to address AI-driven cost increases.

Value-based models aren’t automatically better for the buyer. They require clear definitions, measurable outcomes and firm boundaries around how much value the vendor gets to capture. Procurement also needs finance and the business involved early enough to agree on those terms.

But as AI pricing becomes more consumption- and outcome-oriented, procurement teams will need more options than “buy more, get a bigger discount.”

7. Procurement Teams Are Learning AI By Doing AI

The final topic was talent, and one result stood out: roughly 85% of respondents identified AI fluency as the biggest area their teams need to develop.

One organization created a procurement AI committee staffed by employees already enthusiastic about AI. They’re building prompt libraries, demonstrating use cases and teaching their peers. Other companies have created AI champion programs, run hackathons and agent-building training, and added AI skills and tool knowledge to procurement job descriptions.

Some are going further, including AI usage in performance metrics or hiring roles specifically focused on AI tooling within procurement operations.

There’s a lesson here for procurement leaders trying to figure out how much AI training to buy. Some of the strongest examples we heard were homegrown.

Give people access to the tools. Give them real procurement problems to solve. Then create a way for the people finding useful applications to teach everyone else.

The next AI deal will benefit from what buyers are learning now

What made this conversation interesting wasn’t a big prediction about where enterprise AI is going, but the specificity of what procurement teams are encountering today.

A two-year commitment can disappear in five months. A vendor may struggle to explain the unit it wants you to commit to. Buying hardware ahead of an increase may not protect the price. And some of the easiest AI ROI to prove may come from an accountant automating 10 hours of monthly work.

These are the growing pains of a market where adoption is moving faster than many of the commercial practices around it.

For IT procurement, that makes the learning happening right now especially valuable. Test consumption assumptions before making large commitments. Negotiate what happens when the forecast is wrong. Define the outcome before trying to calculate ROI. Look beyond the AI invoice for the true cost. And give procurement teams room to experiment with the technology themselves.

For years, the conversation about Shadow IT was fairly straightforward: employees found a piece of software they wanted, signed up without going through the normal approval process, and created a visibility and governance problem for IT and procurement.

Shadow AI is making that problem much harder to spot. Today, AI doesn’t necessarily enter the enterprise through a rogue application or an unfamiliar supplier. It can arrive as a new feature from a vendor you’ve worked with for years. It can show up as an AI SKU added during a renewal, a consumption meter inside an existing cloud account, or an agent an employee builds on an approved platform. In many cases, there’s no new supplier, purchase order or sourcing event to alert procurement that something has changed. That changes what IT procurement teams need to look for and, just as importantly, how they manage AI spend.

What Is Shadow AI?

Shadow AI is the use or deployment of AI capabilities without adequate visibility, commercial oversight or governance. The important word here is capabilities. An enterprise may have approved a vendor or platform without fully understanding every AI capability being activated or created within it.

There are two primary paths. The first is vendor-driven AI, where an approved application introduces a new AI feature, pricing meter or commercial model. The second is employee-built AI, where employees use approved platforms to create agents, applications and automations of their own. The entry points are different, but both can leave procurement without a clear view of ownership, permissions, data use, pricing, monitoring and business outcomes. That’s why Shadow AI is, first and foremost, an inventory problem.

Why Traditional Procurement Controls Can Miss AI Spend

Most procurement processes are designed around recognizable commercial events. A business unit wants new software. A vendor needs to be onboarded. Someone submits a purchase request. A contract comes up for renewal.

AI increasingly bypasses those triggers. In our recent webinar, Welcome to the Era of Shadow AI: What Can Procurement Do About It?, we identified five common entry points: embedded AI within existing SaaS applications, AI SKUs and credits introduced during renewals, model and agent services consumed through cloud accounts, APIs purchased by engineering teams, and employee-built agents created on approved platforms.

The common thread is that the vendor may be visible while the AI capability is not. That creates a new challenge for procurement. You can have an approved supplier, negotiated contract and established purchasing process and still lack a complete picture of how AI is being used and what it is costing.

AI Pricing Makes the Visibility Problem More Expensive

AI also introduces a growing number of ways to charge. Depending on the supplier and product, enterprises may encounter per-user licenses, tokens, credits, actions, conversations, premium editions, add-ons and other consumption measures. Several of those meters can exist within the same supplier relationship.

Salesforce Agentforce offers a useful example. Its commercial models have included $2-per-conversation pricing, Flex Credits priced per action, and per-user licenses and add-on editions. Meanwhile, Salesforce increased pricing for some broader editions, citing additional value from AI capabilities.

For procurement teams accustomed to benchmarking seats and subscription rates, the job now includes understanding the meter itself. What generates consumption? Who can generate it? How quickly can usage grow? Where does it appear on the invoice? Can the organization forecast it? What happens when the vendor changes the commercial model? Those questions need answers before usage scales.

Procurement Has More Leverage Than It May Think

The fact that AI features are arriving through existing suppliers doesn’t mean enterprise customers have to accept every change on the vendor’s terms. A complete veto over a SaaS provider’s product roadmap is unlikely, but there are more practical protections that procurement teams are successfully negotiating.

