AI: Coworker or Competitor?

Balancing AI Autonomy and Human Oversight

Imagine this: You've made it to the front of the line at the coffee shop, and with a laptop in one hand and drink in the other, you head over to the only seat available. As you settle your things and log in, you notice seated across the table is "Neurobot," a sleek, silver-haired AI with an affinity for algorithms. It churns out content faster than a caffeinated squirrel on a keyboard. Your marketing agency, once the reigning content czar, now faces a formidable challenge.

Neurobot’s digital ink flows effortlessly, creating blog posts, social media updates, and catchy slogans—all in the blink of a digital eye. Revenue slips away like sand through your fingers. How do you stand out when Neurobot’s content is as ubiquitous as cat memes? How do you differentiate yourself when your competitor's AI whizzes through proposals faster than you can say "machine learning"?

What is Artificial Intelligence (AI) vs all the other software we have used in the past?

What is AI exactly? At TechHouse, we often say that you can't use pronouns like "it" or "them" when trying to solve a problem because the problem can be hidden in the vague term. Similarly, we need to get good, clear terms if we're going to be able to navigate a world where AI can both be a great partner and a significant competitor.

Over the past 50-plus years, software has automated work. With a defined set of instructions, business software applications calculate numbers, generate reports, and store information for later access. We became accustomed to automated billing, payroll, and sales funnels.

Before AI existed, software did what humans told it to do and did it faster than a human. The instructions may have been complex, but they were instructions. We told the computer what to do, and it did it. We tested the software to ensure that the instructions provided were not flawed, so that we could, in turn, ensure the behavior of the software. We could trust the software to perform functions, such as calculating math problems, determining a financial statement, and even in medical devices to monitor heartbeats, blood pressure, and more. Trust in these systems was due to rigorous testing of clearly defined instructions.

AI software is different. AI does not carry out tasks with predefined rules. Instead, AI has a model, and humans train that model to make decisions. There are many different models, each suited for different types of decisions. The better suited the model is to the decision, and the more limited in scope and precise the testing, the more accurate and less probabilistic the answers may be.

AI is a broad term referring to many different specialized technologies. Like in medicine, the AI industry has different specializations to solve different problems. Although a podiatrist is a medical doctor, I would not ask them to cure my migraines. (Though, at this point, I may try that, too!) Similarly, there are many different types of AI.

For example, Optical Character Recognition (OCR) software is trained to scan documents for key information and then store that information. In an accounting system, OCR can scan invoices, identify the invoice number, date, and amount, and then enter that information into the invoice record. How does it accomplish this? Someone trained the OCR Software on what those invoices look like and where to find the invoice number, date, and amount.

Unlike traditional software, which works from a defined set of rules, AI infers the answer from prior learning. It makes its best guess as to the right answer rather than following instructions to create an answer. To get the best guess, AI models data. AI models work from tremendously large datasets, such as the foundational models built on 40 years of digitized information from the Internet.

AI mimics human intelligence's ability to learn from experience and adjust to new inputs. How is AI different from the accounting, spreadsheet, and graphic design software we have all used in the past?

AI that takes us from automation to decision-making, from tool to potential collaborator or competitor.

Let's explore how this affects our understanding of AI and how we plan our business's response.

Testing. Testing. 123 Testing

Two key AI terms are training models and neural networks. A training model, well, you can train it and test the results. For example, with the OCR software we mentioned above, if you receive many documents from your vendors that you need to enter into your accounting software, you can train AI to read them. You would train the AI on each vendor’s invoice layout, the account number, invoice date, amount due. When you scan similar invoices in the future, the AI model takes what it learned, and accurately pulls the invoice data off the invoice and enters it into your accounting system. After you scan in all your invoices, you would check to be sure that the information was read correctly. That there were not any anomalies.

System testing is creating test criteria, testing against it, and comparing the results to what I expected.

If you need to understand whether you have a positive or negative profit on a project, you could highlight your profit number with color to stand out. If the profit is positive, make the numbers green; if negative, red.?

