Misunderstanding of AI in the Workplace

May 27, 2026 - Lane Arnold

The Pullback

In the recent days and weeks, there has been a cataclysmic pullback of AI use in the workforce. Here are just a few of the many examples:

"The cost for my team was more for AI than humans" NVIDIA executive

"I have already blew out my whole budget for 2026 in AI related costs" Uber CTO

(Reportedly)"Microsoft is reportedly reducing internal use of Anthropic’s Claude Code after its AI bills started exploding as employee usage rapidly increased." External report about internal decisions at Microsoft

Most of the admissions of these executives and the alleged report about the internal going of microsoft point to one thing: costs. It points to skyrocketing costs, no thanks to several companies switching to token/usage based billing such as Anthropic, OpenAI, and Perplexity. Some of the rise in costs can also be explained via the Data Centers that supply these services causing an exponential rise in electricity demands, thus ballooning energy prices. Even with those factors, I disagree with them being the causation behind the pulling back of AI use. I believe the root cause is the result of misuse and misunderstanding of these AI tools and pipelines.

AI Tools are not SaaS

One of the biggest misconceptions in this industry has been appropriating modern AI tools and agentic pipelines with SaaS. Lets take a quick history lesson on what the purpose of SaaS was. Back in the early days of software development, software was bought for a one time price and delivered either to a store or mail via DVD. You would be holding the software to use on your personal computer in your hands. Now this came with several disadvantages: to receive updates you would have to purchase an update disk to receive new features, your authentication was often a long pair of strings attached to the container the disc came in. With the increasing commonality of the global internet in the early 2000's, companies began offering Software as a Service rather than an physical product, making the entire process of spinning up software, maintaining, and updating it much more convenient than physical ownership. A great example of this is Microsoft 365. If you were to manually maintain the Office Suite for a small business back in the day, you would have to install each piece of software manually to each computer while manually hosting and maintaining Active Directory (now called EntraID) on-prem. These days, its as simple as signing up for a monthly subscription at Microsoft's website, setting up EntraID & Intune in the cloud with some polices, then have people login to their work machine with their 365 credentials. Their office suite will always have the latest features and will always be updated, and the company will never have to bother maintaining on-prem AD.

AI tools and agentic pipelines do not follow the parameters of a SaaS product. Firstly, when AI models and tools get updated, they often have such an improvement in feature and quality that they are fundamentally different products. A great example of this is ChatGPT. In the early days of ChatGPT (GPT-3.5), the product was a novelty, quite slow, would hallucinate often, but could occasionally spit out useful things like recipes and small-scoped functions in something like python. The followup release of GPT-4, had a completely different feature set, allowing images to be sent to the model as a basis for inputs rather than just text. The follow up release (GOT-4o) could then process and generate outputs including text, audio, and images. These rapid changes in capabilities and features are incredible, and not a bad thing, but the instability and need for rapid pivoting in order to continue to effectively use these products. If I use Adobe Photoshop version 22, and am well versed in it capabilities, I can use Adobe Photoshop version 24 without much thought or need for re-training. The same cannot be said for AI tools and workflows.

Secondly, AI tools do not have consistent, understood pricing. We covered this a bit earlier with multiple companies switching to usage token-based pricing, but one of the hallmarks of SaaS is a predictable, consistent pricing model with transparent (often percentage based) pricing increases. This makes the finance office quite happy as the can fit SaaS into their overhead budgets as a consistent line item, that does not increase much quarter to quarter. AI tools are the antithesis of this: their pricing models change on a whim, and their pricing models vary wildly from different service providers or even different services within the same provider. Granted, current global events do not help the AI industry in its pricing fluctuations, but its pricing models and unpredictable bills are nowhere close the maturity for the finance office to consider them stable line items in a budget.

Thirdly, all users are a part of the AI product. Yes, I know that some of these AI provides come out and swear that they do not use their client's data to improve the product, but let's be real, they do. They use your feedback and data to further help train the model and improve the next version of the product. Now this in and of itself is not a new thing. Going back to the Microsoft 365 example, Microsoft regularly tracks usage metrics for their products to help improve the application and the infrastructure to support the SaaS product. However, microsoft word is an established product, and its users simply help dictate if a new feature to MS word was well received, or needs some tweaks. AI products tracks the information you feed it and the outcomes from the prompt to help improve the product, not the mere use of the product or a specific feature. This is a major concert for both privacy and for Intellectual Property. Because how is a company supposed to weigh the following options: use it and risk your IP to be used to improve the product, or not use the product nad miss out on efficiency gains. Hardly an issue that arrives with Slack or MS 365.

In conclusion, these AI tool and pipelines have the following problems: they are updated with features and capabilities with breakneck pacing, they have inconsistent and unstable pricing modes, and they use the vary data we feed them to design and develop the next iteration of the product. They are NOT SaaS products. To be clear, there is nothing wrong with these tools, nor their pricing models. When utilized correctly, productivity can increase by 10X, but when presented as SaaS to the finance division, it often results in bloated costs, confusing and everchanging procedures to utilized them effectively, and has the potential to leak their most precious IP in the process of improving the product. AI tools and pipelines are not SaaS products, they are incredible toolsets presented via SaaS marketing.