Morning 1 of 3 · today your mail AI assistant answers a mail you send · the Enron mail case
Frank Huettner · 9 November 2026
Today: six blocks and their outcomes
Before the break: n8n, and your first workflow
1Why n8n, and the mail caseWhat you will build today
2WS2.1: your first n8n workflowYour Triage flow, tested on five mails
Break · 20 minutes
After the break: your mail AI assistant runs, and you make it better
3Your mail AI assistant answers your mailn8n workflows WS2.3 and WS2.4Kaminski's answer in your SKKU inbox
4Challenge: all five test mails rightYour score, and the best prompt in the room
5Let's look inside WS2.3, and make it oursYour rule stars an urgent mail in Gmail
6Why we put the human in the loop in WS2.4Why a person approves every draft
1
Why n8n, and the mail case
Why n8n, when you already have Langflow?
A Langflow agent answers when it is asked. A company needs more. It needs someone in the driver's seat.
A trigger starts it
a new mail
a form
a chat
a timer
Connects
hundreds of apps, ready to use
anything else by HTTP, like our Langflow
Checks, retries, human in the loop
checks before it acts
tries again when something fails
human in the loop: waits for a person's OK
Runs and data, built in
every run saved, each step visible
a database for small to medium data sets
Langflow is the brain, n8n is the muscle
Each one does what it was built for.
Langflow: the brain
came with the LLMs
thinks in Python
good for AI and data science: models, prompts, regression
n8n: the muscle
more mature, built for business automation
thinks in JavaScript
good for the web and communication: mail, forms, other apps
Today n8n calls your Langflow flows: Langflow proposes, n8n checks and acts.
Who is behind n8n: the best of both worlds
The company n8n GmbH, in Berlin, is behind n8n.
Its source code is public.
You may run it yourself, for free, also in a company.
Future-proof
a company maintains it
no lock-in: your workflows are files, and you can run n8n yourself
Support
professional support and hosting, for a fee
or the community, for free
Self-hostable
on your own computer, as today in kingo-pod
then your data stays with you
The licence is "fair-code": free to use, also inside a company, but not to sell n8n itself as a service.
Why n8n talks to a database in our setup (not to Gmail as usual)
Any mail provider
n8n has ready nodes for Gmail and Outlook
we use plain mail (IMAP): it works with Naver Works too
PostgresDB in the middle: one place for all mail
every mail is a row: n8n reads and writes rows, not Gmail
the search uses the same rows: no copy in ChromaDB
The mail service guards against accidents
In n8n, Kaminski's contacts have real addresses
any mail to them lands in your own test Gmail account
Align your team around principles for successOne place for the data. Ideally, you decide first where the data lives, and who may read or write it. Here: everything in PostgresDB; n8n writes, the Langflow flows only read.
The n8n editor: where things are
123456
Overview: your workflows, credentials, executions and data tables, one tab each.
··· next to the name: Import, Download, Settings.
Editor: where you build the workflow.
Executions: every saved run of this workflow. Open one: each node with what came in and what went out.
Publish: from then on, the trigger works by itself. Green dot: live. Yellow dot: your changes are not live yet.
+ on the right: a new node (or the key N). Shift+S: a sticky note.
Open n8n now: go to localhost:5678, click Overview, open WS2.0 - Embed Mails.
Double-click a node: in on the left, out on the right
n8n workflow WS2.0 - Embed Mails, node 3 transform_chunks_in_batches, in a run by hand*
12345
Input: what came in from the node before, here get_mails_to_embed: 1 item, one mail.
Schema, Table, JSON: the same items, three ways to look at them.
Parameters: what the node does. Here JavaScript: one text per mail, in groups of 32.
Execute step: runs only this node, on the input on the left.
Output: what goes out to the next node: 1 item with texts and metadata.
* Your list of WS2.0's runs is nearly empty: in homework step 10.4 you set it to keep only failed runs. To see a run like this one: unpublish WS2.0, send a mail to your test Gmail account, click Execute workflow, then publish again.