If you copy the sample code below into a simple notepad text file and call it test.html, you can open that file with a browser and see a red negative five on the screen. These are specific instructions written in a language the software understands to tell the software to do something. The software follows the rules. You can open the file to run that software and confirm it functions as expected.

<!DOCTYPE html>

<html>

<body>

<h2>Number Color Indicator</h2>

<p id="demo" style="color: green;"></p>

<script>

var num = -5; // Change this to any number you want

document.getElementById("demo").style.color = num < 0 ? "red": "green";

document.getElementById("demo").innerHTML = num;

</script>

</body>

</html>

You could change the number from -5 to 5 and see that the number is green when you open the newly saved version. The HTML language is comprised of standard commands. You tested the software to ensure it follows the rules. The more complex the rules and the bigger the program, the more difficult it is to test thoroughly.

Many of us learned about testing hypotheses years ago in school. In the technology industry, system testing is testing the computer or software system. We define what we expect the results to be before running the test and then compare the system actions to see if they meet what we expect. Did the software put the invoice number in the invoice field and the product number in the product field? This testing process works when we have precise test data with clear conditions we are testing.

With system testing, we trust the software we use daily. When something is wrong, we submit it to be "fixed." We expect computer systems to provide repeatable, predictable, and accurate results.

It's not uncommon for people to say, "It seems like nobody tested this." They think the software developers did not go through the process of defining test conditions, running the system to see if it produces those expected results, and comparing the actual to expected results to discern if the system is functioning as it needs to. If we ask software a question, we expect an accurate answer: "The computer can't be wrong." "Numbers don't lie."

If the computer system gives a different answer than we expected, the first thought could be, "What did I miss that my answer is different from the computer?" We may question our thinking before questioning the software. "System of Record" refers to the computer system that holds the "truth"-- accurate records you can trust. It is the place you can go and trust the data in that system to be accurate and true.

Traditional software does what humans do but more accurately and faster, a paradigm left over from the industrial age, when automation increased output and machine labor replaced human labor.

But AI is different.

Precision vs Probability. Testing Limits.

In this world of getting the correct answer, there's an assumption that getting correct answers is what computers necessarily do. We have come to see computers as rigid deliverers of objective accuracy. At the end of the day, a computer's answers are founded on "bits." 010101. Either yes or no. It is intuitive to the human mind to believe that with only the options of yes/no, on/off, and 0/1, there is inevitably only correct and incorrect.

None of us would be using accounting software if they advertised that they were correct 90% of the time. We are not interested in probability when determining how much money we have in our bank account to pay our bills or make payroll. We need precise answers.

When someone says AI hallucinated, the founding assumption is that AI would be accurate because it is a computer system. But that is not necessarily the case.?

AI models make decisions based on their training. Like us, there is a probability that the decision will not be correct. Our OCR example above is an example of classical machine learning AI. Deep Learning AI is different.

The chat/prompt AI tools are great examples of Deep Learning AI models. These AI neural networks consume a tremendous volume of information, such as the 40 years of digitized information on the Internet. The AI neural network absorbs that information and uses it to respond to the requests it receives. It "understands" language using "Large Language Models." Notably, the terms used to interact with Chat tools are prompt and response, not question and answer.? The response from the AI neural network is built based on the probability of being correct based on the information it has absorbed.?

Unlike the OCR invoice scenario above, it is doubtful the AI Neural Network in your chat bot has been trained on your specific prompt. Instead, based on all it has absorbed, it will respond with the most probable answer. Although various AI models will be tested and trained over time to increase accuracy, the sheer complexity of a neural network trained on the Internet's vast data stores makes it impossible to test each condition or even a significant portion of them.

It reminds me of a scene from the movie Armageddon:

President: What is this thing?

Truman: It's an asteroid, sir.

President: How big are we talking?

Scientist: Sir, our best estimate is 97.6 billion…

Truman: It's the size of Texas, Mr. President.

President: Dan, we didn't see this thing coming?

Truman: Well, our object collision budget's about a million dollars a year. That allows us to track about 3% of the sky, and begging your pardon sir, but it's a big-ass sky.