NPI has seen enterprise customers negotiate no-training provisions for customer data, prior written consent requirements for vendor use of generative AI, and notify-and-explain provisions covering new AI features or changes. Supporting protections can include audit rights, AI price caps, exit rights and termination rights for material changes.

One particularly useful approach is requiring vendors to provide advance notice to designated customer contacts when they introduce a new AI feature, change an underlying model, alter how customer data is handled or change the associated fees. Procurement can also push for new AI capabilities to remain off and incur no fees until the customer chooses to activate them. That gives sourcing teams something they badly need in the AI market: time to evaluate a change before it becomes the new normal.

Employee-Built Agents Create Another Procurement Blind Spot

Vendor-driven AI is only half of the issue. Employees can now build apps, agents, copilots and skills in a matter of hours. These tools may read data, write to systems, send information, trigger workflows or even approve actions. An organization can approve the platform used to create an agent without ever reviewing the individual agent itself.

For procurement, the relevant question is expanding from “Which AI tools are employees using?” to “Which AI capabilities and agents can access our systems, data and business processes?” An enterprise AI agent registry can help close that gap by documenting the owner, approved data sources, permissions, security testing, usage monitoring, recertification requirements and a kill switch for each agent.

The goal shouldn’t be to count how many agents the company has deployed. A better measure is whether those agents are part of governed workflows tied to defined business outcomes, with clear decision rights and human approval where it matters. The webinar illustrates this with a procurement workflow spanning spend intelligence, sourcing, risk and contract agents, followed by a human decision gate and measurement of business outcomes.

Procurement Needs Visibility Across the Entire AI Supply Chain

There’s another reason supplier-level visibility is no longer enough: the company named on the contract may represent only one layer of the AI service. Behind that vendor can sit model providers, cloud infrastructure, data and retrieval services, orchestration layers, monitoring tools and other dependencies. Each can affect cost, data handling and commercial risk.

Procurement therefore needs to understand the full AI supply chain: vendor of record, underlying model providers, cloud infrastructure, data paths and retention, pricing meters, change-notice rights, and exit and portability provisions. If an underlying provider changes its pricing, terms or availability, the cost of an AI application can change even though your organization hasn’t signed a new agreement.

6 Questions IT Procurement Leaders Should Be Asking About Shadow AI

A useful test is surprisingly simple: If one of your vendors tripled your AI bill next quarter, would you know before the invoice arrived? If the answer isn’t an immediate yes, start with these questions:

1. How many vendors are charging us for AI? Include embedded AI, add-ons, cloud services, APIs, departmental purchases and expense channels.

2. Where is our AI spend occurring today? Break it down by vendor, business unit, deployment path and purchasing channel.

3. Which costs are consumption-based? Identify whether you’re paying by tokens, credits, actions, conversations or another measure.

4. Where are we paying for overlapping capabilities? Map what your assistants and agents actually do rather than looking only at vendor names.

5. Who owns each AI cost and business outcome? Every meaningful deployment should have both a business owner and a commercial owner.

6. Can we forecast AI spend with confidence? Build forecasts around rates, usage and adoption, and establish thresholds that trigger review when actual consumption moves outside expectations.

These questions move the AI conversation away from a generic inventory of vendors and toward the information procurement actually needs to govern spend.

What Should IT Procurement Do About Shadow AI Right Now?

Procurement shouldn’t try to solve Shadow AI alone. Finance and FinOps can help with allocation and forecasting; IT with architecture and inventory; security and privacy teams with risk controls; legal with enforceable terms; and engineering with optimization and business outcomes. Commercial governance, however, is squarely within procurement’s wheelhouse.

That means maintaining an AI inventory by vendor, capability, owner and deployment path; normalizing different pricing meters; identifying capability overlap; creating renewal triggers when AI SKUs, models, meters or terms change; negotiating stronger contractual protections; benchmarking pricing and terms; and working with FinOps to establish budgets, thresholds and anomaly alerts.

For teams wondering where to begin, NPI recommends four priorities for this quarter: build the AI inventory, instrument upcoming renewals with AI-specific questions and approval triggers, establish baseline unit economics, and begin a recurring cross-functional review of AI spend and anomalies.

The bigger change is moving away from treating Shadow AI as something you audit periodically. AI products, usage and pricing can change too quickly for a once-a-year exercise. Procurement needs an ongoing view of what is being used, what it costs, what changed and where intervention may be required.

How NPI Helps Enterprises Get Control of AI Spend

Getting visibility into Shadow AI is only part of the job. Procurement also needs a way to turn that visibility into better buying decisions, stronger agreements and ongoing cost control.

NPI helps enterprises do that in two ways. AI Agreement Optimization is designed for organizations approaching a specific AI purchase or renewal, combining consumption analysis, forecasting, pricing benchmarks, contract intelligence and negotiation support to help clients buy the right amount of AI at the right price and terms. NPI’s AI FinOps Managed Service provides ongoing analysis across an organization’s AI vendors, including consumption trends by vendor and business unit, optimization recommendations, invoice validation, renewal support and benchmarking against peer organizations.

