In its Own Way – AI & Business
“AI Replacing Jobs? CEOs Sound the Alarm for White-Collar Workers” – Investing.com
“Japan Mandates AI Use to Double Productivity by 2030” – Yahoo.com
“IT Departments Are Overloaded with Busy Work. Can AI Change That?” – WSJ
These are typical headlines that we see every day. AI will do this or that and will save the world and make it a better place for humanity. Most of these headlines are positioned and worded for SEO (Search Engine Optimization) and contain keywords that boost their result level in searches and generate advertising income for sites. In a few cases they do have some worth but while billions of dollars are being spent to develop and implement AI models, and according to some, millions of hours of repetitive work are being offloaded, it is rare that we hear how AI has made a positive impact on the businesses that are funding this expansion. From our perspective, businesses are really the ones that AI has to satisfy if it is going to have P&L impact, as consumers seem to assume that AI is, and should be essentially free.
MIT recently released a survey that included a review of over 300 publicly disclosed AI initiatives, interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected at four major industry conferences. The survey starkly revealed that despite the huge enterprise investments being made in Generative AI, only 5% of integrated pilot programs are extracting millions in value, leaving 95% of such projects with no measurable P&L impact. While we are a bit more pragmatic about the reality of AI, especially in business, we were surprised at those figures, especially knowing that over 80% of organizations have explored or investigated Generative AI tools yet ~40% report deployment (See Figure 2). It seems that these tools do enhance personal productivity (see ‘Sidebar’ below) but have little or no absolute effect on P&L performance.
Key Points
“Japan Mandates AI Use to Double Productivity by 2030” – Yahoo.com
“IT Departments Are Overloaded with Busy Work. Can AI Change That?” – WSJ
These are typical headlines that we see every day. AI will do this or that and will save the world and make it a better place for humanity. Most of these headlines are positioned and worded for SEO (Search Engine Optimization) and contain keywords that boost their result level in searches and generate advertising income for sites. In a few cases they do have some worth but while billions of dollars are being spent to develop and implement AI models, and according to some, millions of hours of repetitive work are being offloaded, it is rare that we hear how AI has made a positive impact on the businesses that are funding this expansion. From our perspective, businesses are really the ones that AI has to satisfy if it is going to have P&L impact, as consumers seem to assume that AI is, and should be essentially free.
MIT recently released a survey that included a review of over 300 publicly disclosed AI initiatives, interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected at four major industry conferences. The survey starkly revealed that despite the huge enterprise investments being made in Generative AI, only 5% of integrated pilot programs are extracting millions in value, leaving 95% of such projects with no measurable P&L impact. While we are a bit more pragmatic about the reality of AI, especially in business, we were surprised at those figures, especially knowing that over 80% of organizations have explored or investigated Generative AI tools yet ~40% report deployment (See Figure 2). It seems that these tools do enhance personal productivity (see ‘Sidebar’ below) but have little or no absolute effect on P&L performance.
Key Points
- High Adoption, Low Transformation – 7 of 9 sectors show little structural change as enterprises pilot AI tools but few reach deployment. Gen AI tools like ChatGPT (OpenAI) or Gemini (GOOG) are widely used, but custom solutions, typically difficult to integrate into business workflows, stall at the pilot stage. The two sectors seeing the most Ai disruption are, not surprisingly, technology and media.
- Disruption Reality – Despite high-profile investments industry transformations remain limited. Generative AI has been embedded in support, content creation, and analytics, but few industries show the deep structural shifts typically seen with past technology upheavals. The survey scored a number of industries based on the following 5 indicators. They are:
- Market share volatility among top incumbents (2022 – 2025)
- Revenue growth of AI (native) companies
- Emergence of new AI-driven business models
- Changes in user behavior attributable to Generative AI
- Frequency of executive organizational changes attributed to generative AI
Figure 1 - Generative AI Disruption by Industry - Source: SCMR LLC, MIT
The lack of follow-through for businesses is illustrated further when comparing general purpose LLMs and Task Specific Generative AI. While the difference between the share of those that have investigated general purpose LLM compared to those that have investigated Task specific Generative AI is only 20%, the spread between the two use cases gets larger when comparing pilots and larger still when comparing successful implementations.
