LONGAN GUANGZHOU · AI FOUNDATIONS

From Asking AI,
to Getting Deliverables

Boyang Li · Beijing Longan (Guangzhou) Law Firm

AI FOUNDATIONS FOR LAW FIRMS2026
SPEAKER

Boyang LiAttorney

Beijing Longan (Guangzhou) Law Firm
Director, Longan Bay Area AI & Law Research Center

  • FOCUSAI and algorithm governance, data compliance, intellectual property, and internet disputes
  • PRACTICELarge language models, agents, legal knowledge bases, and electronic data processing
  • WORKCo-author, ChatGPT for Legal Professionals · Creator of WordOllama

More than a decade of internet law practice

Boyang Li, speaker and attorney
WHERE WE ARE
64%

Still Stuck in
Web Chat

AI is already part of daily work. The next step is to make it deliver more.

79%Use AI every day or almost every day
45%Believe AI can save more than 30% of their time
36%Already use agents to complete tasks automatically
0%Believe AI is essentially unhelpful

Internal survey | n=33 valid responses; multiple-choice totals follow the survey platform’s methodology

n=33 valid responsesInternal survey
LIVE RANKING

Live ranking: where do these AIs belong?

GOAT
ELITE
STRONG
NPC
HARD PASS
WAITING
Chinese models | Drag, or select one and choose a tier
US models | Drag, or select one and choose a tier
PUBLIC BENCHMARKS

Model Score × Cost per Task

Select either chart to enlarge
Artificial Analysis Intelligence Index scores, August 3, 2026; select to enlarge
A · Intelligence Index: model scores
Artificial Analysis Intelligence Index versus cost per task, August 3, 2026; select to enlarge
B · Intelligence × Cost: cost per task
Source: Artificial Analysis · 2026-08-03Select a chart to enlarge
LEGAL CAPABILITY

Legal capability: relative performance on English benchmarks

Select chart to enlarge
Artificial Analysis Legal Index model scores, August 5, 2026; select to enlarge
It combines legal knowledge, agentic knowledge work, long-context tasks, non-hallucination, and reasoning. The benchmark is English-first and rooted in Western legal practice. Use it to compare models—not as a proxy for Chinese-law capability.Artificial Analysis · 2026-08-05
AI IS ALREADY IN LEGAL SERVICES

Individuals Are Already
Using AI to Litigate

UK employment tribunals are seeing a surge in claims. Judges report that more self-represented litigants are using ChatGPT and Grok to draft pleadings, applications, and loss calculations.

UK EMPLOYMENT TRIBUNALSTHE DECODER · 2026.08

Hundreds of pages that look legal—
but may contain invented law and inflated claims

A joint memorandum from UK employment tribunal presidents Barry Clarke and Susan Walker reported that cases received in March were up 39%, while the backlog had grown 55%.

Many AI-generated filings ran to hundreds of pages and included nonexistent authorities and unrealistic claims.

“See? AI is dumb.”
What matters is that client expectations are changing
F
Fast responseClients now expect an immediate first answer
D
Deeper analysisLists of rules and common knowledge are no longer scarce
L
Long-form outputStructured, multi-page analysis can be generated quickly
C
Clear languageComplex issues can be restated in plain language instantly

More people are already using AI. If we only dismiss it, we risk falling behind the world our clients already inhabit.

Chatbot

Chatbot
One Question at a Time

Summarize a document, revise a paragraph, explain a concept, or suggest a few approaches.

  • 01After answering, it usually waits for your next instruction
  • 02Most output remains inside the chat box
  • 03A human still has to copy, organize, and verify
  • 04It is optimized for answers; complex work still needs human decomposition, uploads, and handoffs
  • 05It depends on what you upload and may fill gaps by guessing
A chatbot answersbut rarely owns the full deliverable
COMMON CHATBOT MISTAKES

AI Is Not a Person
Getting Angry
Only Burns Tokens

You carry the case history, drafting experience, and legal judgment in your head. The chatbot may see only the sentence in front of it.

01
Make it sound more like a lawyer“Like a lawyer” says nothing about the audience, position, or document type.
02
Give me the legal basis for receiving 2N compensationMany web chatbots are tuned to please. Give them the desired conclusion first, and they may reverse-engineer the reasoning.
03
It feels wrong—revise it againThis never identifies what is wrong, why it is wrong, or what the revision must achieve.
After several unsatisfying drafts—and still no useful new information—the prompt becomes: You are so stupid
COMMON CHATBOT CAPABILITIES

It Can Do It—
Just Not Reliably

A web chatbot’s browsing, memory, image editing, and file reading are constrained by product permissions and implementation choices.