Together, these services help procurement teams move from trying to piece together where AI spend is coming from to having the intelligence and ongoing oversight needed to manage it as it grows.

Watch: Welcome to the Era of Shadow AI

Shadow AI is already inside most enterprises, and the challenge now is finding it early enough to govern the commercial impact. In our webinar, Welcome to the Era of Shadow AI: What Can Procurement Do About It?, we dig deeper into how AI is bypassing traditional procurement controls, where hidden spend and risk are emerging, what contractual protections enterprise customers are successfully negotiating, and how procurement can build a more effective operating model for AI governance.

If you’re responsible for IT sourcing, software renewals, AI agreements or technology spend, it’s worth watching.

The most useful AI pricing signal of 2026 isn’t in anyone’s rate card. It’s in frontier vendor competition. One frontier vendor is resetting usage quotas so often that developers built a website to track it. The other is rationing access to its flagship model week by week because it can’t serve the demand. That asymmetry, not list price, is where enterprise buyers should be looking. Why? Because it tells you exactly how much leverage you have and how long the window stays open.

Competition is Negotiating on Your Behalf, If You’re Set Up to Receive It

When OpenAI and Anthropic filed for their IPOs, investors flagged the same underlying risk in both: the interchangeability of their products and how easily customers can move between them. That’s a risk for them. It’s a massive opportunity for enterprise buyers.

The Wall Street Journal has reported that OpenAI is weighing significant token price cuts specifically to win customers from Anthropic, in anticipation of Anthropic doing the same. You don’t get pricing behavior like that in a market with real lock-in. You get it when two vendors know the switching cost is low and are fighting to keep it that way.

Every quota reset, lifted rate limit, and preemptive price cut is a competitive concession. The question is whether a customer’s architecture and contracts let them capture any of it.

Watch How Vendors Operate, Not Just How Their Models Score

Don’t judge AI vendors only by who has the “best” model on paper. Pay attention to how they’re actually operating their business, because that has a bigger impact on what you’ll experience as a customer.

Independent intelligence indices put Anthropic’s newest flagship one point ahead of OpenAI’s best model. But the cost story paints a different picture. Recent cost-per-task analysis shows the OpenAI model completing equivalent work at roughly a third of the cost, and OpenAI is pairing that with unusually generous subscription quotas, frequent quota resets, and the temporary removal of rolling usage windows.

Anthropic, meanwhile, has publicly tied flagship availability to compute capacity, extending subscription access in one-week increments and stating it will restore standard access “as soon as capacity allows.” That’s not a criticism of either company. It’s market data. Capacity constraints on one side and aggressive generosity on the other mean the effective cost per unit of work can swing dramatically without a single line of the customer’s contract changing.

Cost Per Task is the Metric That Matters

Per-token price comparisons miss the point. A model that costs more per token but resolves a task in fewer steps can be cheaper. A model that’s cheaper per token but locked behind tight quotas can be more expensive once your teams hit the ceiling and shift to premium API rates.

The unit that should drive vendor decisions is cost per task completed at acceptable quality, measured on your own workloads. That’s the number that reveals whether a two-point benchmark gap is worth a 3x cost difference. In most enterprise workflows, it isn’t.

Five Ways to Convert Competition Into Savings

Make substitutability real. Interchangeability only creates leverage if you can act on it. A routing layer or model gateway that lets you shift traffic between vendors, and down to cheaper tiers, turns “we could switch” from a bluff into a credible negotiating position. NPI’s deal reviews show routing to the right model, not just the cheapest one, is driving 40–85% reported bill reductions.

Don’t lock long into a price war. If frontier pricing is about to fall, a three-year commit at today’s rates is a losing trade. Favor shorter terms and negotiate repricing protection or benchmark-triggered rate reviews for anything longer.

Contract for the volatility you’re seeing. Quota policies, rate limits, and model availability are changing week to week. Push for model-change protection, granular consumption reporting, hard spend caps, and data portability so that operational changes on the vendor side don’t quietly reprice your deal.

Time renewals to competitive events. A renewal that lands while your incumbent’s rival is publicly cutting prices and lifting limits is worth more than the same renewal six months earlier. Track the competitive calendar the way vendors track your fiscal year end.

Treat access risk as part of the cost. The past few weeks have shown that frontier model availability can change on short notice, whether from capacity constraints or policy decisions. A stack with routing and open-weight fallback options isn’t just cheaper. It’s insulated from a risk most buyers haven’t priced in.

The Window is Open, But It Favors the Prepared

Vendors are competing hard for your workloads right now, and that competition is producing genuine concessions: lower cost per task, looser limits, and anticipated price cuts. But none of it accrues to buyers who are architecturally captive or contractually locked in. The savings go to organizations that can measure their real cost per task, move work between vendors, and negotiate with the data to prove both.

If you’d like to see what this looks like against your own AI vendor portfolio, we’re glad to walk through a live example. Contact us today.