LLM Applications for Customers

Topics Covered

Where LLMs Fit

Tasks that suit LLMs

Tasks that do not suit them

Starting from the workflow

Human in the loop

Setting expectations

Choosing a model

Build on the product where possible

Measuring the baseline

Small first, then widen

Explaining LLMs to the customer

Who owns the output

Data the model will see

Integrating with existing tools

Reviewing the use case with the customer

Common early mistakes

When the answer is not an LLM

Planning for model change

Large language models have become one of the most requested capabilities in customer deployments. Customers want assistants that answer questions over their documents, tools that draft replies, extract fields from contracts, summarise cases and route requests. FDEs are often the people who turn these requests into systems that work on real customer data, inside real constraints on privacy, cost and accuracy. This lesson covers where LLMs fit, how to ground them in customer documents, how to design prompts and tools, how to keep them safe and how to run them in production.

Tasks that suit LLMs

LLMs are strong at tasks involving language with some tolerance for variation: summarising long text, drafting replies for a person to review, extracting fields from documents with varied layouts, classifying free text, answering questions from a set of documents and translating between formats. They do well when a human can check the output quickly, and when a small error is cheap compared with the time saved.

Tasks that do not suit them

LLMs are a poor fit for exact calculation, decisions that must be fully explainable by rule, tasks where any error is costly and the output cannot be checked, and problems already solved well by simple rules or a query. Do not use a model to add up invoice totals, to decide who gets a loan with no human review or to look up a value that sits in a database column. A rule or a query is cheaper, faster and right every time.

Starting from the workflow

Begin with the customer workflow, not with the model. Who does the task today, how long does it take, what does a good result look like and what happens when it is wrong? A support team that spends six minutes reading each case before replying may gain more from a two line summary at the top of the case than from a chatbot. The workflow tells you where the model output appears and who checks it.

Human in the loop

Most customer LLM applications should keep a person in the loop, at least at first. The model drafts and the person approves, edits or rejects. This protects the customer from errors, gives users control and produces a stream of corrections that show where the system is weak. Over time, as accuracy is measured and trust grows, some outputs may be applied automatically, such as routing low risk requests.

Setting expectations

Customers often arrive with expectations set by demos and headlines. Be clear early that model outputs are probabilistic, that they will sometimes be wrong and that the system will be measured against agreed accuracy targets on their own data. Explain what the model can see and what it cannot, such as documents it was not given. Setting expectations honestly at the start avoids disappointment later and makes evaluation, covered in the next lesson, a shared goal.

LLMs fit tasks where a person can check the output quickly and a small error costs less than the time saved.

Choosing a model

Choose the model for the task, the data rules and the budget. Larger models usually handle complex reasoning and long documents better, while smaller ones are faster and cheaper for classification and extraction. Check where the model runs, what the provider does with inputs, whether the customer allows that provider and which regions are available. Many customers have approved providers and contract terms already, and choosing outside them adds weeks of review.

Build on the product where possible

Many products now include LLM features such as summaries, search and drafting. Before building something custom, check whether a product capability can be configured to meet the need. Custom LLM pipelines carry ongoing work: prompt changes, evaluation, model upgrades and monitoring. Building on the product shares that work with the product team and other customers.

Measuring the baseline

Before building, measure how the task performs today: time per case, error rate, backlog size and how often work is redone. Without a baseline, the customer cannot tell whether the LLM application helped, and debates about value turn into opinions. The baseline also sets the bar the system must clear in evaluation.

Small first, then widen

Start with one task, one team and one type of document. A summary for one support queue, extraction of five fields from one contract type or drafts for one category of request. Prove the value, learn the failure modes and build the evaluation before widening to more teams and tasks. LLM projects that start with an assistant for everything usually struggle to show value on anything.

Explaining LLMs to the customer

Many people the FDE works with will not know how LLMs work. A short, plain explanation helps: the model predicts likely text based on its training and on what it is given in the prompt, it does not look things up unless the system gives it documents and it can produce confident text that is wrong. This helps users judge outputs sensibly and helps sponsors understand why grounding, evaluation and review matter.

Who owns the output

Agree who is responsible for what the system produces. If an agent sends a drafted reply, the agent and their team own it, just as they would own a reply they typed. If the system routes requests automatically, the customer process owner owns the routing rules and the accepted error rate. Writing this down early prevents arguments after the first visible mistake, and it shapes the design of review and audit features.

Data the model will see

List exactly what data the application will send to the model: case text, customer names, account details, attachments. Check each item against the customer data rules and the provider terms. Remove or mask fields the task does not need, such as payment details in a support case being summarised. Many security reviews of LLM applications focus on this list, so having it ready speeds approval.

Integrating with existing tools

Users adopt LLM features that appear inside the tools they already use: the case system, the contract repository, the email client. Plan the integration with these systems early. A separate application that users must open, copy text into and copy results back from rarely survives the first month, however good the model is.

Reviewing the use case with the customer

Before building, write a one page use case and review it with the customer owner, a few users and their security or compliance contact: the task, who uses the output, what data the model sees, how errors are caught, the baseline and the target. This page becomes the reference for scope, evaluation and the security review, and it often surfaces a blocking concern before any code is written.

Common early mistakes

Teams new to LLM projects often make the same mistakes: they demo on a handful of hand picked examples, skip evaluation, let the scope grow to every task the sponsor mentions and ignore cost until the first invoice arrives. Each of these is avoidable with the steps in this lesson, and each one has ended promising deployments.

When the answer is not an LLM

Sometimes the use case review shows that a search box, a better form or a simple rule would solve the problem. Say so. Recommending the simpler tool builds credibility, saves the customer money and frees the LLM budget for a task where it really helps. Customers remember the FDE who told them the truth about a fashionable technology.

Planning for model change

Models improve and change quickly, and providers retire old versions. Design the application so the model can be swapped with a configuration change and an evaluation run, rather than a rewrite.

Keep the prompts and settings in versioned configuration.

Workflow first, model second

Use LLMs where a person can check the output quickly and an error is cheap, and use rules or queries for exact answers. Start from the workflow, keep a person in the loop, set honest expectations, pick a model the customer allows, build on the product where possible, measure the baseline and start small.