15. Real Use Cases and Workflows

In this chapter no new theory. Just where agents are changing real work, with patterns you can adopt. Think of it as a menu to pick your next project from.

15.1 Coding and software development

Intensive pair programming

Tools: Claude Code, Cursor, Aider, Windsurf.

What changes: the speed of writing "boilerplate" code (CRUD, integrations, mechanical refactors) plummets. Devs focus on architecture, edge cases, code review.

Typical pattern:
1. Write a spec (the "what" and "why").
2. The agent proposes plan + diff.
3. You review, correct, iterate.
4. Tests run automatically.
5. Commit.

Realistic result: 2-4x productivity on standard tasks, marginal on truly new problems.

Automatic code review

Agent that runs on every PR and comments on:
- Potential bugs.
- Patterns inconsistent with the rest of the codebase.
- Missing tests.
- Security issues (SQL injection, hardcoded secrets).

Examples: GitHub Copilot Code Review, Anthropic /ultrareview, CodeRabbit.

Debug and incident response

Agent that, given a log or stack trace:
- Identifies the probable cause.
- Searches relevant code.
- Proposes a fix.

In on-call, reduces "time to first hypothesis" from 30 minutes to 1.

Migration and refactor

Real examples: Python 2 → 3, AngularJS → React migration, monolith → microservices.

Approach:
1. The agent analyzes the codebase, maps patterns.
2. Proposes a migration strategy.
3. Executes small steps, with tests at each step.
4. Human reviews, approves.

For big projects, agents like Devin, Coursive, or dedicated frameworks can work autonomously for days.

15.2 Customer support

Automatic Tier-1

Agent handling ~70% of standard questions:
- FAQ.
- Order status.
- Password reset.
- Data change.

When it doesn't know, escalate to a human with context already prepared.

Pattern:
- RAG on the company knowledge base.
- Tools for CRM, order system, billing.
- Conversation memory for continuity.
- Confidence threshold: below X, escalate.

Triage and classification

Agent reading incoming tickets and:
- Classifying (billing, tech support, sales).
- Assigning priority.
- Routing to the right team.
- Suggesting the first response to the human agent.

Reduces triage time by 80-90%.

Conversation summary

After a long conversation, the agent produces:
- Summary.
- Action items.
- Customer sentiment.
- Follow-up suggestions.

15.3 Research and analysis

Deep research

Tools like ChatGPT Deep Research, Perplexity Pro, Gemini with Workspace.

Pattern:
1. Complex question ("analyze the X market over the last 5 years").
2. Agent navigates the web, reads dozens of sources, cross-checks.
3. Produces a structured report with citations.

Work that required 1-2 days of a junior analyst is done in 30 minutes, with quality sufficient for first draft.

Data analysis

ChatGPT Code Interpreter, Claude with code_execution tool, Hex Magic, Julius AI.

Pattern:
1. Upload a CSV/Excel.
2. Explain what you want to understand.
3. The agent writes Python, executes, produces charts.
4. Continue the conversation: "now segment by region", "do statistical test", "export Excel".

For exploratory analysis it's a revolution.

Document review

For legal, finance, due diligence:
- Upload a set of contracts / documents.
- Agent extracts specific clauses, flags anomalies, compares with templates.
- Produces a checklist to review.

Tools: Harvey, Hebbia, Legora (legal); Hebbia, Anvilogic (finance/security).

15.4 Writing and content

Long-form writing

Pattern that works:
1. Clear brief: topic, audience, tone, length.
2. Outline before text.
3. Iterations on the outline.
4. Section-by-section expansion.
5. Final editing (human or assisted).

For blog posts, articles, guides: half a day of human work reduces to an hour of review.

Agent that every morning:
- Reads the sources you follow.
- Synthesizes into 5 bullets.
- Sends via email.

Basic setup: 100 lines + cron + LLM.

Localization

Translation + cultural adaptation of web content, manuals, e-commerce. Specialized agents with corporate glossaries reach final-editing quality, no longer rebuilding.

15.5 Operations and automation

Email management

Always-on agent (Ch. 4) that:
- Classifies incoming emails.
- Replies to repetitive ones (automatic or with draft).
- Extracts action items into a task manager.
- Flags urgent ones.

Tools: Superhuman AI, Shortwave, custom agents on Gmail API.

Calendar and scheduling

Agent that, given a goal ("meeting with Marco and Giulia by Friday"), finds slots, sends invites, manages conflicts, reschedules.

Consumer tools: Reclaim, Motion. For companies: custom agents on Outlook/Google Calendar.

Process automation

Business-as-usual workflows with agents that orchestrate heterogeneous steps:
- Customer onboarding: read docs, validate them, create account, send welcome.
- Procurement: request → quotes → comparison → order.
- Reporting: collect data from N sources, format, send.

Pattern: orchestrator + specific tools for each system.

