The 3 Mindsets Shaping Every AI Decision

As LLMs advance, companies are finding more creative ways to put AI to work. Many companies are introducing AI features such as AI customer support or virtual assistants every other day. This speeds up the adoption of AI, but can leave people with polarized opinions about its actual value. 

Like most major technological shifts, AI has created two loud camps.

Optimists want to push the technology as far as it can go. They love to see AI everywhere because they believe it can automate repetitive work, increase speed, make better decisions, and free people to focus on the work that matters. When a new model, agent, or AI tool promises to save time or unlock new capabilities, they want to try it.

Pessimists are far less convinced. Seeing the abundance of AI output stresses them out because they believe the real cost of AI goes beyond the subscription fee. From buying the tool to ROI you must integrate it, train your staff, monitor its output, and correct its mistakes, etc. Eventually, you may find yourself relying on a black box you don’t fully understand and wondering whether it is making people more productive or just more dependent on it. If you add hallucinations, privacy risks, work slop, and the fear of becoming replaceable, the promised AI future starts to look much less attractive.

However, there is a third perspective: that of the AI realists.

Realists don’t feel the need to try everything, nor do they reject AI right away. They recognize both the potential for success and the risks of failure. What makes them different is that they ask a more practical question: What needs to happen first for AI to work successfully?

This article explores today’s ideas, opportunities, and concerns about AI and what each perspective can teach us about building AI that works reliably.

The AI Optimist POV

AI optimists are as aware as the rest about the problems surrounding AI. They know that as companies grow so does the work, data, and complexity people operate in on a daily basis. They just believe that AI can fix the ever-increasing scale of data so people can operate within it more efficiently. 

Here are some of the ideas and opportunities they see

Vibe coding

“Vibe coding” is when a user prompts AI to write code for them, rather than write it themselves. Most AI optimists welcome this reality of low- or no-code because it can produce any idea faster and cheaper than ever. They believe you no longer need a team of programmers to build an app – just a couple of AI tool subscriptions, and a good idea will suffice.

This turns coding into an iterative cycle of prompting, evaluating, and editing, with developers increasingly shifting their attention from writing every line themselves toward defining intent and orchestrating the development process. 

This can be beneficial even if you have developers nearby. For example,  non-technical managers can prototype their ideas directly using AI, rather than spending time persuading colleagues and potentially losing sight of their original creative spirit. 

AI-driven automation

The optimists are not satisfied with AI only assisting occasionally. They want it to manage entire workflows and automate systems. If AI can understand what needs to happen next, use the right tools, and complete work with less human oversight, processes can move much faster.

That is the promise of autonomous workflows. Instead of following fixed rules, AI agents can presumably work toward a goal across multiple steps, adapt to changing situations, and interact with different systems or data sources.

For example, fast-food companies in the US have introduced AI drive-throughs to make ordering faster and reduce repetitive work for employees. On paper, it makes sense because the process is predictable – customers order, AI processes the request, and staff focus on preparing the food. In reality customers have deliberately tested the systems with absurd requests, including an attempt to order 18,000 cups of water at Taco Bell, while others have become frustrated when AI repeatedly misunderstands relatively simple orders. The idea of automating the workflow still makes sense, but the challenge is that real customers do not always behave like the workflow expects.

24/7 availability

AI can work 24/7 while people need to eat, sleep, and move to function well. Naturally, people cannot work efficiently around the clock. Most people finish work around 5 or 6 p.m., which is also when many businesses close. But those people may still need help with shopping, bookings, or customer support after work. That often means waiting until the weekend or using their lunch break to sort things out. Add different time zones, and it gets even more complicated. 

Klarna offers one of the clearest examples of why optimists get excited. Its AI customer-service assistant operates across 23 markets, in more than 35 languages, 24/7. In its first month, it handled 2.3 million conversations—around two-thirds of Klarna’s customer-service chats—and reduced average resolution time from 11 minutes to under two. Klarna estimated the system would contribute $40 million in profit improvement in 2024.

Disclaimer: While these initial numbers look impressive, Klarna later admitted that relying too heavily on AI hurt its service quality. In 2025, the company shifted back to human workers because the AI struggled to resolve complex customer problems. The AI remained in place, but it turned out that a faster response time didn’t always mean the issue was actually fixed.

The AI Pessimist POV

Pessimists are concerned about worst-case scenarios, especially when AI makes irreversible decisions. They are afraid that AI will create more problems and complexities for ordinary people. They also notice that when people trust AI too quickly, those users can become complacent about its limitations and less vigilant about security and data privacy.

Here are some of their concerns.

Hallucinations

Pessimists worry that AI can be most dangerous when it sounds convincing while being wrong. A clear, confident answer can make false information harder to notice, especially when people trust the system without checking its sources.

This is the problem with hallucinations in large language models (LLMs). LLMs generate responses by predicting what comes next based on statistical patterns. They don’t automatically check every statement against a database of verified facts. When information is missing, rare, or ambiguous, a model can produce something plausible but incorrect, including false facts, invented studies, or fake citations. In enterprise settings, those mistakes can become especially costly when they affect legal, financial, compliance, or operational decisions.

