You've used one. Maybe you opened ChatGPT to draft a proposal on a Friday afternoon when the words weren't coming. Maybe a colleague showed you how Claude summarized a 40-page title commitment in two minutes. Maybe you asked Gemini to translate a dense state minimum standard into plain English and it did a surprisingly decent job.
Large language models are the AI tools most professionals encounter first, and for good reason. They're the most accessible, the most immediately useful for everyday tasks, and the most likely to produce a result that makes you think, "Alright, that actually saved me time."
They're also the most misunderstood. Not because people think they're magic (though some do), but because most professionals don't realize that the tool they're using isn't just a language model. ChatGPT, Claude, Gemini: these products look like a single AI, but under the hood, each one is a system of specialized components working together. The language model handles text. A code interpreter handles math. A web browser handles real-time information. A vision module handles images and documents. The LLM acts as the orchestrator, deciding which tool to call, when, and how to combine the results.
That distinction matters more than almost anything else in this post, and it connects directly to how the most effective AI solutions in this industry are being built.
In the first post in this series, we established that AI is pattern recognition, not magic. In the second, we walked through the different types of AI models and why "we use AI" tells you almost nothing about what a tool actually does. This post goes deep on the model type you're most likely to interact with directly: large language models. How they work, who's building them, what they're good at, where they fall short on their own, and why the real value comes when they're part of a larger system.
How LLMs Actually Work (Without the Computer Science Degree)
A large language model is a system trained on enormous amounts of text to predict what words should come next. That's the foundation of everything these tools do, and understanding it will save you from both underestimating and overestimating them.
During training, the model reads billions of pages of text: books, articles, websites, legal documents, scientific papers, code, government filings. It builds a statistical map of which words and phrases tend to appear near each other, in what order, in what context. "Bearings and distances" tend to appear together. "Property described in Deed Book" tends to be followed by a number. "Subject to easement" tends to precede a description of restrictions.
When you give an LLM a task, it's not looking up an answer in a database. It generates text one word at a time, predicting the most likely next word given everything that came before it. The output reads like it was written by someone who knows the topic because the model learned the patterns of how knowledgeable people write about that topic. It learned the shape of expertise without possessing the expertise itself.
Three technical concepts are worth knowing because they directly affect how you use these tools.
Tokens are the units LLMs process. A token is roughly three-quarters of a word. When a model has a "200,000-token context window," that means it can process and reference roughly 150,000 words in a single conversation, approximately 300 pages of text. When vendors quote pricing per million tokens, that's the unit of measurement.
Context windows determine how much information the model can hold in its working memory during a conversation. A small context window means the model forgets earlier parts of a long conversation. Larger windows let you feed in more reference material. Some current models support context windows of a million tokens or more, enough to process an entire project research package in one session. A year ago, the standard was 8,000 to 32,000 tokens. The expansion has been rapid.
Training data is what the model learned from, and it determines what the model knows and doesn't know. Most LLMs were trained on general internet text. They know a lot about a lot of things, but their depth in any specialized field depends on how much quality material about that field existed in their training data. Surveying-specific content, the kind of technical precision you need, represents a tiny fraction of the internet. That's why a general-purpose LLM can produce text that sounds authoritative about boundary law but gets the substance wrong.
Who's Building Them and Why It Matters
The LLM landscape has expanded rapidly. Understanding who builds the major models helps you evaluate the tools built on top of them, because every chatbot, every document assistant, every "AI-powered" product is running on one of these underlying models.
OpenAI (ChatGPT) pioneered the consumer chatbot and continues pushing the frontier. Their current lineup centers on the GPT-5 series for general-purpose work and o3 reasoning models for complex multi-step tasks. Most people's first exposure to LLMs was through ChatGPT, and it remains the most widely used platform.
Anthropic (Claude) emphasizes safety and reliability. Claude models currently support context windows up to one million tokens, enough to process hundreds of pages of project documents in a single session.
Google (Gemini) offers some of the largest context windows available, up to two million tokens on Gemini 1.5 Pro. Google has also focused on multimodal capability, meaning the same model processes text, images, and documents together. Their Gemini Flash models are positioned as cost-efficient options for high-volume processing.