15.6 Sales and marketing

Lead enrichment

Agent that, given a raw lead (email, company):
- Searches public info.
- Profiles the company (sector, size, buy signals).
- Suggests outreach angle.

Tools: Clay, Apollo, Crystal.

Personalized outreach

Generation of 1:1 messages based on:
- Prospect profile.
- Your product.
- Applicable use case.

Caution: without human touch, falls into spam. Best practice: AI does the draft, human refines.

Content velocity

From one cue, the agent produces: LinkedIn post, X thread, blog post, newsletter. Each channel with adapted tone.

15.7 Education and training

Personalized tutors

Khanmigo (Khan Academy), Duolingo Max, corporate GPT-tutors.

Pattern:
- Student asks.
- The agent doesn't give the answer, asks Socratic questions.
- Adapts difficulty to the level.
- Keeps memory of progress.

Corporate onboarding

New hires chat with an agent that has access to all internal documentation. Typical questions ("how do you do X?") answered 24/7 without bothering colleagues.

Simulated training

Agents impersonating difficult customers, candidates in interview, crisis situations. Trainees practice safely.

15.8 Healthcare (with caution)

Cases that work today:
- Clinical documentation: listening to a consultation, generating SOAP notes, ICD coding. Tools: Abridge, Suki, Nuance DAX.
- Primary triage: chatbot directing to specialist or ER. Under clinical supervision.
- Paper research: synthesis of medical literature for the clinician.

Cases that don't work today:
- Autonomous diagnosis: huge risks, complex regulation.
- Therapeutic decisions: AI assists, doctor decides.

Rules: in EU the AI Act classifies many medical applications as "high risk" → requires specific certifications.

15.9 Legal and compliance (with caution)

Cases that work:
- Contract review: clause extraction, comparison with templates, flag anomalies.
- E-discovery: analysis of large document volumes for litigation.
- Legal research: case law search, judgment summaries.
- Drafting standard clauses: lawyer refines.

Caveat: AI can invent case law. Mandatory verification. Famous cases of lawyers sanctioned for presenting non-existent rulings generated by ChatGPT.

15.10 More "agentic" cases

Examples of products pushing the autonomy level:

Devin (Cognition): coding agent that works autonomously for hours, completing complex tasks.
OpenAI Operator / Anthropic Computer Use: agents that use browser/desktop like a human (see screen, click, type).
Replit Agent: build full-stack apps from prompts.
AutoGPT, BabyAGI (early): first attempts at generalist agents, with results more demonstrative than productive.
Aria (Opera), Arc Search: native-AI browsers.

They are often still impressive demos that betray fragility in production. The direction is clear, the maturation speed isn't.

15.11 Cross-cutting patterns

Regardless of domain, winning workflows share:

Human-in-the-loop on costly decisions. AI does most, human validates the critical.
Structured briefs instead of free prompts.
Concrete tools to talk to real systems (CRM, DB, API).
RAG to make the agent speak with authority on specific data.
Context memory so as not to repeat setup every time.
Measurement of output and outcome.
Disclosure that it's AI-generated when relevant.

15.12 When NOT to add agents

Worth remembering:

If the workflow is simple and codified, traditional automation is better.
If the error is unacceptable and there's no human verification, careful.
If the data is too sensitive and you don't have dedicated infra, wait.
If costs aren't justified, don't scale.

AI is a multiplier, not a replacement of judgment.

15.13 To pick as your next project

If you're starting and want a project to do to learn, here's a list in order of ease:

Q&A bot on your PDF (RAG + chat). 1-2 hours. Ch. 7, 10.
Automatic summary of your day's emails. 2-3 hours. Email tool + LLM.
CSV analysis agent. 3-4 hours. Code interpreter pattern.
Customer support on company FAQ. 1-2 days. RAG + dev deploy.
Coding agent specialized for your stack. 1-2 weeks. Claude Agent SDK + custom tools.

Starting to do > reading another 100 pages.

15.14 Key takeaways

Coding, support, research, writing, ops, sales, education: agents are changing everything.
Cross-cutting patterns: structured brief, tools, RAG, human-in-the-loop, measurement.
High-stakes domains (health, law, finance): assistance yes, autonomy no.
Impressive demo ≠ robust production. Always verify.
Start from a small, personal project. Living an end-to-end agent is more formative than 10 courses.

15.15 Common mistakes

"AI will solve X." Without concrete patterns and workflows, it doesn't.
Building the most ambitious agent on the first try. Frustration guaranteed.
Neglecting integration. Agent quality depends on the quality of tools and data you give it.
Launching without measuring real impact. "The user is happy" without data.
Not capitalizing on existing workflows. AI shines when it slips into processes you already do, not when you have to reinvent them.

Last chapter: the glossary to remember terms, and a list of resources to go beyond the guide.