Capability
What you expect
What happens
Web access
Search the whole web and eliminate bad answers
Usually reads only a small result set—and sometimes only sources available through its own channels.
Context
Remember everything—“I will train my own specialist model”
Free web products often expose less usable context than advertised. Long chats and many attachments make earlier requirements easier to miss.
Image editing
Precisely edit the existing file like image software
Most products regenerate a new image. Upscaling, compression, transparency, and exact dimensions remain unreliable.
Legal research
Query authoritative legislation and case databases
It may browse low-quality articles instead, then confidently cite invented cases or outdated rules.
File reading
Upload it and the model understands everything
Weak OCR and broken extraction can misalign text and structure, producing nonsense downstream.

Common skepticism about AI often comes from being burned by a free chatbot.

A BETTER WAY
Markdown

Markdown:
Structured Documents
for AI

Separate goals, problems, steps, and examples with headings, hierarchy, lists, and quotations.

AI can read plain text, but clear structure makes important requirements harder to miss.

# Level-one heading
## Level-two heading

This is **bold**, *italic*, `inline code`, and a [link](https://example.com).

> This is a quotation: give the conclusion first, then the authority.

- Unordered list
- Second item

1. Ordered list
2. Second item

| Field | Content |
| --- | --- |
| Conclusion | First |

```text
Give the conclusion first, then list the authority
```
Toggle between source and rendered output
Markdown is not an SEO/GEO shortcutbut it is a strong source format for structured publishing and AI
ONE DRAFT, MANY USES
# Injured on the Commute: Is It a Work Injury?

## Conclusion First
It may qualify when **time, route, and responsibility requirements** are satisfied…

## Analysis
The analysis cannot stop at “after work.” It must consider the **purpose of the commute, route deviations, and responsibility for the accident**.

If the employee detoured to collect a child, explain the reason, distance, and time before assessing whether the commute remained reasonable.

## Common Disputes
### Is a detour to collect a child reasonable?

## Practical Steps
- Preserve the accident determination
- Preserve attendance and commuting evidence

## Authorities and Sources
Toggle between source and rendered output

For Every Article,
Keep a Markdown Master

  • 01Clear heading hierarchy
  • 02Easy conversion for different publishing platforms
  • 03Easy for AI to summarize, split, and adapt
  • 04Supports structured web publishing and SEO/GEO extraction
SEO/GEO still depends on quality, sources, author credibility, and crawlabilityMarkdown = content master
MARKDOWN · CONTENT SOURCE

Markdown Is Readable
and Token-Efficient

As plain text, Markdown moves naturally between AI tools and websites without constant reformatting.

CONTENT SOURCE

One Structured
Markdown Master

  • Heading hierarchymakes the article structure explicit
  • Body contentis not locked to one visual format
  • Emphasisuses bold, quotations, and lists
  • Source linkssupport verification and citation
OUTPUTS

Convert the Format
for Each Use Case

  • HTML pageadds site styles, images, navigation, and interaction
  • Markdown textis read directly by search, agents, and LLMs
  • Content stays consistentinstead of being rewritten
  • Maintain one sourceand update both outputs together

Write the content once.
Publish it many ways.

Markdown preserves content structure; HTML handles presentation. A website, newsletter, knowledge base, and AI reader can all start from the same master.
CLOUDFLARE · TRAFFIC SHIFT2026.08

Non-human traffic, including AI crawlers,
exceeds human traffic

A person shops for a camera · perhaps 5 sites
→
Agent research · perhaps 5,000 sites
TIME · DESIGNED FOR AISAME ARTICLE
303 KBHuman browsers receive full HTML: images, menus, styles, and interaction
→
13 KBAI crawlers receive Markdown: headings, body, links, and sponsorship

TIME now identifies ClaudeBot and PerplexityBot and serves them dedicated text/markdown pages.

Visual design competes for human attention.
Structured content competes for AI citations.

Even TIME is adapting for AI.
Before TIME did, I had already added automatic Markdown delivery to our new law-firm website.
GLOBAL WEB · AI DELIVERY

Websites Are Already
Rebuilding Content for AI

CLOUDFLARE · 2026
>50%of internet traffic is non-human

The share crossed 50% for the first time in 2026. Agents and crawlers are now a major web audience.