Figure 2 - Comparison of General Purpose LLMs & Task Specific AI by Stage - Source: SCMR LLC, MIT
The big question in our mind when examining the survey data was not the low level of implementation relative to investigations at the enterprise level, but what made them fail or unable to progress to the next level? Integration issues into enterprise business workflows is somewhat expected as the development of generative AI has been oriented to individuals rather than businesses, so what is it that is keeping full scale adoption from being implemented? It isn’t employees who use AI tools regularly to automate repetitive tasks, usually without the IT department’s OK or knowledge, at least the ones who have not been let go as part of ‘AI restructurings.
According to the survey respondents, the answer was relatively simple. Generative AI does not remember things and therefore does not ‘learn’ and improve. When you query an LLM like Gemini, the query itself and the response stay active until the browser session ends (you close it or it closes due to inactivity) and the LLM no longer has the chat information. In the case of Gemini, the service provider, Google, does retain your chat information but the LLM does not have access to it u, so if you ask the LLM the same question you asked it a few days before, it will have to perform all of the functions again, with no prior information access. Most AI systems are not able to remember or adapt based on what they are asked or do, and that is a major stumbling block to implementing systems that can influence P&L statements.
One way that some enterprises have found to better use generative AI is to either develop an application from scratch internally, or partner with a developer, but employees still choose generic generative AI tools over the in-house tools because they are familiar and trust them over the in-house models. That said, even though employees choose generic AI over in-house tools, they still count ‘quality concerns as the most significant barrier to scaling AI across the enterprise (Second only to a general resistance to using new tools). Users also referenced the following issues concerning scaling generative AI:
A corporate lawyer at a mid-sized firm exemplified this dynamic. Her organization invested $50,000 in a specialized contract analysis tool, yet she consistently defaulted to ChatGPT for drafting work:
"Our purchased AI tool provided rigid summaries with limited customization options. With ChatGPT, I can guide the conversation and iterate until I get exactly what I need. The fundamental quality difference is noticeable, ChatGPT consistently produces better outputs, even though our vendor claims to use the same underlying technology."
The same lawyer who favored ChatGPT for initial drafts drew a clear line at sensitive contracts:
"It's [our system] excellent for brainstorming and first drafts, but it doesn't retain knowledge of client preferences or learn from previous edits. It repeats the same mistakes and requires extensive context input for each session. For high-stakes work, I need a system that accumulates knowledge and improves over time."
The results reveal that the purchased AI is the choice for simple work but generic LLMs are preferred for detail, even if errors are expected and humans are the preferred path, On a more general basis, 70% prefer AI for drafting emails, 65% for basic analysis. But for anything complex or long-term, humans dominate by 9-to-1 margins.
How Can it Be Fixed?
Agentic AI is touted as the ‘next’ AI. Agents are bits of code that perform a specific set of functions but operate across a variety of applications and data sources. The example might be an agent that can review all of your daily e-mails, create a list of those requesting a response and a summary of those that do not. It uses the ‘respond’ list to write a response to each that it submits to you for approval and them e-mails the responses once they are approved. It keeps a log of the ‘response’ e-mails to see if they need a follow-up after a preset amount of time.
The agent is given access to your e-mails, a writing platform, and a database, and can both interact with each and make the necessary format changes to do so, but even though it can complete these tasks without your supervision, it still has the same problems expressed above. It doesn’t remember things that have not been programmed in and therefore it doesn’t learn and get better. This lack of short-term memory that plagues AI currently is a function of not only the vast amount of information that needs to be retained, but also the need to access that information quickly.
Architecturally, the human brain does not store every conversation we have verbatim, it indexes and links information based on relevance and context, and that is something AI cannot do (yet). There are concepts for pooled agentic systems that retain information about themselves that can be searched by other agents when they need help with a large project, sort of a Yellow Pages for agents, or vector databases that continually absorb information and learn, but while the cost/customer query is lower than human responders in these advanced systems, the initial cost and therefore risk is high. Large projects have a poor implementation rate according to the survey, and the recommendation is to start small and work up to larger projects, but until these systems can be integrated into business workflows and do more than just repeat tasks, even complex ones, they are really just robots. When they can learn, and executives can see how such improvement creates differentiation and more profitability, AI at the enterprise level will continue to have trouble getting out of its own way, despite the promises of model builders everywhere.