Meta (Llama) took a fundamentally different approach. Meta releases its Llama models as open-weight, meaning anyone can download, modify, and run them without paying a licensing fee. Llama 4, their latest generation, is natively multimodal and uses a mixture-of-experts architecture. Meta's decision to make frontier-quality models freely available has reshaped the economics of the entire industry.
Mistral AI, based in France, builds models focused on European data sovereignty and multilingual capability, with infrastructure independent of U.S. cloud providers. DeepSeek, a Chinese research lab, demonstrated that frontier-quality LLMs can be trained for a fraction of established costs. Their V3 model achieved GPT-4-comparable performance with a reported training budget of approximately $5.6 million, versus $100 million or more for comparable models. Other notable players include xAI (Grok) and Cohere, which focuses on enterprise deployments that run on minimal hardware.
The field moves fast. Pricing has dropped dramatically, with some models costing less than $0.50 per million input tokens. Processing a 50-page title commitment might cost a few cents.
Open-Source vs. Closed-Source: A Decision That Matters
This distinction directly affects data privacy, cost, and control for professional work.
Closed-source models (ChatGPT, Claude, Gemini) run on the provider's servers. You send your data to their API, their servers process it, and they send back the result. Enterprise tiers offer stronger data protection agreements, but the fundamental architecture means your data leaves your environment.
Open-source models (Llama, Mistral, DeepSeek) can be downloaded and run on your own hardware or private cloud. Your data never leaves your environment. The tradeoff: you need the technical capability to deploy and maintain the model, or you need to work with a provider who runs it in isolation on your behalf.
For surveying professionals handling client data, deed records, and title commitments, this isn't academic. Your professional obligations around client confidentiality don't disappear when you paste a legal description into a chatbot. The RICS Professional Standard on responsible use of AI, effective March 9, 2026, explicitly requires surveyors to assess data handling practices of any AI tool they use in professional work. We'll cover security and data confidentiality in depth later in this series, but the key point for now: know where your data goes when you use an LLM, and make sure that answer aligns with your professional obligations.
Both approaches are viable. Closed-source models via enterprise agreements offer the most accessible path. Open-source models offer the strongest data control for firms with the technical resources to deploy them. The right choice depends on your firm's data sensitivity requirements, technical capabilities, and budget.
Where LLMs Create Real Value for Surveyors
Enough theory. Where does this technology actually earn its keep in your practice?
LLMs excel at tasks that involve processing, generating, or transforming text. In a profession that produces and consumes enormous amounts of written documentation, that's a significant portion of office work. The key is matching the tool to tasks where its strengths apply and its limitations don't create unacceptable risk.
Drafting and refining legal descriptions. An LLM can produce a first draft of a metes-and-bounds description from field data, or restructure an existing description into a different format. Every bearing, every distance, every call to a monument still needs verification by a licensed professional. But starting from a structured draft and verifying it is faster than writing from scratch, particularly on routine descriptions where the format is predictable. The LLM handles the boilerplate. You handle the precision.
Summarizing title commitments and deed packages. A 40-page title commitment with 25 Schedule B-2 exceptions is a lot of reading. An LLM with a large context window can ingest the entire document and produce a structured summary: property owner, legal description, exceptions listed with recording references, easements identified, relevant conditions. That summary doesn't replace reading the document. It gives you a roadmap so you know where to focus your attention.
Generating report narratives from structured data. You have the field data, the measurements, the findings. The LLM drafts the narrative sections of your survey report, environmental assessment, or project summary. You review and refine. For firms producing high volumes of reports where the narrative framework is consistent across projects, this saves meaningful time.
Interpreting regulations and standards. State minimum standards, zoning ordinances, subdivision regulations, county recording requirements: the regulatory environment is dense and varies by jurisdiction. LLMs can translate this language into plain English summaries, compare requirements across jurisdictions, and identify provisions relevant to your project. Confirm what the model tells you against the actual statute, but using it as a research accelerator beats reading 80 pages of legalese to find the three paragraphs that apply.
Proposals, client communication, and business writing. This is the lowest-risk, highest-return application for most firms. Drafting scoping documents, writing proposal narratives, composing client emails, building RFP responses. A proposal that used to take three hours of writing time on a Sunday night gets a solid first draft in 15 minutes. You refine it in 45. That's two hours back, every time. For firms looking to scale without proportional hiring, the time recovered on business tasks alone can shift how leadership spends their weeks.