CRAWLER PURPOSE
52%of crawler requests support AI training

The share rose from 22% in spring 2025 to 52% by June 2026.

USER-TRIGGERED FETCHES
×21more than 21× growth in one year

When users ask AI to search, research, or answer, agents fetch material directly from the web.

TIME · DESIGNED FOR AI

One Story,
Different Versions for Humans and AI

Readers see the full web page. AI crawlers receive easier-to-parse Markdown, now including sponsorship written for AI audiences.

Web page · Human→Markdown · AI
CLOUDFLARE · OUTLOOK

The “Agentic Web”
Has Arrived

Over time, human and AI visits will be identified and served separately. Content must be worth reading for people—and easy for AI to extract, verify, and cite.

Even TIME is adapting for AI.
Law firms must now ask: who are our articles really written for?

We prepared early: our new law-firm website already supports automatic Markdown delivery.
The page serves human readers; the structured master also serves search engines, agents, and LLMs.
Agent

AGENT: Executes Autonomously to Deliver Results

LAW-FIRM EXAMPLE

“Review the case folder, organize the facts and evidence, and draft a complaint.”

AGENT / COMPLAINT TASK● RUNNING
USERDraft a complaint from the case-folder materials.
AGENTUnderstood. I’ll review the folder first.
AGENTFound 27 files. Classifying them and extracting key facts.
AGENTI’ll verify the parties, claims, facts, and evidence.
AGENTThe defendant’s service address is missing. Flagging it for follow-up.
AGENTDrafting from the firm template and checking names, amounts, and dates.
AGENTDraft complete: complaint.docx
01

Read Files

02

Extract Facts

03

Flag Gaps

04

Use Template

05

Verify Output

An agent is not always smarterbut it works closer to the job
ONLINE AGENT

Online Agents
Work Continuously in the Cloud

The provider supplies a controlled workspace. The agent can run code, process files continuously, and return finished deliverables.

InputsUpload files or connect approved cloud drives and accounts.
ToolsRun approved code, browsers, and data-processing tools.
LimitsExternal programs, network access, logins, and system permissions remain policy-controlled.
PROVIDER-HOSTED
KIMIGeminiChatGPT Work
CONTROLLED CLOUD WORKSPACE
case-files.zip
RUN CODE / OCR
REPORT · SHEET · FILE
ISOLATED · TEMPORARY · CONTROLLED
LOCAL AGENT

Local Agents
Use Files and Apps on Your Computer

Within the access you grant, an agent can locate materials, use installed tools, and save results to the specified location.

InputsRead approved folders directly—no file-by-file uploads.
ToolsUse Word, browsers, scripts, and internal tools to complete the task.
LimitsThe user still controls file scope, app permissions, and final actions.
COMMON LOCAL OPTIONS
WordbuddyQworkCodexCoWorkOpenCode
LOCAL / SELF-HOSTED AGENTS
OpenClaw (“Lobster”)Hermes Agent
AUTHORIZED LOCAL WORKSPACE ● APPROVED WORKSPACE
LOCAL
AGENT
Case FolderContracts · Receipts · Audio
WordTemplates · Pleadings
BrowserWebsites · Portals
ScriptsBatching · Validation
Internal ToolsCases · Databases
AGENT LIMITS

What Can an Agent Do?
Where Does It Stop?

MOST TASKS CAN MOVE FORWARD

If Code or a Computer Can Do It,
an Agent Can Usually Move It Forward

The prerequisites: permitted access, a clear goal, decomposable steps, and a result that can be checked.

READ→PROCESS→USE APPS→DELIVER
01

Subjective Taste

It can imitate references and an established style, but people must define—and ultimately judge—what feels polished or right.

02

Unreadable Materials

A model cannot reconstruct missing information from badly blurred or damaged scans. Image enhancement, OCR, and human review are often required.

03

Unverifiable Requests

“Make it better” and “it feels wrong” have no defined boundary. Without examples, rules, or acceptance criteria, the agent can only guess.

An agent’s limits often depend on whether the task is clear, readable, and verifiable.

HOW TO BRIEF AN AGENT

Skip Prompt Tricks.
Explain the Work.

