Follow
Follow
News
June 15, 2023 by Daniel W. Rasmus, Serious Insights
Technology like generative AI generates hype. Because of its close affiliation with science fiction, the concept of AI has been front and center in many plots involving humanity’s extinction, but probably not as often as humanity has been its own worst enemy.
If we step back from the science fiction tropes, practical uses for generative AI exist, though they are hard to find even among the less existentially threatening use cases.
First, I outline two different but similar use cases that focus on the summarization feature of generative AI, and then provide a list of other use cases, along with a brief analysis of each one.
[Cover image via my prompt to DALL·E]
Singularity Digital Enterprise (SingularityDE) worked with a contact center as part of its continuous effort to improve after-call analytics.
As with most call centers, tagging and summaries fall into manual processes. These processes often reflect assumptions by the operators that might not prove correct, including the introduction of bias based on several factors, including identifying patterns that may not exist.
Perhaps most importantly, operators multi-task, primarily with the charge to solve or ameliorate customer issues. Back-office work, such as tagging and summarization, fall as a secondary priority, as they should, to customer-facing work.
Photo by Tima Miroshnichenko on Pexels.com
SingularityDE’s solution summarizes phone calls in near real-time using Azure OpenAI. The technology also augments phone calls by detecting intent, performing sentiment analysis, and identifying calls-to-action. They can even verify that the operator followed the appropriate script.
Their proposed operations display includes three panels, one with the audio of the call available for replay, a transcript panel, and an analytics panel. This data can, of course, be easily integrated into CRM systems to enhance the overall visibility of customers and their concerns.
Using generative AI in call centers in this way does not threaten the call center operators. It offers them more time to manage customer expectations and problem-solve. What it does eliminate is the need for people to do jobs that detract from their highest-order tasks and goals—overall reducing workloads by removing not just unliked tasks, but tasks for which the AI may be better suited to provide results that the operators can apply to make their interactions more effective.
Using AI for this also elevates the timeliness of aggregate analytics by making summaries, sentiment and intent more accessible so that the teams accountable for engineering and other customer solutions can proactively reduce recurrent issues.
Singularity Digital Enterprise is a Serious Insights client.
Ripcord sells enterprise document management software. They recently demonstrated their early foray into ChatGPT for content summarization and data extraction.
Content summarization has proven an elusive feature, although it has existed for many years, often alongside automatic tagging tools that attempt to automate taxonomic features by inferring categories and subcategories associated with content. These categories can include the type of content, for instance, a sales receipt, or the topic of the content such as a report about politics and the environment—two concepts commonly tagged in the analysis of an EPA report.
Photo by Polina Tankilevitch on Pexels.com
Because summarization has traditionally been computationally restrictive, it becomes a metadata feature, a field filled algorithmically by background processes. The field may also be filled out manually.
The summaries become static once posted. The problem with static summaries is that they are written by a reader, not many readers. They summarize what that reader believes is essential, whether algorithm or human. And they usually do so in one language, as summarization is expensive, though some may invest in multi-lingual summaries.
Ripcord is testing the elimination of static summaries in favor of generative AI summaries. The summaries can appear in the speaker’s language, created in real-time. The demonstration showed summary text flowing into place as the query results were revealed.
Auto-summarization provides not only the benefit of eliminating summarization as a costly function, but it can also create a summary in any language for which the generative AI system is proficient, instantly localizing content that might otherwise never be localized.
Perhaps more interestingly, and not in the demo, is the idea of creating points of view as a part of the summarization prompt. Rather than a generic summary, a CFO could create a financial point-of-view that would emphasize the financial aspects of a piece of content over its other features.
To further extend this idea, consider real-time exclusions from a document set returned based on a concept but for which no point-of-view exists within a piece of content. If the document does not include, for instance, financial information, it could be dropped from the query, saving the CFO time from reviewing even the summary of content that is not relevant.
Another implication of generative AI in summarization is a movement away from a codified enterprise truth. Some may see this as risky. With generative AI, it is improbable that any two encounters with the same content by the same reader or another reader will produce the same summary. In a static summary world, the summary becomes metadata, part of the record, and, therefore, part of the “source of truth” associated with the data.
In a generative AI world, the summary could differ at each encounter, informed by the algorithm’s generative text design and additional training between encounters. A better-trained AI may offer a more nuanced summary—over time—the summaries may not only be different but better.
It is also possible to tie personal preferences into summarization so that the financial point-of-view discussed above plays a more infusive role. Summary clusters could evolve, informed by the reader’s work, to identify convergent ideas within the query. The summaries could inform queries that reinforce or complement existing research.
The long-term effectiveness of generative AI will be its inclusion in what I call the “last mile of automation”—those tasks that require knowledge of the end user and access to their data to guide them through routine tasks.