Every n8n workflow: a trigger, then nodes
A trigger starts it
A timer (e.g. WS2.0 - Embed Mails)
An external event (e.g. WS2.3 - Triage New Mail)
A page is opened (e.g. WS2.4 - Approvals Page)
You click Execute workflow in the editor (e.g. WS2.1 - TEST - Triage)
Langflow runs when you call it.
Many items
a list of items (an "array") goes from node to node
WS2.0: one item per mail, up to 320 in a run
each node works on every item of the list
Langflow sends one message at a time through the whole flow.
Left to right, top before bottom
from the trigger to the last node
when it splits: the upper branch runs first
never back, unless you loop on purpose
Langflow does not care where a component sits: only the connections count.
Runs you can open
double-click a node in a run
left: what came in
middle: what happens in the node
right: what went out
Langflow shows the last run.
The case: Vince Kaminski's mailbox
Kaminski ran the research group at Enron.
His mailbox from spring 2001 is public (released after Enron's collapse at the end of 2001 by the US energy regulator (FERC) during its investigation).
The emails and the names are real.
In your test Gmail account every address ends in .example; n8n and Langflow see the real ones. No mail reaches these people.
What comes in: questions, meetings, papers, students, hiring, and a lot of FYI.
Some mails are confidential, like resumes. The mail AI assistant may sort them, but never quote them.
3,244real mails in your test Gmail account, January to April 2001
451of them confidential: Draft Reply, the Langflow flow that writes drafts, cannot read them
140test mails of May 2001, newer than your mailbox: five of them test your Triage flow today
Nobody appreciates cleanly sorted files, until they find themselves in chaos
n8n · your workflows at the end of today
WS2.0 - Embed Mailshomework
WS2.1 - TEST - Triagetoday
WS2.2 - Ask Mailboxextra
WS2.3 - Triage New Mailtoday
WS2.4 - Reply with Approvaltoday
WS2.4 - Approvals Pagetoday
WS2.5 to WS2.9later
Search for “WS2.1” in the tabs Workflows and Data tables: the test and its two tables come up.
Workflow
WS2.<step> - <purpose> WS2.3 - Triage New Mail
A test
WS2.<step> - TEST - <what> WS2.1 - TEST - Triage
The test's data tables
WS2.1 - TEST - Triage - Cases WS2.1 - TEST - Triage - Runs
Files
small letters and _ ws2_1_test_triage_cases.csv
Nodes
what the node does, first fetch_triage · post_result · if_confidential
Your team
your team's name in place of WS2.<step> Blue - TEST - Intake
Shift+S in n8n: a sticky note with what the workflow does, your name, the date.
2
WS2.1: your first workflow
Ready? Three checks
kingo-pod runs: in the terminal, in your kingo-pod folder, ./kingo up. n8n opens at localhost:5678.
Your test Gmail account is open in another browser tab.
You downloaded ws2_s1.zip and unzipped it: the folder ws2_s1 holds the workbook, these slides, two CSV files and the folder n8n.
Everything else you set up in the homework.
Before you build: two data tables
What is a data table?
A table that lives in n8n, like a small spreadsheet.
Why two?
Cases: the five test mails, each with the action expected
Runs: empty for now; in block 4 it gets one line per test run
Important: everyone does this now, first. Your first workflow, WS2.1, reads its test mails from Cases. Without the two tables, it cannot run.
In the workbook: block 2, Make the two data tables, from the two CSV files in your folder ws2_s1.
n8n can reach out to Langflow with so-called HTTP requests
n8n sends the mail to your Langflow flow Triage and gets its answer back.
Align your team around principles for successMake a CONTRACT about what goes in, what comes out. Ideally, you decide this before you build: what is the structure of the input to a receiving flow, and what is the structure of its answers. Here: the Langflow flow Triage gets one mail and answers with:
action: calendar, reply, forward or archive
events: what the mail asks for or delivers, such as a meeting
deadline: as written in the mail
why: the reason for the action
checks_ok: did the answer pass the fixed checks?
problems: what the checks found
n8n then reads exactly these.
Your 1st n8n workflow will test the Langflow flow Triage
How many of five test mails does your Langflow flow Triage sort right?
