Large language models in business: what works, what fails

Large language models, the AI behind tools like ChatGPT, Claude and Gemini, are overhyped until they’re not. This post covers where they still fall short in a business, where they deliver real results, and how to put them into production safely.
Nearly every company now has an AI initiative, or at least a slide about one in the board deck. These models can write code and automate parts of a workflow. They can also state something false with total confidence. The difference comes down to where you use them and how.
Where large language models fall short
1. General-purpose chatbots
A chatbot that answers anything, with nothing to ground it, is still unreliable. It might do well in a demonstration. In front of customers or employees, it can:
- make things up and present them as fact,
- give different answers to the same question,
- create regulatory risk, especially in finance, legal and health care.
Unless the model answers only from your approved sources, a general chatbot does more harm than good. The companies I’ve seen succeed use these models internally first, long before customers ever see them.
2. Making decisions on their own
A model can lay out options and summarize the tradeoffs. Handing it a sensitive decision, like a financial approval or a legal assessment, with no person checking it, is a fast path to a costly mistake. That’s why Bennett ONE sends any action that moves money, changes a record or reaches a customer to a person on your team for approval.
3. Complex code
AI coding assistants do make developers faster. For core systems and code that connects many services, the generated code still needs careful review, testing and rework. Otherwise small bugs turn into expensive problems later.
Where they deliver
1. Answers from your own documents
The most useful setup I’ve seen pairs a language model with a search system over your own documents. The system first finds the relevant passages, then the model writes an answer from those passages only, with citations. This approach is called retrieval-augmented generation.
It works well for:
- internal assistants for legal, compliance, human resources and operations,
- searching and summarizing procedures, contracts and policies,
- answers that cite their sources, so anyone can check them,
- plain-English questions across documents and databases.
2. Sorting and routing requests
Service desks, compliance teams and support teams receive messages that are messy and unprioritized. Language models can classify each one by intent and urgency, route it to the right team, suggest a response, and pull out details like dates, names and product codes. People spend less time sorting and more time solving.
3. Assistants for specific teams
Internal tools are a safer place to start than anything customer-facing. Good examples include legal teams searching their clause library, marketing teams summarizing campaign results, product teams reviewing feedback, and finance teams asking about forecast scenarios. With access controls and answers grounded in your data, these assistants become tools people use every day.
How to deploy them safely
A project succeeds on careful design and a clear business goal. On every system we put into production, we do the following.
Security and compliance first. Each user and each agent can see only the data it’s allowed to see, and every step is logged.
Answers checked against your sources. We set up search over your documents, guide the model’s tone and format, and test every output against criteria we agree with your team.
Measurement and feedback. People review results early on, quality is scored automatically, and the system improves from real use.
That’s the job of Bennett ONE: it checks every answer, routes approvals to your team, and records every step with its source.
Final word
Language models won’t replace your judgment, your team or your systems. In the right setup, they make all three faster.
If you’re deciding where a language model could help your business, the first call with one of our senior engineers is free and takes 30 minutes.