A well-known example is the 2023 Mata v. Avianca case. A legal team submitted a brief containing six nonexistent judicial decisions generated during AI-assisted research. The citations looked convincing, but the cases did not exist — showing why fluent output should never be mistaken for verified information.

Productivity issues

Pessimists question whether generating work faster always means becoming more productive. AI can produce reports, presentations, code, and research in minutes, but someone still has to decide whether the output is accurate enough to use.

This may change the focus of work from creation to verification. In a Microsoft Research survey conducted in 2025, 319 knowledge workers reported 936 instances of generative AI being used in the workplace. Researchers discovered that AI altered the context in which people used critical thinking based on these self-reported experiences. Workers were more required to validate AI outputs, incorporate them into their work, and ensure that the outcome was appropriate rather than expending as much effort creating knowledge.

For pessimists, this creates a productivity problem. The people best equipped to catch subtle mistakes are often subject matter experts whose time is already valuable. AI may therefore make the first draft dramatically faster while moving part of the workload downstream to checking, correcting, and approving what it produced.

The loss of agency

A deeper concern for pessimists is that people may become dependent on AI to produce work they no longer fully understand.

Vibe coding is an obvious example. AI can help someone build software much faster, even if they do not understand every line of code. That is useful while everything works. But when something breaks, requirements change, or the original developer leaves, someone still needs to understand what was built and why.

The same concern applies to human thinking. A 2025 MIT study found lower neural connectivity among participants writing essays with an LLM compared with those writing without digital tools. The study was small and focused only on essay writing, so the results should not be generalized too far. 

However, pessimists are concerned that this goes beyond just coding or writing. If AI researches the topic, structures the argument, writes the report, and makes the recommendation, people can gradually lose ownership of the reasoning behind their own work. That becomes obvious the moment they have to explain, defend, or adapt that work without the AI beside them.

The AI Realist POV

AI realists understand that AI is neither good nor bad but a reflection of the people using it. 

When people are prudent and meticulous about how they query AI, they can get better results. For enterprises, that same level of care means putting the right processes and context in place before AI can be used effectively. It’s not realistic to assume that AI tools will know an organization’s terminology, business rules, or data by default.

Realists therefore focus on three things:

• Giving AI Domain-Specific Knowledge
• Keeping human-in-the-loop where it matters the most
• Building a foundation that allows both to work at scale

Giving AI domain-specific knowledge

Generic LLMs are trained on billions of data points. However, they don’t automatically know about your company. They don’t know what your internal acronyms mean, how any of your products relate, or which policies apply in different situations, and so on.

Realists therefore ground AI in domain-specific knowledge instead of expecting it to get everything correct without it. That’s done by using a knowledge graph that consists of taxonomies and ontologies. Taxonomies define the language and concepts used by the organization, while ontologies go further by defining how those concepts relate and which business rules connect them. 

This approach obliges AI to first check the knowledge graph before using its training data. If done correctly it AI can reach 90 to 100% accuracy while optimizing token usage.

Keeping human-in-the-loop

Realists also accept something optimists want to avoid – not everything should be fully autonomous.

A human-in-the-loop approach keeps people involved at the points where judgment, expertise, or accountability is crucial. That doesn’t mean a subject matter expert has to manually approve every AI response. The consequence of getting something wrong determines where organizations will deploy human judgement.

Radiology provides a good example. In a real-world use case, AI analyzed chest X-rays and flagged cases that might contain a pneumothorax (a medical condition where air leaks into the space between the lung and the chest wall) so those cases could be reviewed sooner. Radiologists still made the final diagnosis, but the AI helped prioritize their workload. For confirmed cases, median reporting time dropped by 46%, from 186 to 100 minutes.

Building a foundation

Realists know that a strong foundation is what makes AI scalable. If company data is fragmented across databases, departments, and applications, adding an LLM on top can’t solve the problem. AI will simply have the same fragmented view. That’s why they build the foundation that makes the organization’s knowledge connected, consistent, and understandable. 

Once the business language and relationships are defined in a knowledge graph, the next step is to connect them to the company’s actual data. Documents, databases, reports, product information, and other sources can be enriched with semantic metadata and linked through a knowledge graph. The point is to create a connected semantic layer that explains what everything means and how it relates.

Together these elements form a Semantic Backbone. Applications can use it through GraphRAG to retrieve trusted facts and the context around them before generating an answer. Instead of building a new knowledge layer for every chatbot or agent, the company has one reusable foundation that different AI applications can work from.

That is how AI realists scale AI, by first organizing what the company knows, then letting AI use it.

Finding the balance

To wrap up, AI doesn’t have to be viewed through only one of these perspectives, but each can teach us a lesson. AI optimists remind us what becomes possible when we push the technology forward. AI pessimists remind us to question the risks and hidden costs. And AI realists ask what needs to be in place to make those possibilities work reliably.

In practice, successful AI probably needs a little of all three: ambition to experiment, skepticism to ask the uncomfortable questions, and the right knowledge, infrastructure, and human oversight to make it work.

And if you recognize yourself—or someone you work with—in one of these three personalities, we’ve created something for you. 

Explore our free booklet for a lighter and slightly less serious take on AI optimists, pessimists, and realists