Align your team around principles for successOne task per piece. Ideally, you give each piece one task. For example, the Langflow flow Triage only sorts. It doesn't reply or anything else. Breaking it down makes it easy to test and to reuse.
Build WS2.1: a trigger, then four nodes
Createa new workflow, named WS2.1 - TEST - Triage
Adda Manual Trigger, then the four nodes of WS2.1, one after the other
RunExecute workflow. In Cases, Triage's answer stands next to the expected one.
In the workbook: block 2, WS2.1: every step, and every setting to copy.
✓ Done: your first n8n workflow calls your Triage flow and tests it.
Extra, if you are fast: build WS2.2 - Ask Mailbox
A chat that searches your mail: the playground of homework step 13, now in n8n.
1Chat Triggertrigger_chat
2HTTP Request → Langflowfetch_ask_mailbox
3Edit Fieldsset_answer
trigger_chat
your question, typed into the chat
fetch_ask_mailbox
asks the Langflow flow Ask Mailbox
set_answer
takes out the answer, for the chat
What you see
What is Steve Leppard's email address?
Steve Leppard's email address is steven.leppard@enron.com. This comes from your own email of January 31, 2001 ("Re: Book order") … she had previously tried steve.leppard@enron.com, and that address bounced back.
In the workbook: block 2, Extra: build WS2.2 - Ask Mailbox: every step, and every setting to copy.
Extra, if you are fast: let an AI chat write the SQL
Who wrote to Kaminski most often? The answer is in PostgresDB. An AI chat writes the query in SQL, the language of databases: you do not need to know it.
Ask an AI chatCopy the request from the workbook into ChatGPT, Claude or another AI chat.
Paste the queryinto a Postgres node, in a new workflow WS2.2 - Top Senders.
Run and checkExecute workflow: ten senders. Does the list make sense? If not, tell the AI chat what is wrong.
The credential may only read and write what n8n needs, so a wrong query cannot break the mailbox.
In the workbook: block 2, Extra: let an AI chat write the SQL: the request to copy, and every step.
Break · 20 minutes
Your first workflow runs. After the break: your mail AI assistant answers your mail.
3
Your mail AI assistant answers your mail
The path of your mail
Your SKKU maila question, to your test Gmail account
PostgresDBthe mail service copies it in, and n8n wakes up
WS2.3 - Triage New Mailtakes the mail, asks the Langflow flow Triage
WS2.4 - Reply with Approvalasks the Langflow flow Draft Reply for a draft, checks it
WS2.4 - Approvals Pageyou approve
Your SKKU inboxKaminski's answer
Importthree ready workflows from the folder n8n
ConnectWS2.3 to WS2.4: WS2.3 hands every mail that needs a reply to WS2.4
Publishboth WS2.4 workflows first, then WS2.3: WS2.4 must already run when WS2.3 calls it
In the workbook: block 3: the three files, every click, and the address of the approvals page.
Try it: a mail from your SKKU account
your SKKU mailbox
Steve's address
You to your test address
Dear Vince, what is Steve Leppard's email address? I need to reach him.
later, after your approval
Vince J Kaminski to you
Kaminski's answer, drafted from his own mails, approved by you.
localhost:5678/webhook/approvals
Drafts waiting for you
To: your SKKU address Subject: Re: Steve's address
The draft of Draft Reply (Enron), checked by WS2.4
Open the approval form
Send from the SKKU address you allowed in homework step 6, and ask a question. A mail that only informs is archived: no draft. The mail to copy: workbook, block 3.
✓ Done: your mail AI assistant answers mail, and you stay in control.
A ready workflow to import: WS2.1 - TEST - Triage - Improved. New, compared to your WS2.1:
1 · Settings
Which flow, how many rounds, which part of the answer to compare: in set_test_settings, or on its page.
2 · Rounds
Each test mail once per round, three rounds: one round can be lucky.
3 · One line per run
In Runs: what you changed, the score of each round, the seconds, every answer.
4 · A dashboard
Better than before? Which mail broke? A page where you set the test and run it.
3–4 / 5
Our rounds on 2 and 8 Nov, model GLM 5.3 Flash: 3 or 4 of 5, once 5. Can you make it 5 in every round?