What an LLM Can't Do on Its Own, and Why That's the Wrong Question
An LLM, on its own, doesn't know what it doesn't know. It has no mechanism for checking whether its output is factually correct. It produces the text pattern that best matches your request based on what it learned during training. When its training data contained enough relevant, accurate information, the output is good. When it didn't, the model fills in the gaps with plausible-sounding text that can be completely fabricated.
This behavior, called hallucination, is the most important limitation for any professional to understand.
An LLM asked to generate a metes-and-bounds description will produce something that reads like a perfectly formatted legal description. The language will be correct. The structure will follow standard conventions. But the bearings might not close. The distances might contradict the deed. It might reference a monument that doesn't exist or cite a recording reference it invented. The output looks authoritative because the model learned the pattern of what authoritative legal descriptions look like. It didn't verify any of the content.
On its own, an LLM doesn't do math. It predicts text. When it produces a number, it's generating the number that pattern-matching suggests should appear in that position, not computing a result. Boundary law interpretation carries similar risk. An LLM can quote statutes and cite case law, but studies of AI-generated legal analysis have found models producing fabricated citations and misstated legal holdings at significant rates. Factual claims about specific properties or records are unreliable unless the model has been given the actual source documents in its context window.
Those are real limitations. But "on its own" is doing important work in those sentences, because modern LLM platforms don't operate on their own.
The Bigger Picture: LLMs as Orchestrators, Not Solo Performers
If you've asked ChatGPT or Claude to solve a math problem and gotten the right answer, you're not imagining things. You're seeing the system at work, not just the language model.
Modern LLM platforms can call external tools: code interpreters that execute real calculations, web browsers that retrieve current information, vision modules that read images and documents, file analyzers that process uploaded data. When ChatGPT writes and runs a Python script to solve an equation, the math is real. The LLM recognized that the task required calculation, wrote code to perform it, and passed that code to an interpreter that executed it. The language model orchestrated the workflow. A specialized tool did the computation.
This is the insight that changes how you should think about LLMs and their role in professional work.
The products you interact with, ChatGPT, Claude, Gemini, are not single models. They are systems. The language model sits at the center, acting as a coordinator that decides which specialized capability to invoke for each part of a task. Need to read a scanned document? The vision module handles that. Need to compute a coordinate transformation? The code interpreter handles that. Need to search for current regulatory information? The web tool handles that. The LLM ties the pieces together, interpreting results and producing a coherent response.
Sound familiar? It should. Your firm works the same way. A party chief doesn't personally manufacture the total station, calculate the adjustment, draft the plat, check the title, and stamp the survey. Different specialists handle different steps, each contributing the capability they're built for, with a project manager coordinating the workflow and a licensed professional reviewing the result.
The reason this matters for surveyors and geospatial professionals goes beyond understanding how ChatGPT works. It reveals something fundamental about how AI creates value in professional workflows. A single model trying to do everything is like asking one person to handle every phase of a complex project. The results will be inconsistent. Some tasks will be handled well. Others will be handled poorly, and you may not know which is which until something goes wrong.
The alternative is a system where specialized components handle what they're designed for. A computer vision model reads the scanned plat. An LLM interprets the legal description and extracts structured data. A computation module verifies the math. A second LLM pass reviews the output against the source documents and flags inconsistencies. Each component does the thing it's best at, and the pipeline produces a result that no single model could achieve alone.
This is exactly how Enspectri approaches workflow automation for surveying and geospatial teams. Not one model trying to handle a complex processing pipeline, but purpose-built systems where specialized AI capabilities handle different steps, with professional review at every decision point. The same principle that makes ChatGPT more useful than a raw language model, giving it tools to compensate for its weaknesses, is the principle behind building AI systems that actually work in production environments where errors have real consequences.
The catch is that the LLM is still deciding what to do, which tool to call, how to interpret inputs, how to combine results. If the language model misreads a bearing from a document and passes that wrong value to a code interpreter, the computation will be precise and precisely wrong. The math executes flawlessly. The answer is still incorrect because the orchestrator made a bad decision upstream.
That's why the professional stays in the loop. Not because the technology doesn't work, but because the value of a multi-component AI system, like the value of a well-run survey crew, depends on someone with expertise and judgment overseeing the process. The AI handles volume and speed. You handle verification and accountability.