President: Is this, going to hit us?

Truman: We're efforting that as we speak sir.

President: What kind of damage?

Truman: Damage? A total, sir. It's what we call a global killer. The end of mankind. Doesn't matter where it hits, nothing would survive, not even bacteria.

President: My God. What do we do?

A trained model for a specific task will be precise and accurate in a way that the neural network with a chat (large language model) prompt and response cannot. An AI Neural Network absorbs a tremendous volume of information, and then it will use probability based on what it's read and what it thinks might be an answer to whatever question you have. The critical distinction here is that the prompt and response are likely not system tested. It is a big system and testing would have a limited budget.?

So why would anyone use software that does not feature accuracy as its main feature? When working with Deep Learning AI, we must consider probability and likelihood. Is a Deep Learning foundational model without specific training on my prompt the right tool to decide how to handle an employee issue? How to write an email, even? No.

When is Deep Learning best?

When do we think about neural network prompts and responses in our business? When we can improve our capabilities and augment human capabilities. Rather than automation, it's collaboration. ?

AI as Collaborator: Guardrails, Ethics and Responsibilities

Having an AI model available for prompt and response interaction is like having another team member to brainstorm with. With more perspectives, you can increase the probability of solving the problem, especially if that second person brings specific knowledge and understanding.

Like the digitized, non-tested data from the internet which informs the model to make the decision, the AI model could provide flawed decisions and guidance.

With humans, we assume some of the information we receive is flawed. We assume that the data the person is working from could be better and that their problem-solving ability and decision-making processes could be better.

We know that humans are not system-tested. What they're saying may or may not be accurate, so we combine our experience with our five senses of what we see in the real world with what the human is saying and try to discern what information to proceed with. Is what they're saying entirely correct? Partially, correct? ?

In the past, we could assume that if a computer provided an answer, it was accurate, or the programmers did not test it sufficiently, and it needed to be fixed. The bar for well-run software was accuracy. We would even use a computer to test a human's work, such as solving a complex math problem.?

In this new world we need to be aware that Chatbots which are built on Deep-Learning AI and Large Language Models use probability to gauge an answer. Unlike traditional software, and even unlike Classical Machine Learning like our OCR invoice above, the Chatbots with Deep Learning AI are not system tested. And yet, we find many people are thinking of these tools as if they are. We often hear our customers asking a Chatbot a question or asking a Chatbot to create content and then using that content as is, as though the results have been system tested somewhere for accuracy. As though if they just ask the question in the right way, they will get an answer. When in fact, it will always be a prompt with a response. A response based on probability not system-tested accuracy.?

I recently spoke with Data Privacy Architect Swati Popuri on this topic during our Women in Cloud Bright Talk session last week, “Algorithmic Overlords: Balancing AI Autonomy and Human Oversight.” Swati walked us through the concept of a “Consequential Decision.”

A Consequential Decision has a significant impact. It’s a decision we need to be sure we get right. And, if the answers are potentially flawed, then we need human oversight.

This brings us to another critical area for our strategy – Ethics, responsibility, and potential liability.

If the model is trained on digitized internet data, what biases and errors are reflected in that information? How will it help or harm our business and decisions if we rely on it for responses derived from biased information? What happens to decisions that would shift given information discovered by science or elsewhere in 2024 if the foundational model was last updated in 2022?

How about legal risks? In August 2023, the Equal Opportunity Commission settled its first AI hiring discrimination lawsuit. Three companies violated the 1967 Age Discrimination Act because the AI hiring program "automatically reject[ed] female applicants age 55 or older and male applicants age 60 or over." There is ongoing litigation for a class action lawsuit involving Workday. Amazon stopped using its AI hiring tool, because having been trained on a database of primarily male applicants, it preferred resumes that used words that are more commonly used by men in their resumes like "executed" and "captured." Navigating the AI Employment Bias Maze: Legal Compliance Guidelines and Strategies ( americanbar.org )

Most AI and Standards organizations have identified the necessity of trustworthy and responsible AI. The National Institute of Standards and Technology (NIST.gov ), the ones who decide an inch is an inch, have defined these key "building blocks" Trustworthy and Responsible AI | NIST :

