Top 5 AI Concerns and How to Mitigate Them Responsibly

Before I get started on the blockers, I want to make one thing clear that is a critical pre-supposition to all AI efforts: an AI effort needs to support your organizational strategy and be aligned with your mission. If the “why” of your AI efforts is not aligned to your business strategy and your executive team, don’t even start. This is the number on reason I insist on having executive coaching sessions on AI. I want the executive team to understand and not underestimate the critical impact that AI will have on their organization. I’ve found that most orgs are over-estimating some aspects of AI, get let down by a simple feature, then underestimate the long-term impact they need to prepare for.

With that said… now to the blockers. The blockers for AI efforts tend to follow the same patterns from organization to organization.  Let’s talk about those blockers and how to clear them responsibly.

The first blocker is around the general concern of data privacy.  Companies are concerned about data exfiltration and have a concern that their data will become everyone’s data.  Even executive teams have heard about “don’t put your data in ChatGPT”.  The problem is that often a healthy alternative hasn’t been provided.   The good news is that the first blocker is one of the easier to resolve.  The first thing I communicate to companies is that everything we’re going to talk about in the AI space is using a private instance of their data and that their data will remain their data.  This is true for Copilot scenarios, Copilot Studio, Azure Open AI, or other ML models that a company might build. The most successful companies will create these alternatives to free ChatGPT-like platforms that are built to sell your data for advertising or data mining.  A few examples of how to take ground here and ensure your AI data is private, not public:

• Using Microsoft Copilot for true ChatGPT replacement
• Using Microsoft M365 Copilot for Office 365 productivity assistant
• Building [Company GPT] for internal data with high quality responses
• Building a Customer Service chatbot for answering customer questions
• Building an external product manual look-up

All of these are examples of AI using private instances and should not be confused with the free ChatGPT or other advertising-based chatbots. You should NOT be allowing your employees to use free instances of AI tools or those that pay for their product via advertising, especially through search. You’ll quickly find those requests showing up in ads or other retrained models later.

The second blocker is data readiness.  This is often used as a “get-out-of-jail-free” card for tech executives that want to kick the can down the road.  The reality is much more nuanced, since data readiness is a use-case specific problem, not a generic problem. Don’t allow your tech organization to take a “if we build it they will come” approach to data. Instead, ensure that a data architecture exists that data sources and outcomes can be mapped into with a scalable governance approach.  In many instances, the data necessary to address key needs for the organization already exists.  In others, the moon-shot idea might be based on having X data ready, but as a result of understanding the idea, the path to getting the data ready is more easily understood and prioritized.  A couple examples of data readiness:

• Customer Service chatbots are mostly dependent on previously answered questions or documents which include the answers.  In many cases the customer support or technical documents are sufficient to answer many of the customer questions in a manner that “shifts left” a good percentage of customer queries.  Success does not need to mean 100%.

• [Company]GPT solutions for HR, onboarding, policies, support, general questions, etc. are rarely dependent on data readiness.  In most cases the employees just don’t know where to go for the answers.

• Supply Chain Forecasting is an example of a use case with significant data readiness needs, especially understanding the correct data elements, influences on demand and inventory, as well as external factors.  The most important data here is historical sales and historical influences.  The company-specific nature of these models is why most out-of-box supply chain forecasting is challenged to arrive at sufficient answers.

The summary of data readiness is… it might be a problem, it might not… but you only know if you are thinking about the end-in-mind vs. data readiness as a monolith. Think about your data critically and with the necessary focus on usage, ethics, and quality in the context of the outcome you are trying to drive. In some cases data readiness is iterative. For example, you might start an internal chatbot with some high quality data on one use case, then add other use cases as the data is ready. In this case it is more about setting clear expectations with the customer than having all data ready at once.

The third blocker is a concern around human displacement.  The goal here is to turn the topic around into an opportunity or even a responsibility.  Like it or not, human displacement as a result of AI is coming for the job market.  The opportunity or responsibility we have is to prepare the people under our care for the changes and to enable them with forward-looking skills in our organizations.  The challenge in previous eras of human history is that retraining was under invested and the impacted humans were hit hard by the change.  The result of transitions like the industrial revolution, transportation, the internet, and smartphones was a tremendous difference between “haves and have nots” in the context of talent and place within the new economy.  The AI economy will transition jobs in a similar fashion, with some skills being picked up intuitively and others needing serious personal investment.  The challenge is that similar to other transitions, many will be born into or make the jump, whereas others will need to relearn skills they had before or never had.  It’s critical that every person have the opportunity to participate in the AI economy.