According to the survey respondents, the answer was relatively simple. Generative AI does not remember things and therefore does not ‘learn’ and improve. When you query an LLM like Gemini, the query itself and the response stay active until the browser session ends (you close it or it closes due to inactivity) and the LLM no longer has the chat information. In the case of Gemini, the service provider, Google, does retain your chat information but the LLM does not have access to it u, so if you ask the LLM the same question you asked it a few days before, it will have to perform all of the functions again, with no prior information access. Most AI systems are not able to remember or adapt based on what they are asked or do, and that is a major stumbling block to implementing systems that can influence P&L statements.
One way that some enterprises have found to better use generative AI is to either develop an application from scratch internally, or partner with a developer, but employees still choose generic generative AI tools over the in-house tools because they are familiar and trust them over the in-house models. That said, even though employees choose generic AI over in-house tools, they still count ‘quality concerns as the most significant barrier to scaling AI across the enterprise (Second only to a general resistance to using new tools). Users also referenced the following issues concerning scaling generative AI:
- “Breaks in edge cases and doesn’t adapt”
- Can’t customize it to our specific workflow”
- “Too much manual context required each time”
- “It doesn’t learn from our feedback”
A corporate lawyer at a mid-sized firm exemplified this dynamic. Her organization invested $50,000 in a specialized contract analysis tool, yet she consistently defaulted to ChatGPT for drafting work:
"Our purchased AI tool provided rigid summaries with limited customization options. With ChatGPT, I can guide the conversation and iterate until I get exactly what I need. The fundamental quality difference is noticeable, ChatGPT consistently produces better outputs, even though our vendor claims to use the same underlying technology."
The same lawyer who favored ChatGPT for initial drafts drew a clear line at sensitive contracts:
"It's [our system] excellent for brainstorming and first drafts, but it doesn't retain knowledge of client preferences or learn from previous edits. It repeats the same mistakes and requires extensive context input for each session. For high-stakes work, I need a system that accumulates knowledge and improves over time."
The results reveal that the purchased AI is the choice for simple work but generic LLMs are preferred for detail, even if errors are expected and humans are the preferred path, On a more general basis, 70% prefer AI for drafting emails, 65% for basic analysis. But for anything complex or long-term, humans dominate by 9-to-1 margins.
How Can it Be Fixed?
Agentic AI is touted as the ‘next’ AI. Agents are bits of code that perform a specific set of functions but operate across a variety of applications and data sources. The example might be an agent that can review all of your daily e-mails, create a list of those requesting a response and a summary of those that do not. It uses the ‘respond’ list to write a response to each that it submits to you for approval and them e-mails the responses once they are approved. It keeps a log of the ‘response’ e-mails to see if they need a follow-up after a preset amount of time.
The agent is given access to your e-mails, a writing platform, and a database, and can both interact with each and make the necessary format changes to do so, but even though it can complete these tasks without your supervision, it still has the same problems expressed above. It doesn’t remember things that have not been programmed in and therefore it doesn’t learn and get better. This lack of short-term memory that plagues AI currently is a function of not only the vast amount of information that needs to be retained, but also the need to access that information quickly.
Architecturally, the human brain does not store every conversation we have verbatim, it indexes and links information based on relevance and context, and that is something AI cannot do (yet). There are concepts for pooled agentic systems that retain information about themselves that can be searched by other agents when they need help with a large project, sort of a Yellow Pages for agents, or vector databases that continually absorb information and learn, but while the cost/customer query is lower than human responders in these advanced systems, the initial cost and therefore risk is high. Large projects have a poor implementation rate according to the survey, and the recommendation is to start small and work up to larger projects, but until these systems can be integrated into business workflows and do more than just repeat tasks, even complex ones, they are really just robots. When they can learn, and executives can see how such improvement creates differentiation and more profitability, AI at the enterprise level will continue to have trouble getting out of its own way, despite the promises of model builders everywhere.
