Agentic AI -- Can It Solve Resource Challenges?

Even though Agentic AI concepts go back to the 1950s, the term "Agentic AI" hadn't sunk into my IT awareness until a recent webinar by Dr. Daniel Barulli for the Association of Independent Information Professionals (Disclaimer: I'm AIIP's Director of Membership).

Dr. Barulli's presentation on Agentic AI and Deep Research tools intrigued me, as I pondered whether it could help small businesses and nonprofits alleviate their traditional resource challenges of Time, Money and Knowledge by turning hours of research into an hour or less.

Agentic AI agents achieve desired outcomes by determining and performing the tasks needed to achieve those outcomes. They conduct their own research instead of using training data, find external tools to complete tasks, pivot based on new data, and apply reasoned decision-making. They do so with little human input after the initial prompt.

Deep Research, such as offered by ChatGPT and Gemini, are among the capabilities Agentic AI agents can use in multi-step research of online sources -- traditional search engines, databases, media, other AI agents, etc.

In this blog, let's start with a look at some differences between Traditional and Agentic AI tools:

• Traditional AI Tools 
  • Short prompts
  • Shorter responses
  • Immediate responses
  • Uses training data
  • Doesn't reason
  • Doesn't automatically cite references
  • Workflow is not autonomous
  • Doesn't use much energy
  • Doesn't cost much

• Agentic AI Tools
  • Can use longer prompts
  • Much longer responses
  • Longer response times
  • Searches for knowledge rather than using training data
  • Engages in reasoning
  • Have memories (ChatGPT remembers the name of my upcoming book when it responds to my prompts related to it)
  • Autonomous workflows to a large degree
  • Hallucations, errors
  • Extremely energy hungry
  • Costs for SMBs range from $20 a month to much higher

An Experiment:

Aware of Gartner's 2024 study on Software Buyers' Regret, in which 59% of SMB decision-makers with less than 250 employees stated they have experienced long-term business performance issues due to a recent poor software selection, I wanted to know the cause of that level of dissatisfaction.

ChatGPT's Deep Research tool first asked to clarify if I wanted only U.S. businesses, wanted only information on critical SaaS apps, and if it should include indirect factors (poor support, integration issues, etc.) and / or direct factors (bugs, lacking features, slowness, etc.).

After clarifying, ChatGPT said it understood the assignment and proceeded to carry out the task without any further intervention from me. After six minutes, it reported its findings from 24 sources culled from 119 searches.

It returned data about the percentage of small businesses with less than 250 employees and then detailed the common causes of regret, both direct and indirect, including financial loss, operational disruptions, and productivity declines from poorly designed, slow apps, and staff having to revert to manual tools.

Not Perfect:

Rather than click all 24 sources, I followed up on its report to see if the sources cited were in the public domain or would require permissions to use in my upcoming book, marketing materials, or a blog or website.

It reported that the statistic from Gartner's 2024 study on Software Buyers' Regret would require permission from Gartner to use for commercial purposes. 

It then asked if I'd like to draft a formal permission request to Gartner for using the statistic in my upcoming book. I said yes. It crafted a short permission request email. However, the email address it provided bounced as non-existent. I then contacted Gartner online and received permission.

Traditional AI:

For comparison, I ran that same prompt through ChatGPT's regular LLM and didn't receive the depth of information about the causes of small business remorse with SaaS applications.

Instead, it returned outdated U.S. Census Bureau data from 2016, other data from unnamed sources, identified some common causes of negative performance related to new software, and offered a few tips on mitigating the risks.

My Current Opinion:

Agentic AI and tools like Deep Research have the potential to help SMBs with their major resource challenges -- Time, Money and Knowledge. It's probably good enough for a lot of people who won't be bothered with fact-checking now. 

Thinking about the use of these tools in my business, I plan to test these agents to see if they can gather data on SaaS vendors and populate that data into appropriate spreadsheet tables without my help. That would save my SMB clients significant time creating meaningful comparisons of competing SaaS providers.

You still need humans to fact-check their work, though, as they remain prone to hallucinations and errors such as the incorrect email address for Gartner's permissions department.

Despite the inaccuracies, I felt the 5 minutes of my time crafting the prompt and waiting for its response, plus the 30 minutes fact-checking its citations (mainly to see if the citations actually existed -- most did), saved me hours of searching and rummaging through long reports to find the information I wanted.

While I will spend more time fact-checking the report before using any of the data in marketing materials, it definitely provided a good, quick starting point for my research into buyers' remorse with SaaS solutions.