5W2Hstill applies: who, what, why, where the files are, what period the sources cover, how to proceed, how far to go—and how to deliver.
  1. 01People & FilesWhose matter? Where are the files?
  2. 02Task & PurposeWhat is needed, and who will use it?
  3. 03Source PeriodWhat dates must cases, laws, and public sources cover?
  4. 04Method & LimitsWhat to follow, what it may do, and what it must not infer
  5. 05DeliverySave locally, email it, or send it through a connected messenger
USER / TASK BRIEFREADY

Read all materials in X:\Cases\Zhang-San-Employment-Dispute. Draft an employment-dispute complaint for filing with the court (Word). If more cases are needed, search only Guangdong judgments from January 2023 to the present, and state each decision date and source. Follow the existing folder template; use only facts in the materials, highlight missing information in yellow, and do not fill gaps by inference. When complete, save to X:\Cases\Zhang-San-Employment-Dispute\Output. Do not overwrite original files. Include a materials inventory and a list of missing information.

FILES → TEMPLATE → VALIDATE → WORDVERIFIABLE · DELIVERABLE

Run one first: confirm it can do the job correctly before scaling the same task.

YOUR FIRST AGENT TASK

One Sentence In
Three Legal Documents Out

A real example: after receiving the case folder, document types, counsel, and save location, the agent read the materials, invoked the right Skill, generated the documents, and validated them—without intervention.

01One instruction: the agent reads the files and invokes the right Skill
Natural-language instruction asking an agent to generate a retainer agreement and supporting documents (content blurred) 【INTERNAL SCREENSHOT】
02Generate, verify, and package the deliverables
Agent response after generating and validating the agreement (content blurred) 【INTERNAL SCREENSHOT】
03Deliver three Word documents at once
Civil retainer agreement, power of attorney, and law-firm letter generated by the agent (content blurred) 【INTERNAL SCREENSHOT】
ONE BRIEF→INVOKE SKILL→GENERATE & VALIDATE→THREE DOCUMENTS NO MID-TASK INTERVENTION
HOW AGENTS DELIVER

No Copy-Paste
AI Finds It.
AI Fills It.

An agent can locate parties, case numbers, facts, and other required information in the case folder, then place each item in the correct field.

  • 01Read the case materials and locate required information
  • 02Identify template fields, map the data, and fill them
  • 03Preserve headings, type sizes, paragraphs, and formatting
  • 04Check gaps and conflicts before saving the final document

No manual fact-hunting—and no repeated copying and pasting between source files and Word.

Civil retainer agreement auto-filled by the agent with formatting preserved (content blurred) 【INTERNAL SCREENSHOT】
CHOOSING AN AGENT

Which Agent Should You Choose?
Start with How You Want to Work

BEGINNER-FRIENDLY

Workbuddy

A visual interface for office files and everyday computer tasks—no command line required.

BEST FOR
Word, PDFs, spreadsheets, and file organization
WHY
Low learning curve; feels like everyday office software
LOW COST · TRY CODING

OpenCode

The open-source tool connects to local models or a provider of your choice—an inexpensive way to try a coding agent.

BEST FOR
Scripts, project edits, and automation experiments
WHY
Flexible model choices and strong value
TECHNICAL CHOICE

Codex

Built for complex coding, scripts, and sustained engineering work—and for turning experiments into repeatable workflows.

BEST FOR
Complex projects, tool development, and deep automation
WHY
Full engineering capabilities for long-term use

You do not need the “most powerful” agent on day one. Choose one you will open every day and that solves the work in front of you.

SAME MODEL, DIFFERENT RESULTS
Why Does “Your” AI Invent Law
While “Mine” Does It Far Less?
MCP

MCP Gives an Agent
Approved Sources

For legislation, query the legal database. For case information, read the case system. For evidence, access an approved file directory.

AGENT“Verify the cited provision, including its version and effective status.”
MCP
LEGAL DATABASEProvision · version · effective date · source link
CASE SYSTEMCase number · parties · milestones · case data
APPROVED FILESContracts · evidence · exhibits · firm templates

Unify the interfaces, and the agent can locate and invoke the capabilities it needs.

Skill

Skill:
Turn “I Suppose”
into “This Is the Process”

An instruction manual that teaches AI how the work must be done.

