The market is officially three years post ChatGPT and many of the pundit bylines have shifted to using terms like “bubble” to suggest reasons behind generative AI not realizing material returns outside a handful of technology suppliers. 

In September, the MIT NANDA report made waves because the soundbite every author and influencer picked up on was that 95% of all AI pilots failed to scale or deliver clear and measurable ROI. McKinsey earlier published a similar trend indicating that agentic AI would be the way forward to achieve huge operational benefits for enterprises. At The Wall Street Journal’s Technology Council Summit, AI technology leaders recommended CIOs stop worrying about AI’s return on investment because measuring gains is difficult and if they were to try, the measurements would be wrong. 

This places technology leaders in a precarious position–robust tech stacks already sustain their business operations, so what is the upside to introducing new technology? 

For decades, deployment strategies have followed a consistent cadence where tech operators avoid destabilizing business-critical workflows to swap out individual components in tech stacks. For example, a better or cheaper technology is not meaningful if it puts your disaster recovery at risk. 

While the price might increase when a new buyer takes over mature middleware, the cost of losing part of your enterprise data because you are mid-way through transitioning your enterprise to a new technology is way more severe than paying a higher price for a stable technology that you’ve run your business on for 20 years.

So, how do enterprises get a return on investing in the latest tech transformation?

First principle of AI: Your data is your value

Most of the articles about AI data relate to engineering tasks to ensure that an AI model infers against business data in repositories that represent past and present business realities. 

However, one of the most widely-deployed use cases in enterprise AI begins with prompting an AI model by uploading file attachments into the model. This step narrows an AI model’s range to the content of the uploaded files, accelerating accurate response times and reducing the number of prompts required to get the best answer. 

This tactic relies upon sending your proprietary business data into an AI model, so there are two important considerations to take in parallel with data preparation: first, governing your system for appropriate confidentiality; and second, developing a deliberate negotiation strategy with the model vendors, who cannot advance their frontier models without getting access to non-public data, like your business’ data. 

Recently, Anthropic and OpenAI completed massive deals with enterprise data platforms and owners because there is not enough high-value primary data publicly available on the internet. 

Most enterprises would automatically prioritize confidentiality of their data and design business workflows to maintain trade secrets. From an economic value point of view, especially considering how costly every model API call really is, exchanging selective access to your data for services or price offsets may be the right strategy. Rather than approaching model purchase/onboarding as a typical supplier/procurement exercise, think through the potential to realize mutual benefits in advancing your suppliers’ model and your business adoption of the model in tandem.

Second principle of AI: Boring by design

According to Information is Beautiful, in 2024 alone, 182 new generative AI models were introduced to the market. When GPT5 came into the market in 2025, many of the models from 12 to 24 months prior were rendered unavailable until subscription customers threatened to cancel. Their previously stable AI workflows were built on models that no longer worked. Their tech providers thought the customers would be excited about the newest models and did not realize the premium that business workflows place on stability. Video gamers are happy to upgrade their custom builds throughout the entire lifespan of the system components in their gaming rigs, and will upgrade the entire system just to play a newly released title. 

However, behavior does not translate to business run rate operations. While many employees may use the latest models for document processing or generating content, back-office operations can’t sustain swapping a tech stack three times a week to keep up with the latest model drops. The back-office work is boring by design.

The most successful AI deployments have focused on deploying AI on business problems unique to their business, often running in the background to accelerate or augment mundane but mandated tasks. Relieving legal or expense audits from having to manually cross check individual reports but putting the final decision in a humans’ responsibility zone combines the best of both. 

The important point is that none of these tasks require constant updates to the latest model to deliver that value. This is also an area where abstracting your business workflows from using direct model APIs can offer additional long-term stability while maintaining options to update or upgrade the underlying engines at the pace of your business.

Third principle of AI: Mini-van economics

The best way to avoid upside-down economics is to design systems to align to the users rather than vendor specs and benchmarks. 

Too many businesses continue to fall into the trap of buying new gear or new cloud service types based on new supplier-led benchmarks rather than starting their AI journey from what their business can consume, at what pace, on the capabilities they have deployed today. 

While Ferrari marketing is effective and those automobiles are truly magnificent, they drive the same speed through school zones and lack ample trunk space for groceries. Keep in mind that every remote server and model touched by a user layers on the costs and design for frugality by reconfiguring workflows to minimize spending on third-party services. 

Too many companies have found that their customer support AI workflows add millions of dollars of operational run rate costs and end up adding more development time and cost to update the implementation for OpEx predictability. Meanwhile, the companies that decided that a system running at the pace a human can read—less than 50 tokens per second—were able to successfully deploy scaled-out AI applications with minimal additional overhead.

There are so many aspects of this new automation technology to unpack—the best guidance is to start practical, design for independence in underlying technology components to keep from disrupting stable applications long term, and to leverage the fact that AI technology makes your business data valuable to the advancement of your tech suppliers’ goals.

This content was produced by Intel. It was not written by MIT Technology Review’s editorial staff.