Note that there are a variety of possible jobs where upskilling might happen. The largest (and noted by the above) is AI Practitioners… people who can use, not build AI. The lesser roles, but still important are those that prepare data, build models, or combine the lego-blocks together to create outcomes. You can see these in the image below.

I also want to remind an important point here… this is exactly why AI is an executive issue, not an IT issue. AI will change every role in your organization and those that ignore that will have negative results, both for the employees and the organization itself. Do not underestimate the change this will have on your organization.

The fourth blocker is the concern of Quality and Hallucinations.  This is a very interesting blocker because it forms a general mistrust of AI solutions. The reality I’ve experienced is that most companies are terrible at monitoring the human processes we seek to augment.  They don’t know TODAY if they are answering questions correctly or if they are improving over time.  For example, most companies I talk with have no idea if the Demand and Inventory Forecast is accurate or not.  They don’t know if Customer Service reps are answering questions correctly or just getting off the call.  They don’t know if they are setting accurate expectations with customers on order delivery.  So… the first step is to monitor it, measure it, and manage it.  This is a natural component of any AI project because we want to prove that value was created as a part of the engagement, but also we need to gauge the success vs. the historical methods.

The biggest concern on quality comes from the idea of hallucinations, which are essentially answers provided from LLMs that are “made up”.  Similarly with the question in general, the reality of the situation is much more nuanced.  The AI agent is just answering a question the best it can based on the data that is most likely to be correct.  Interestingly, once an AI agent incorrectly answers a question, the likelihood of continuing down the wrong path is high, similar to a human who is building on an already-wrong answer.

So, what to do about it?  This is a major concern for sure. The good news is that mitigations to this concern are coming a long way.  Hallucination modules are able to be built into chat experiences that are augmented by the need to reference source data where answers were provided from.  There are good ways to mitigate directly incorrect answers by preparing the model prompt with a system control that applies certain guardrails.  For example, if I don’t know the answer, I’ll generally say, “I don’t know”, or “this might be the answer”.  An AI agent can be built to do something similar, or just not answer at all in cases with lower confidence.  This lets us build controls that mitigate it. We’re also seeing ways to fact check original responses for “groundedness” and those controls are getting better all the time. In some cases we need to build specific validations between the LLM and the customer to ensure the responses are factual.

Another concern surrounding quality is tied to security and/or jailbreaking.  This is an area of major investment from the big tech firms since jailbreaking is one of the best known issues with LLMs.  Thankfully great forward progress exists here, where event out-of-box guardrails can be placed around a LLM for security, language, obscene content, and security. Be sure to check out the Microsoft safety capabilities built right into Azure AI Studio. There are huge advancements in the ability to build, secure, and test AI systems at scale for both hallucinations and jailbreaking/ethics.

The fifth blocker is bias and represents potentially the most serious, since it directly disempowers communities.  The impact of AI and bias is about understanding the intended, or unintended consequences of your data or AI solution.  For example, I was building an AI solution that was tied to occupancy rates of large commercial apartments.  Early in the project a concern was quickly surfaced around what data was being used, how the data could be applied, and making sure that communities were treated fairly as a result.  The important focus here was to understand the nuances that go into building a system for that prediction, but also ensuring that critical data was not used, or was not used improperly as a part of the overall solution.  Going to pains to capture flows and controls was important so the system is not a mystery to those using it.

The advice here is to understand the end-in-mind and how data might be used to achieve that goal, but looking carefully to think about where it might be mis-used.  Also, its important to understand how data platforms that exist to serve the model might be used by a completely different and adjacent need.  In some cases the model is built on a dataset that is appropriate for the initial use, but an unsuspecting use case will use it to unknown or tainted results.  This is where data readiness meets generic bronze-silver-gold data environments and the responsibility to manage the dataflow.

Wrapping up, it’s critical that executives and AI builders are aware of these blockers, but also don’t see them as impediments to progress, but rather responsible controls to ensure exist in every solution that is built.