I’m still looking for AI that will prepare me for a meeting based on what it knows about me and the meeting and my content. Preparing me for a meeting, however, requires access and pre-processing (and ongoing processing) of all of my content.
Automating passive tasks is well and good, but being able to manage active workflows is where AI will really pay off. When I can ask a prompt a request like: Read this document on scenario planning and identify, with references, missing topics by analyzing all of my scenario planning research stored in OneDrive, (assume many PDFs scattered across several folders at various levels of depth with no structure or metadata), AI will truly be adding value to my process. I don’t need generic answers to generic prompts. I need specific actions in specific situations.
For now, I’m just going to list the possible solutions of generative AI for the enterprise with a brief description. As Serious Insights briefs with more generative AI companies, we will expand the use case set and provide a deeper analysis of the solutions, their benefits, and their pitfalls.
Photo by luis gomes on Pexels.com
Developers will need to determine on a case-by-case basis if generative AI, particularly ChatGPT, improves their productivity; it can write code and suggest improvements or fixes to bug-ridden code. ChatGPT understands the syntax for several languages, from Python to old-school languages like Lisp.
Generative AI could use similar technology to summarization to create user guides that meet the needs of particular users. Some users might want specific, expert guidance on sophisticated functions; others might command a system to tell me how to set up my toaster oven like I’m a 5-year-old.
For new assignments, generative AI might, as it does in tools like Miro, create a starter list of ideas for assignments, such as creating an outline for a social media campaign.
Bard’s response to the prompt: create an outline for a social media campaign
Here are some additional tips for creating a successful social media campaign:
By following these tips, you can create a successful social media campaign that will help you to achieve your business goals. |
Another demonstration from Ripcord demonstrated the extraction of data from a receipt in a language other than the target system—for instance, a Mandarin receipt as input to a U.S.-based expense system.
In more sophisticated situations, the system can identify data associated with promoted classes from long-form text. For example, ChatGPT could extract data classes for company and stock symbols from the report and write them out in JSON format.
The SingularityDE solution included sentiment analysis, as do tools like Grammarly. Generative AI can be presented with text and asked to identify the sentiment associated with that text. Employee and customer-facing systems that receive text-based responses and social media posts will likely widely adopt sentiment analysis.
Predictive analytics is an area where I will profess some skepticism because of my scenario planning background. While presenting ChatGPT with market data can result in a model that could predict which new product features, for instance, might appeal to which existing customers. Unfortunately, context limits all predictive analytics.
Photo by PhotoMIX Company on Pexels.com
If context changes, and ChatGPT’s rapid rise to prominence is an example, the historical data will prove increasingly inaccurate as the new context establishes itself. Predictive analytics in volatile markets requires more than AI. The same is true, for instance, in predictive analytics for supply chains where unusual disruptions may invalidate underlying assumptions and the models built on those assumptions.
Predictive analytics will likely prove more reliable for other uses, such as patterns in machine failure to inform maintenance interventions inside of enterprises and for commercial and consumer products.
Modeling, model testing, and model synthesis could help speed innovations.
Fraud detection requires intimate customer data, such as their use and travel patterns. Many such systems already exist. Generative AI may be another tool for identifying activities that sit outside of expected behaviors. Given the wide range of existing systems, it is likely that fraud detection vendors will incorporate generative AI-based enhancements to their systems rather than enterprises employing tools like the ChatGPT API to develop new fraud detection capabilities.
Photo by Yan Krukau on Pexels.com
Generative AI offers many opportunities, including chatbots, product personalization, and targeted promotions and markdowns. Organizations must monitor and balance their AI-based engagements. Existing solutions already contribute significantly to consumer frustration and dissatisfaction.
If generative AI can do a better job, great, but it appears many firms don’t care about customer frustrations, or they would be addressing current issues.
Photo by Antoni Shkraba on Pexels.com
In Chatbots And Automations Increase Customer Service Frustrations For Consumers At The Holidays, by Chris Westfall for Forbes, he shares that “over 72% of respondents in a recent UJET survey reported that interaction with a chatbot is a “complete waste of time.”” Organizations need to get closer to their customers, not using tools to push them further away.
I leave this one for last as it is the first for most. It is too easy to generate content with generative AI, and therefore, it may prove difficult to check the content for accuracy, in facts or tone. Of course, generative AI will generate content, but it may also demand, over time, the adoption of new editorial practices aimed at keeping the easily generated content relevant, ensuring its adherence to facts, and ensuring that it doesn’t expose the organization to unintended legal or political situations.
Read more at SeriousInsights.net >
Daniel W. Rasmus, the author of Listening to the Future, is a strategist and industry analyst who has helped clients put their future in context. Rasmus uses scenarios to analyze trends in society, technology, economics, the environment, and politics in order to discover implications used to develop and refine products, services and experiences. He leverages this work and methodology for content development, workshops and for professional development.