• Validity and Reliability
• Safety
• Security and Resiliency
• Accountability and Transparency
• Explainability and Interpretability ?
• Privacy
• Fairness with?Mitigation of Harmful Bias

Microsoft requires that all its customers commit to responsible and ethical AI. Their landing page can be found here: Empowering responsible AI practices | Microsoft AI . Their 2025 Responsible AI (RAI) Transparency Report can be found here: Responsible AI Transparency Report ( microsoft.com )

So, if this sounds like a lot, can I just ignore AI for now?

One of our customers asked if they could just create a "No AI Allowed" policy and thereby avoid the AI challenges until everything is settled. Unfortunately, that could be quite hard to implement. Microsoft and LinkedIn published the results of their study on AI in the workplace in May 2024. In general, AI is in use by 75% of workers worldwide. AI at Work Is Here. Now Comes the Hard Part ( microsoft.com )

How could that be? There are numerous free and low-cost browser extensions that utilize AI. Someone recently shared with me a list of over 1,000 programs available with AI. How common is AI? Even very common tools like Google search and Edge Bing use AI, as do Word, Excel, QuickBooks and more.?It is unlikely a no-AI policy is possible at this stage.

And so, from here? Now what?

How do we proceed with our business in a universe that includes not only AI-trained models but also foundational models with probabilistic responses to our prompts?

Changes in thought and understanding can rattle more than our cage. It can shift the ground beneath our feet and upset our understanding of the world order. Does it change that order? Or just our understanding of it?

In the 1970s and 1980s, software programming combined with powerful machines to automate assembly lines and contributed to the rust belt. Job loss is a genuine concern, and creating new jobs is a real opportunity.

How do we plan for a future that appears to be such a deviation from the past that we can feel like we are on a completely different planet?

I do not believe the answer to that question is known. We don't have a finite scenario we can test to ensure an accurate outcome. But we deal with similar scenarios in our business every day. Each interaction with an employee, customer, client, or board member requires decisions that are probabilistic, not certain.

We gather the facts of the situation, arm our team with the tools we can, and make our best guess, then adjust as needed to drive forward.

Gather the facts of the situation.

Like other Macro Trends, understanding the fundamental principles of AI is essential. Identify a trusted advisor, preferably a few who can provide perspective.

Prepare your team

Ensure your team has critical thinking skills. Asking the right prompt increases the likelihood of getting a better response.?

Setup the necessary technical platform

AI eases decision-making. However, it requires a vast amount of data to do it well. Ensure you have a data governance platform in place.

Make your best guess.

Consider AI in your business, your industry, among your customers, and in your community. What will change? How will it change? Consider 3 outcomes and prepare a rough plan should they arise. Revisit this plan as you would any other business plan, only with the fast rate of change, a monthly or quarterly revisit may be required rather than annual.

Track and Adjust

More than traditional software, AI models are constantly evolving, as are our adaptation to a world with AI. To understand where we are, we need to know where we have been. Identifying measures for performance and risk and continually monitoring those to respond is needed. Additionally, create feedback loops from your team to ensure you benefit from the variety of perspectives and experiences as quickly as possible. There are many tools available. A good start is Microsoft's RAI Impact assessment template: Microsoft-RAI-Impact-Assessment-Template.pdf

Some ways TechHouse can help

• Webinars: Check our “Algorithmic Overlords: Balancing AI Autonomy and Human Oversight” recording on Bright Talk. We also hold regular webinars on AI, Data and Cybersecurity topics
• Adoption: Our Aware for Copilot Solution on Azure Marketplace contains curated assessments, sample policies, communications, and guides for your AI Adoption journey. Now available on Azure Marketplace!
• Training and mentoring for you and your team: From Cybersecurity to critical thinking workshops, our team is dedicated to transferring skills to help your team thrive in this new world.
• Technical Preparedness: Engage us for an AI preparation, data governance, cybersecurity, or Copilot/AI rollout in your organization.