---
name: legal-research
description: Query legislation and validate citations
---

# When to use
When the user asks about legal authority, provisions, or limitation periods.

# Procedure
1. Query the legal database
2. Extract the provision, version, and source link
3. Verify the citation against the source text

# Failure handling
stop_if: Missing or inconsistent → stop and report.
A Skill defines how to research and deliverbut Skills vary greatly in determinism
TWO WAYS TO WRITE A SKILL

Same Task, Different Execution

TEXT-ONLY SKILL
SKILL.md · RULE-BASED ---
name: weather-query
description: Look up weather by city
---

# WEATHER LOOKUP SKILL

## Step 1
Open Baidu and search for “China Weather.”

## Step 2
Find the search field and enter the city supplied by the user.

## Step 3
Read the page, organize the results, and return them to the user.
HOW THE AGENT EXECUTES

Read Skill → open Baidu → search China Weather → locate the input → read and organize the page.

The model must decide how to navigate and whether to trust the result. Any step can drift.
SCRIPTED SKILL
SKILL.md · SCRIPTED ---
name: weather-query
description: Retrieve structured city weather
---

## Required argument
script.py --city CITY_NAME

## Example response
json:
{
  "city": "Guangzhou",
  "condition": "sunny",
  "temperature_c": 28,
  "high_c": 31,
  "low_c": 25,
  "wind": "southeast, force 2",
  "source": "China Weather"
}

## Failure handling
Source unavailable or field missing: stop; do not guess.
HOW THE AGENT EXECUTES

Receive city → call script.py → receive JSON → validate fields → return result.

The boundaries live in code: stop if a field or source is unavailable. The agent makes fewer judgments, uses fewer tokens, and produces more stable results.
Rules explain the procedureScripts fix inputs, outputs, and failure conditions
TWO TYPES OF SKILL

Rules Resemble Prompts;
Scripts Create Constraints

COMPARE
RULE-BASED
SCRIPTED
FORM
Text instructions
Scripts, tools, validators
AI ROLE
Interpret rules and choose next steps
Call script, pass arguments, receive result
FAILURE
Skipped steps, improvisation, inconsistency
Argument or validation failure; stop
BEST FOR
Voice, output preferences, flexible formats
Queries, batching, checks, document generation
VALUE
Less repetition; still model-dependent
Fixed limits, fewer judgments, fewer tokens, easy QA
A text-only Skill is a“saved prompt”; a scripted Skill is a“fixed workflow.”
If code can guarantee the step, do not rely on AI to “remember” itA practical distinction, not a formal industry standard
SCRIPTED SKILL · DETERMINISTIC OUTPUT

How “Fixed” Is a Scripted Skill?
It Will Faithfully Reproduce a Template Typo

TEMPLATEORIGINAL
Typo in the original power-of-attorney template (content blurred) 【INTERNAL SCREENSHOT】
OUTPUTAGENT DELIVERABLE
Agent-generated power of attorney reproducing the template typo (content blurred) 【INTERNAL SCREENSHOT】

Wait… how did nobody spot that typo?

RULE-BASED SKILL · ANOTHER USE

Another Use for Rules:
“Distill” the Partner

An associate gives the agent before-and-after drafts of the same document. It extracts the supervising lawyer’s stable preferences and continuously updates a reusable Skill.

Complaint · First Draft

The factual narrative comes first; the claims appear later.

Uses evaluative terms such as “obvious” and “malicious.”

Facts are not mapped to evidence item by item.

Complaint · Final Draft

Move the claims forward; present facts chronologically.

Remove evaluative language; describe facts objectively.

Cite the corresponding evidence number for each fact.

AGENT · DIFF ANALYSIS COMPARE EDITS · EXTRACT RULES
  1. Identify deletions, additions, order, and comments
  2. Track recurring editing preferences
  3. Separate stable rules from case-specific edits
  4. Create or update SKILL.md
SKILL.md · partner-style ---
name: partner-drafting-style
description: Draft to the supervising lawyer’s preferences
---

# STRUCTURE
- Put claims first
- Present facts chronologically

# LANGUAGE
- Remove evaluative language
- Map each fact to evidence
NEXT DRAFT→APPLY SKILL→PARTNER REVIEW→FEED NEW VERSION→UPDATE SKILL ↺

After a few cycles: the partner’s recurring preferences have been “distilled” into a Skill.

REDUCING HALLUCINATIONS

The Same AI Model Remains at the Core
Tools and Engineering Make It “Smarter”

The same model can perform very differently across agent frameworks—but it will generally outperform a plain web chat.

