Leadership · 5 minute read
LLMs Explained for Executives
A large language model is software trained on vast amounts of text to generate language, which makes it capable of reading, writing, summarizing, classifying, and reasoning over text and code. It is a general reasoning engine, not a database: it does not know your business, can be confidently wrong, and is priced by text processed.
Every AI agent, assistant, and copilot runs on a large language model, and every executive is now expected to have a view on them. This explainer gives the working understanding a leader needs: what LLMs are, what they are genuinely good and bad at, why they make things up, how cost works, and which decisions actually matter.
What is a large language model?
A large language model is software trained on a vast amount of text to predict the next fragment of a sequence. That training objective, applied at enormous scale, produces a system that can read, write, summarize, translate, extract, classify, and reason through problems expressed in language, including code.
Two properties follow that executives should hold onto. First, an LLM is a general engine: it can work on almost any language task without being built for it. Second, it is not a database: it does not contain your customer records, your policies, or your contracts. It knows your business only through the information placed in front of it for each task, which engineers call context. FISTA's glossary entry what is an LLM has the technical definition; this piece is about implications.
What are LLMs good and bad at?
| Strong | Weak |
|---|---|
| Reading and understanding unstructured text | Exact recall of facts not in the context |
| Drafting and rewriting in a specified style | Arithmetic and calculation at scale |
| Summarizing long documents | Knowing the limits of its own knowledge |
| Extracting structured data from documents | Consistency across runs without constraints |
| Classifying and routing | Tasks requiring current information it was not given |
| Multi-step reasoning through described problems | Judgment about your business without your policies |
The pattern is that LLMs excel at language and reasoning and fail at recall and precision. Well-built systems compensate: they give the model the right documents (retrieval), hand calculations to tools, constrain outputs to formats, and test the results. The model is one component; the system around it does the rest.
Why do LLMs hallucinate, and what manages it?
An LLM generates plausible language. When the fact it needs is in its context, it usually reports it accurately. When it is not, the model may produce a fluent, confident answer anyway. This is a property of how the technology works, not a defect to be patched away.
Three practices manage it: grounding (retrieving the right documents into context before the model answers), constraints (requiring citations, allowing "I don't know," restricting outputs to approved sources), and evaluation (measuring how often the system is wrong on real cases, before release and on a schedule). The AI hallucinations explained for executives piece covers this in depth.
How does cost work?
LLMs are priced by tokens, which are fragments of words, counted on the input sent and the output produced. More capable models cost more per token. Two trends matter for budgeting: per-token prices for comparable capability have fallen repeatedly, and total usage tends to rise as agents take on more work, so total spend can grow while unit cost falls. Well-run programs track cost per task and manage it by routing routine work to cheaper models and caching repeated content. The LLM token cost explained guide gives the arithmetic.
What decisions actually matter?
Model selection attracts attention, but it is rarely decisive. The decisions that determine outcomes:
- What work to apply LLMs to. Language-heavy, high-volume tasks with a measurable baseline.
- How to ground them. Connecting the model to correct, current company information is most of the engineering.
- How to test them. An evaluation set of real cases, run before every change.
- How to constrain them. Output formats, approved sources, permissions, and escalation.
- How to keep the model replaceable. A gateway that lets the company switch models as capability and pricing change.
The how to choose an LLM for enterprise agents guide covers selection; the LLM vendor lock-in guide covers replaceability.
How do LLMs fit into agents and assistants?
An assistant puts an LLM behind a chat interface with your documents retrieved into context; the person acts on the answer. An agent gives the same model tools and a loop so it can act on the answer itself, under permissions and approval gates. The model is the same in both; what changes is how much authority the system around it grants. That distinction is why an "LLM strategy" is really a decision about grounding, tools, and controls, which the how AI agents work explainer walks through part by part.
What should executives ask?
- What tasks is this model proven on, in our evaluation set?
- Where does it get our information, and how current is it?
- How often is it wrong, and how do we know?
- What is the cost per task, and how does it trend with volume?
- How quickly could we switch models if we needed to?
How can FISTA Solutions help?
FISTA Solutions builds systems around LLMs, through its AI enablement practice and its AI agents practice, with grounding, evaluation, constraints, and model replaceability designed in, so the strengths of the model reach the business and the weaknesses do not. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries.
If you want a briefing on where LLMs fit your business and what the evidence should look like, talk to FISTA on WhatsApp, or continue with RAG explained for executives.
Share-ready article cover
Download the generated social format.
Clear answers
Questions raised by this field note.
Straightforward guidance for evaluating scope, fit, and the next step.
01What is a large language model in simple terms?
Software trained on enormous amounts of text to predict what comes next in a sequence, which turns out to give it the ability to read, write, summarize, translate, extract, classify, and reason through problems expressed in language. It is a general engine for language work, made useful for a business by connecting it to company data and tools.
02What are LLMs good at and bad at?
Good at: understanding unstructured text, drafting, summarizing, extracting structured data from documents, classifying, and reasoning through multi-step problems. Bad at: exact recall of facts not in front of them, arithmetic at scale, knowing what they do not know, and consistency without constraints. Systems around the model compensate for the weaknesses.
03Why do LLMs make things up?
Because they generate plausible language rather than retrieve verified facts. When the information is not in the context they have, they may produce a fluent, confident answer anyway. The remedy is grounding (giving them the right documents), constraints (requiring citations or abstention), and evaluation that measures how often they are wrong.
04How are LLMs priced?
By tokens, which are fragments of words, counted on both the input sent to the model and the output it produces. Larger and more capable models cost more per token. Prices for comparable capability have fallen repeatedly, but total spend often rises as usage grows, so companies manage cost per task with routing and caching.
05Which LLM should a company choose?
The one that performs best on your evaluation set for your tasks at an acceptable cost and under acceptable terms, and the architecture should keep it replaceable. Most companies use more than one: stronger models for hard tasks, cheaper ones for routine work. Model choice matters less than grounding, evaluation, and controls.
Continue exploring
Related capabilities
Start with the hard problem
Need the outcome owned, not merely analyzed?
Tell us where delivery is constrained. We’ll map the fastest credible path from intent to verified production.