AI for Non-Technical Leaders (Video Course / Guide)

AI is software that makes predictions based on patterns in data. That's it.

When you ask ChatGPT a question, it's not "thinking." It's predicting the most statistically likely next word, then the next, then the next, based on billions of examples it saw during training. When your email filters spam, it's predicting whether a message matches patterns it learned from millions of labeled emails.

This matters because it changes how you should evaluate AI tools. Don't ask "Is this intelligent?" Ask "Does this prediction solve my problem?"

What AI does well:

• Pattern recognition at scale (reviewing 10,000 resumes for keywords)
• Generating text that follows learned formats (drafting engagement letters)
• Classifying information into categories (routing support tickets)
• Extracting structured data from unstructured sources (pulling dates and amounts from invoices)

What AI does poorly:

• Tasks requiring true reasoning or logic chains
• Anything where being 95% accurate isn't good enough (legal compliance checks)
• Understanding context it wasn't explicitly trained on
• Knowing when it doesn't know something

If you remember nothing else: AI is a prediction engine, not a reasoning engine. Use it where predictions add value.

You don't need to understand transformers or neural networks. You need to understand three things.

1. Training: Learning patterns from text

An LLM

reads billions of documents (books, websites, code repositories) and learns which words tend to follow which other words in which contexts. It builds a massive statistical model of language patterns.

When you see "The attorney filed a motion to..." your brain predicts "dismiss" or "compel" might come next. An LLM

does the same thing, but across millions of pattern variations simultaneously.

2. Prompting: Activating the right patterns

When you write a prompt, you're not giving instructions to a person. You're activating specific statistical patterns in the model.

Bad prompt: "Write something about client onboarding."

Good prompt: "You are a senior operations manager at a mid-sized law firm. Write a 3-step client onboarding checklist for new corporate clients. Include specific documents to collect and systems to update."

The second prompt activates more relevant patterns because it provides context (law firm, corporate clients) and structure (3 steps, specific format).

3. Generation: Predicting one token at a time

The model generates responses one "token" (roughly a word or word fragment) at a time. Each token is predicted based on all previous tokens in the conversation.

This is why LLMs

sometimes "drift" in long responses. Early predictions constrain later ones. If the model starts down the wrong path, it keeps going because each new word is predicted based on the words before it.

Practical implication: Break complex tasks into smaller prompts. Don't ask for a 10-page document in one shot. Ask for an outline, then expand each section separately.

An agent is an LLM

with three additions: memory, tools, and a decision loop.

Standard LLM

interaction:

1. You send a prompt
2. Model generates a response
3. Done

Agent interaction:

1. You send a goal ("Find all clients we haven't contacted in 90 days")
2. Agent breaks this into steps (query CRM

   CRM

   Click to read the full definition in our AI & Automation Glossary.

   , filter by last contact date, format results)
3. Agent uses tools to execute each step (CRM

   CRM

   Click to read the full definition in our AI & Automation Glossary.

   API

   API

   Click to read the full definition in our AI & Automation Glossary.

   , spreadsheet formatter)
4. Agent checks if goal is met; if not, tries another approach
5. Agent returns final result

Real example: You ask an agent to "prepare a conflict check for Acme Corp."

The agent:

• Searches your document management system for "Acme"
• Queries your CRM

  CRM

  Click to read the full definition in our AI & Automation Glossary.

  for related entities and contacts
• Checks your matter management system for adverse parties
• Compiles findings into a structured report
• Flags potential conflicts for human review

You didn't tell it each step. You gave it a goal, and it figured out the steps.

Key difference from chatbots: A chatbot follows a decision tree you built. An agent decides its own path based on the goal you set.

Traditional chatbots are if-then scripts. You map every possible conversation path in advance.

User says "billing question" → Route to billing script → Ask "What type of billing question?" → If "invoice" then show invoice options → If "payment" then show payment options.

This works for narrow, predictable interactions. It breaks when users ask anything you didn't script.

LLM

-powered agents are different in four ways:

1. No predefined paths The agent interprets intent from natural language. Users can ask "Why is my invoice higher this month?" or "I think you charged me twice" or "Can I get an itemized breakdown?" The agent understands these are all billing questions without you mapping each variation.