AGENTSAME MODEL + ENGINEERED CONSTRAINTS
RECEIVE TASK→ USE MCP
GET SOURCES→ USE SKILL
READ RULES→ VALIDATE
IN A LOOP→ FINAL
OUTPUT

Every step has a source, rule, and checkpoint. If evidence is missing or a citation fails, stop and obtain support.

CHATBOTSAME MODEL + IMPROVISED ANSWER
RECEIVE TASK→ SEARCH &
SELECT SOURCES→ COMPOSE
ON THE FLY→ OUTPUT
DIRECTLY

What to search, trust, and say is decided on the fly, usually without fixed validation or stop conditions.

Hallucinations can be reduced substantially—but never to zero. What an agent adds is consistency.

LOCAL DEPLOYMENT

Running a Model Locally?
Let an Agent Handle the Setup

State which model you want and what it will be used for. An agent can inspect the machine, choose a build, install, launch, test, and hand off the working setup.

  • INSPECTOS, CPU/GPU, RAM, VRAM, and disk space
  • SELECTRuntime, quantization, context length, and acceleration
  • DEPLOYDownload model, install dependencies, configure, and launch
  • HAND OFFRun live tests; leave logs and a one-click launcher
LOCAL DEPLOY / QWEN3.6 27B● RUNNING
USERDeploy Qwen3.6 27B locally on this computer.
AGENTUnderstood. I’ll inspect the hardware first.
AGENTDetected the OS, GPU, memory, and available disk space.
AGENTThis machine is suited to a 4-bit GGUF. I’ll use llama.cpp with hardware acceleration.
AGENTDownloading the model, installing the runtime, and generating launch settings…
AGENTLocal service is running. Testing responses, context handling, and restart flow.
AGENTAcceptance tests passed. Created one-click launcher: start-qwen.bat

You do not need to learn deployment first. Delegate “inspect → select → install → test” to an agent, then review what it delivers.

OPEN-SOURCE MODELS

Text and Documents:
Common Open-Source Projects

Open-source projects evolve daily and already cover most routine tasks.

LLM · LARGE LANGUAGE MODEL

Case synthesis, contract comparison, and legal drafting

  • Qwen3.6 27BLocal Q&A, summaries, and stable text work
  • Qwen3.6 35B-A3BFast batch organization and drafting
OCR · DOCUMENT RECOGNITION

Convert scanned files, receipts, and complex PDFs into text

  • PaddleOCR / PP-OCRScans, receipts, and conventional tables
  • MinerULong PDFs, tables of contents, and complex layouts
  • Unlimited-OCRBatch recognition for many files and long documents
  • OvisOCR2Complex pages and page-level layout understanding
OPEN-SOURCE MODELS

Speech: Understand and Speak

Turn recordings into searchable material—or produce short-video narration and voice-preserving translation.

ASR · SPEECH TO TEXT

Transcribe meetings, hearings, interviews, and recordings

  • FunASR · paraformer-zhLong Chinese recordings and batch transcription
  • FunASR · SenseVoiceSmallMultilingual speech, emotion, and sound-event recognition
TTS · TEXT TO SPEECH

Short-video narration and voice-preserving translation

  • Qwen3-TTSNarration, voice design, and voice cloning
  • IndexTTS2Preserve timbre and emotion for voice translation
DISCUSSION
Q

Open Discussion

What remains unclear—and which tasks are most worth unpacking together?

Which concept or example would you most like to explore further?
Which part of legal work would you most like an agent to take over?
Which agent, Skill, plugin, or workflow are you using or testing?
What concerns you about using agents in daily practice?
LIVE DISCUSSIONQ&A
NEXT SESSION
NEXT SESSION08.28
AI AGENT · HANDS-ON

For Anyone Finally Ready
to Move Beyond Doubao

Start with installation, complete a real task live, and learn the practical techniques behind effective agent use.

SET UPChoose an agent, install it, and complete the basic configuration
RUNRead folders, use apps, and create and revise documents
REPEATBrief tasks, review results, and turn experience into a Skill

Bring: a laptop and one anonymized real-world task to work through live.

08.28 · HANDS-ON AI AGENTSNEXT
THANK YOU

Thank
You

Let’s keep talking about AI tools, workflows, and practical use cases for law firms.

Boyang LiLong An Law Firm · Guangzhou
Boyang Li personal WeChat QR code
PERSONAL WECHATScan to connect and continue the conversation
BOYANG LI · LONG AN GUANGZHOUTHANK YOU
Chinese
← → Navigate · F Fullscreen