Effective Use Cases for Generative AI: When to Choose AI Over Traditional Tools

I love shiny new tools, but I also love things that actually solve problems. Lately that tension keeps showing up in conversations about generative AI, because everyone assumes the answer to every problem is, sure, throw an LLM at it. Spoiler, that is not always the answer. The real question I keep asking, and the one I pushed on during a long chat with Ben Hoyle, is simple, blunt, and oddly underused, “what are your tasks?” If you cannot answer that, nothing else matters.

When not to use generative AI

This is where most people trip up. We saw teams paste giant Excel sheets into chat windows and ask for analysis. Ben put it bluntly, “If you can do a task in Excel using Excel functions, then jumping to generative AI to solve the same task for you is actually a bad use of that technology.” He was exactly right.

• Use traditional, well-tested tools for deterministic, repeatable tasks. Excel, SQL, or a simple script will be faster, safer, and more auditable.
• If your outcome needs 100 percent reproducibility, don’t use a black box as your primary engine. Large language models can be deterministic if you control seeds and infrastructure, but for most users that level of control is neither available nor practical.

Generative AI is not a sledgehammer for every problem. Sometimes it’s a hammer for some nails, and sometimes you need a precision screwdriver.

When to choose generative AI

There are clear, practical cases where GenAI wins, and wins hard.

• Knowledge retrieval from messy internal docs: If you want to chat with a pile of requirements documents and extract meaning, GenAI excels. Ben described how embedding these models into tools like Microsoft Copilot, inside workflows people already use, suddenly makes “a mass of our data accessible” and useful.
• Orchestration and tool-calling: Use LLMs as the brains that decide which tool to call and when. Ben framed this nicely, “you’re using large language models as an orchestrator or as the brains of an orchestration software.” Let the LLM call a deterministic function, run a reliable calculation, then turn around and explain the result in plain language.
• Prototype and ideation, especially across domains: If you don’t fully know the problem space, GenAI can expand your mental toolkit. People will surprise you with creative approaches, like using image generation to simulate a new wall color for a living room. Inspiration is underrated.

Generative AI moves the needle most when it reduces friction between the human question and the answer, especially when the answer requires synthesizing many sources or translating between vocabularies.

Practical checklist to decide

Here is the litmus test I now use when someone asks, should we use GenAI for this?

1. Is the task inherently deterministic and already well served by existing tools? If yes, do not use GenAI.
2. Does the task require synthesizing across messy, natural language sources, or acting as a translator between human language and software? If yes, GenAI is worth trying.
3. Does the task require access to sensitive data? If yes, stop and build a compliant, internal workflow first.
4. Can you design the system so the model orchestrates calls to deterministic functions or tools, rather than trying to be the deterministic oracle itself? If yes, great.
5. Do you have a plan to pilot with measurable outcomes, then scale carefully? If no, you risk a lot of noise and little value.

These are not theoretical exercises. Ben and his teams ran internal data-hackathons to surface real problems and then pick the best technical approach. That sequence works, because it starts with the problem, not the tool.

Building trust and scaling adoption

This is the political engineering as much as the tech. Two things matter more than clever models.

First, top-level sanction. If a CEO or a senior leader demonstrates an internal GenAI tool, people stop treating it as forbidden, secret, or a black market hack. Ben explained it bluntly, get the highest ranking member to use the system, show a live demo, and then the rest will feel legitimate doing it too.

Second, grassroots access and training. Showcase internal examples where someone in a similar role used the tool successfully. That closes the translation gap between “knowledge retrieval” as a concept and “this is how it helps me write a spec.”

And yes, you need guidelines, not just hype. Employees want to know what data they can safely put into a model, and how to phrase requests to get useful outputs. Prompt craft used to be everything, now it’s less critical, but teaching people to use models as sparring partners is essential.

About determinism and reality

A frequent objection is randomness, and it’s valid for most users. Models appear non deterministic because of changing seeds, different servers, evolving models, and context drift. Ben pointed out you can host a model locally, set the seed, and achieve deterministic behavior. For most businesses that means you need infrastructure and governance if determinism matters.

A practical compromise is to design hybrid workflows, where the LLM does the creative or synthesizing work, and deterministic code handles exact calculations and compliance.

Final thought

Generative AI is incredibly capable, but capability is not a mandate for universal use. The right move is less about choosing AI over traditional tools in absolute terms, and more about choosing the right combination for the job. Start from the problem, not the sparkly model. Use GenAI where it reduces friction, enhances synthesis, or acts as an orchestrator. Use Excel and tried and tested functions where you need precision and auditability. Build trust from the top, give people safe ways to explore, and let the technology slot into workflows, not replace them by decree.

If you do that, you get the upside without the chaos. And honestly, that sounds like a good day at work.