2. Context retention Agents remember the conversation. If a user asks "What about last month?" the agent knows "last month" refers to the billing period you were just discussing. Chatbots forget context between steps unless you explicitly program memory.

3. Tool use Agents can call external systems. When a user asks about their invoice, the agent queries your billing system, retrieves the data, and formats a response. Chatbots can only display information you pre-loaded.

4. Failure recovery If an agent's first approach doesn't work, it tries another. If the CRM

API times out, it might try a database query instead. Chatbots just error out.

When to use each:

Use a chatbot when:

• The interaction is simple and fully predictable (password resets, appointment scheduling)
• You need 100% control over every response
• Compliance requires exact wording

Use an agent when:

• Users ask questions in unpredictable ways
• The task requires multiple steps or system integrations
• You want the system to improve based on new data

1. Intake and Qualification

The task: A potential client fills out a web form or sends an email. Someone needs to determine if they're a good fit, what service they need, and who should handle it.

The AI approach: An agent reads the intake form or email, extracts key information (industry, issue type, urgency, budget), checks against your qualification criteria, and routes to the appropriate partner or practice group.

Specific implementation:

• Connect the agent to your intake form (Typeform, Google Forms, website contact form)
• Give it access to your client qualification rubric
• Set up routing rules (corporate M&A → Partner A, employment disputes → Partner B)
• Configure it to draft a preliminary engagement scope for partner review

Time saved: 2-3 hours per week per intake coordinator.

2. Document First Draft Generation

The task: An associate needs to draft a standard document (engagement letter, NDA, demand letter, audit planning memo).

The AI approach: The associate provides key details in a structured prompt. The agent generates a first draft using your firm's templates and style guide.

Specific implementation:

• Create a prompt template for each document type
• Include [CLIENT_NAME], [MATTER_TYPE], [KEY_TERMS] placeholders
• Store your firm's standard language and clauses in the agent's knowledge base
• Set up a review workflow (agent drafts → associate reviews → partner approves)

Example prompt for engagement letter:

Generate an engagement letter for [CLIENT_NAME] for [MATTER_TYPE].
Scope: [BRIEF_SCOPE]
Fee structure: [HOURLY/FLAT/CONTINGENCY]
Key terms: [SPECIAL_TERMS]
Use our standard limitation of liability and dispute resolution clauses.

Time saved: 30-60 minutes per document.

3. Client Communication Summarization

The task: After a client call or email thread, someone needs to update the matter file with a summary and next steps.

The AI approach: The agent reads the call transcript or email thread and generates a structured summary with action items.

Specific implementation:

• Use a transcription tool (Otter.ai, Fireflies.ai) to capture call audio
• Feed transcript to agent with this prompt: "Summarize this client call. Include: decisions made, open questions, action items with owners and deadlines, and any concerns raised."
• Agent outputs structured summary
• Associate reviews and saves to matter file

Time saved: 15-20 minutes per call or email thread.

4. Research and Analysis Assistance

The task: You need to analyze a large document set (discovery materials, financial statements, contract portfolio) to find specific information or patterns.

The AI approach: The agent reads all documents and answers specific questions or generates a summary report.

Specific implementation:

• Upload documents to a vector database (Pinecone, Weaviate) or use a tool with built-in document analysis (Claude, ChatGPT with file upload)
• Ask targeted questions: "Which contracts have auto-renewal clauses?" or "Summarize all references to intellectual property ownership."
• Agent searches all documents and compiles findings

Time saved: 3-5 hours per research project.

Pick one task that meets these criteria:

• Takes 30+ minutes each time it's done
• Happens at least weekly
• Follows a consistent pattern
• Doesn't require perfect accuracy (human review is acceptable)

Map out the current process in 5-10 steps. Identify which steps involve pattern recognition, text generation, or data extraction. Those are your AI opportunities.

Start with the simplest possible implementation. If you're testing document drafting, start with one document type. If you're testing intake, start with one practice area.

Run it in parallel with your current process for two weeks. Compare outputs. Measure time saved. Adjust prompts based on what works and what doesn't.

AI won't replace your judgment. It will give you more time to apply it.