Using LLMs Effectively: The Professional's Framework
Understanding the limitations isn't the end of the conversation. It's the beginning of using these tools well. The professionals getting the most value from LLMs have developed workflows that leverage the model's speed while guarding against its weaknesses.
Start with tasks where errors are easy to catch and low in consequence. Proposal drafts, client email templates, report narrative frameworks, research summaries of regulations you'll read yourself anyway. Build comfort and develop judgment about where the model performs well and where it doesn't. Every firm's workflow is different, and the best way to learn what works is to test it on low-stakes output first.
Provide the source material directly. Don't ask an LLM to tell you what's in a deed. Paste the deed text into the conversation and ask the model to summarize, extract, or reorganize what you gave it. When the model works from your source documents rather than its training data, the outputs are grounded in actual records rather than statistical patterns. This single practice eliminates most hallucination risk on document processing tasks.
Use LLMs to check their own work. One of the most powerful techniques is running a second pass where you ask the model to review its own output for errors, inconsistencies, or unsupported claims. Provide the original source material and the model's first draft, then ask it to compare the two and flag anything in the draft that isn't directly supported by the source. This iterative review loop catches errors that a single pass misses. We'll go much deeper on techniques like this in upcoming posts on prompt engineering and building verification workflows.
Build reusable context for your firm's specific work. An LLM produces better output when you give it context about your firm's standards, templates, preferred formats, and quality requirements. A document that describes your standard report structure, your certification language preferences, your client communication tone: these context files turn a generic tool into one that produces output closer to what you'd write yourself. Later in this series, we'll cover how to build training materials, style guides, and custom instructions that shape AI behavior for your specific practice.
Never skip professional review on work that carries liability. This applies to every tool in your workflow, not just AI. A CAD drawing gets checked before it's stamped. A legal description gets reviewed before it's recorded. LLM output follows the same principle. The model produces a draft. A licensed professional reviews, verifies, and takes responsibility for the final product. The professional's judgment is the quality control layer that makes the tool safe to use.
The firms already integrating LLMs into their workflows aren't treating these tools as replacements for professional judgment. They're treating them the way they treat every other force multiplier: something that handles volume so their people can focus on the work that actually requires their expertise.
What Comes Next
This series continues with deep dives on vision and vision-language models (the AI that processes your imagery, scanned documents, and plats) and geospatial-specific foundation models (AI built from the ground up to understand spatial data). Later, we'll cover the practical skills that make all of these tools more effective: prompt engineering techniques for technical precision, building context and training materials that teach AI your firm's standards, and constructing verification workflows that catch errors before they leave your office.
The tools are here. The question is whether you'll learn to use them effectively or wait while firms that do pull further ahead. LLMs won't replace your judgment. But they'll handle a lot of the text-heavy work that buries your best people in tasks that don't require it.
This is Part 3 of the AI for Surveyors and Geospatial Professionals blog series by Enspectri. Built by industry insiders who've run survey crews, scaled geospatial operations, and shipped production software. See how Enspectri's AI-powered tools deliver real results for real workflows.
Sources
- Bluebeam, "Building the Future: AEC Technology Outlook 2026" (October 2025) (27% AEC AI adoption, 94% plan to increase usage)
- RICS Professional Standard, "Responsible Use of AI in Surveying Practice" (effective March 9, 2026) (AI data handling and accountability requirements for surveyors)
- OpenAI, Platform Documentation: Models (GPT-5 series and o3 reasoning model specifications)
- Anthropic, "Introducing Claude Opus 4.6" (February 2026) (1M token context window, adaptive thinking capabilities)
- Google, Gemini API Models Documentation (Gemini model lineup, 2M token context on 1.5 Pro)
- Meta AI, "The Llama 4 Herd" (2026) (Llama 4 open-weight release, mixture-of-experts architecture)
- Euronews, "Mistral, Europe's AI champion, releases new, smaller frontier models" (December 2, 2025) (European sovereignty positioning, multilingual frontier models)
- South China Morning Post, "DeepSeek kicks off 2026 with paper signalling push to train bigger models for less" (January 2026) (DeepSeek V3 training cost efficiency)
- LLM.co / Manila Times, "Study on the Growth of Open Source vs. Closed Source LLM Adoption" (March 2026) (41% of enterprises plan to expand open-source LLM usage; hybrid strategies emerging)