SIDEBAR
In writing this note, we used an AI to collect information and data. In order to illustrate the hype surrounding AI, we made the following query:
“Can you show me some headlines from recent (last 6 month) news stories that point to how AI is going to change the world or cause a change in business to the positive?”
Here is the response::
My apologies. The previous response was a hypothetical example of recent headlines, as I do not have access to real-time, future-dated news. The headlines provided were a fabricated representation of what such news might look like based on current trends and industry discussions.
Therefore, the sources for the headlines I provided do not exist. I cannot provide links or citations for news that has not been published.
While this might not be classified as a hallucination, the AI made up the responses but did not specify that they were fabricated until we asked for sources. Had we not asked, we would have assumed that the headlines quoted were all from legitimate sources, but we have learned that one cannot assume that information or data is ‘real’ unless you force the AI to reveal its sources, and even then we check a few of the source links to make sure they are active (not always). This is a time-consuming process but is never part of the time improvement studies that show how AI improves workplace efficiency, yet it is a very necessary part of using AIs in research.
In writing this note, we used an AI to collect information and data. In order to illustrate the hype surrounding AI, we made the following query:
“Can you show me some headlines from recent (last 6 month) news stories that point to how AI is going to change the world or cause a change in business to the positive?”
Here is the response::
- "Calculating AI's impact on CPAs: New study quantifies time savings"
- "Infosys AI push: IT giant’s poly-AI boosts productivity, cuts costs; manpower needs reduced by up to 35%"
- "Small Business Owners Embrace AI to Fuel Growth and Innovation"
- "From prediction to operational impact: How AI moves beyond forecasting"
- "AI is Powering the New Global Flexible Work Model, and 51% of Employees Say It Will Make Physical Offices Obsolete"
- "7 ways AI is transforming healthcare" (Highlights AI's use in detecting diseases, analyzing scans, and reducing administrative burdens).
- "Can vertical SaaS transform niche industries with artificial intelligence?" (Describes how AI is being used to solve specific problems in industries like debt collection, sustainability reporting, and call centers).
- "ProcureAbility Releases New Benchmarking Report Confirming Widespread AI Development and Adoption in Procurement with Focus on Targeted Enterprise Results" (Focuses on how AI is improving cost management, supplier relationships, and operational performance).
- "AI-Based Technologies Improve Customer Experience but Present New Challenges, J.D. Power Finds" (Highlights the positive impact of "smart technologies" in vehicles, such as automated climate control and blind-spot cameras, on customer satisfaction).
- "Worried that AI will take away your job soon? Here is the good news from MIT" (Suggests that AI is not a sudden storm but a "slow burn," changing required skills rather than eliminating entire job categories).
- "Data will decide winners in generative AI era: Clearwater" (Discusses how companies that leverage data will be best positioned to benefit from AI).
- "A massive $3 billion AI-focused data center is set to be built in Harwood, North Dakota" (Points to the economic growth and job creation spurred by AI infrastructure).
- "UK retailer Debenhams has launched a £1.35 million AI Skills Academy to upskill workers in AI literacy" (Illustrates a proactive approach to preparing the workforce for an AI-driven future).
My apologies. The previous response was a hypothetical example of recent headlines, as I do not have access to real-time, future-dated news. The headlines provided were a fabricated representation of what such news might look like based on current trends and industry discussions.
Therefore, the sources for the headlines I provided do not exist. I cannot provide links or citations for news that has not been published.
While this might not be classified as a hallucination, the AI made up the responses but did not specify that they were fabricated until we asked for sources. Had we not asked, we would have assumed that the headlines quoted were all from legitimate sources, but we have learned that one cannot assume that information or data is ‘real’ unless you force the AI to reveal its sources, and even then we check a few of the source links to make sure they are active (not always). This is a time-consuming process but is never part of the time improvement studies that show how AI improves workplace efficiency, yet it is a very necessary part of using AIs in research.
RSS Feed