The AI Credit Blowout: Why So Many Teams Are Spending Big and Getting Little

Is your AI spend actually under control? And are you getting the value you expected for it? Depending on how you're using AI, whether that's improving business processes or building software, it's worth watching closely.

Uber reportedly burned through its entire 2026 AI budget in just four months. Microsoft hit a similar wall, opening up AI tools to thousands of employees, only to course-correct when the sheer scale of adoption became the problem. These are some of the most sophisticated tech operators in the world, and they still got caught out.

During one of our recent Agentforce implementations, our consultants noticed during the testing phase that the AI token consumption was running significantly higher than expected. Our customer had uploaded user guides directly into a Data Library so the AI agent could reference them. However, because cost was tied to file size and processing indexing, those PDF guides, which were packed with heavy screenshots and high-res diagrams, were quietly driving up credit usage every single time they were indexed.

Because we caught it early, we optimised the documentation format and resolved the issue. Left unchecked, that single detail would have drained thousands of credits a month.

First Principle: Knowing When AI Isn't the Answer

Our Head of Technology puts it simply. Not every task needs AI. In fact, most don't. Understanding the difference between automation and AI is one of the easiest ways to protect your budget. 

• Automation follows fixed rules. Formatting a document, moving a file when a checkbox is ticked, or triggering a scheduled reminder. Rules get the job done for a fraction of the cost.
  
• AI makes judgments. Reading open-ended customer feedback, summarising a complex call, or spotting patterns across unstructured data.

The trap is reaching for AI when a plain rule would work. It gets the job done quickly, so nobody notices, until you realise you’re paying premium rates for basic, deterministic tasks.

A new perspective: Measure Value Per Credit, Not Raw Prompts 

The question “What's actually driving the spend?” led one of our consultants to explore a different lens on this. 

At small scale, AI cost is easy to contain. You can see what's running and what it's for. At scale, that clarity disappears. The spend isn't just bigger, it's harder to trace back to anything. 

To control spend without killing employee momentum, teams need to shift from chatting with AI to managing structured AI workflows.

A Level 3 agentic workflow might consume more credits on a single run than a simple chat prompt. But if that workflow replaces three hours of manual data entry or case management, the value per credit is exponentially higher. 

The real problem isn't increased consumption. It's consumption with no clear line to output. Not "how do we use less AI" but "how do we make sure every credit is doing something worth paying for."

What Smart Leaders Should Do Next

Getting control of AI spend doesn't mean locking everything down and killing momentum. It requires treating AI like cloud infrastructure rather than a static software licence.

One of our consultants, who's been close to several of these rollouts, shared a few things worth flagging. Here is how forward-thinking leaders manage it: 

• Treat AI spend like cloud consumption. Set active management practices, establish clear allocation budgets by team, and assign account owners to review usage.
• Watch the first 90 days closely. Teams often build and test using free trial credits, which hides real operational costs. Monitor usage spikes aggressively during the first three months of live deployment.
• Audit for low-hanging fruit. Audit existing setups for wasted spend, whether that’s oversized files in Data Libraries, redundant agent steps, or background tasks firing too frequently.
• Match the tool to the job, per action. One approach doesn't fit every business process. Some tasks suit plain rule-based automation. Others work better as an enhanced workflow where AI speeds up a person. Others justify an autonomous agent. Picking the right model per action makes the cost easy to justify at the board level later.
• Move people up the levels, not just into the chat window. If your AI strategy stops at "everyone has access to AI," you're optimising for adoption, not value. The goal is to get people thinking like AI managers, defining the task, setting the boundaries, reviewing the output, rather than just firing off prompts and hoping for the best. 
• Set spend allocations per level. A casual Level 1 chat habit across hundreds of employees hits the bill very differently to a Level 3 agent running a critical workflow. Allocate budgets accordingly, set thresholds, and adjust as your usage matures.

The Bottom Line 

Nobody has fully worked out the economics of AI yet. But AI shouldn't be judged by how many prompts it processes or how many credits it consumes; it should be judged by what it actually produces.

The organisations that win won't necessarily be the ones spending the least on AI. They’ll be the ones that know exactly what every credit is buying.

Let's Compare Notes 

At ProQuest, we're genuinely curious about getting this right and happy to work it out alongside you. Being part of Decision Inc. means we spend our days testing what works, solving implementation traps, and figuring out the messy bits together.

If any of this sounded like your organisation, come have a chat. AI will keep evolving, so let's keep comparing notes. What are you seeing in your business?