In the workbook: block 4: every step, and a question to check yourself.
Make it better: the prompt
In Langflow, open your flow Triage (Enron), then its component Prompt Template. The actions are described there:
"calendar" a meeting, call, visit or interview is
asked for or set: to Shirley Crenshaw …
"reply" someone wants information, an answer
or a decision from Kaminski …
"forward" someone wants something done …
In Langflow: change only the prompt, then save the flow. On the page of WS2.1 - TEST - Triage - Improved: write what you changed, then Run the test.
Read test mail 30: its whole text is in the column mail of the data table Cases. Why does the model say reply, not forward?
Who has 5 of 5 in all three rounds? Read the prompt aloud.
Did your fix for one mail break another?
Is five enough to tell a better system from a worse one?
Test your project's Langflow flows from n8n, too
Copy WS2.1 - TEST - Triage - Improved, give it your test cases and your flow: it tests every Langflow flow of your team's project.
Run it again after every change, and compare in its dashboard.
It also avoids the Langflow 1.8 memory bug. In the Playground, chats with the same name share one memory, even across flows: one test can remember the one before. From n8n, every call of the test starts with an empty memory, a session of its own.
In the workbook: block 4, Test your project's Langflow flows from n8n, too.
✓ Done: you test your Triage flow the right way: three rounds, one line per run.
5
Let's look inside WS2.3, and make it ours
What WS2.3 did with your mail
Your mail from block 3, “Steve's address”, went through five steps in WS2.3.
Wake upYour mail lands in PostgresDB. PostgresDB tells n8n at once, and WS2.3 starts. Just in case, WS2.3 also looks every 15 minutes.
Take it onceWS2.3 notes the mail as taken. No other run handles it again, so it never gets two answers.
ProtectCard and bank numbers are masked before a model sees the mail. A mail that tries to give the assistant orders goes to a person.
Ask TriageYour Langflow flow from the homework reads the mail and proposes what to do.
ActA fixed rule first. Then the action in Gmail: a label, the archive, or a draft for your OK.
Open WS2.3, tab Executions, the run of your mail. Double-click each node: what came in, what went out.
In the workbook: block 5, Look inside WS2.3: each step and its nodes.
Triage proposes an action, n8n checks it, then acts
A model answers with text, and text can come out broken. So n8n checks before it acts.
The answer of your Langflow flow Triage to “Steve's address”, in Frank's run:
{
"action": "reply",
"why": "Frank Huettner asks Kaminski for Steve Leppard's email address, so a reply with the information is needed.",
"events": ["Request_Data"],
"deadline": "",
"checks_ok": true,
"problems": []
}
action: what to do. One of four: reply, calendar, forward, archive.
checks_ok: is the answer in order? The component Check Triage in your flow checks it: one complete answer, with one of the four actions.
Whether the action is right, the check cannot tell. Your test in block 4 measures that.
In order
n8n does what action says.
Not in order
No action. The mail gets the label Assistant/Check, and a person looks at it.
What WS2.3 does in Gmail: a rule first, then Triage's action
Your test Gmail account, inbox, after WS2.3Triage's action: what WS2.3 did
☆Dorothy DaltonAssistant/Confidentialmessage from Ken Ricemarked confidential: the rule, whatever Triage said
☆Jason SokolovPapersreply: WS2.4 drafts an answer, you approve it
☆lopezAssistant/CalendarRisk Systems Enhancement Mtg - EB50M03calendar: a label
★Lon DraperAssistant/ForwardStrip of Spread Optionsforward: a label and a star
☆Stinson GibnerJury Dutyarchive: out of the inbox, not deleted
☆any senderAssistant/Checkan answer not in order, or a mail that gives ordersnot in order: a person looks at it
Align your team around principles for successWhat a rule decides, not the model
Ideally: before you build, decide which decisions a fixed rule makes. A rule decides the same way every time, and you can check it.
Here: the mailbox marks some mails as confidential, such as a resume or an evaluation. Such a mail only gets its label, never a draft, whatever Triage proposes.
n8n never touches Gmail itself: it writes one row into the table changes, such as “star this mail”. The mail service does it in Gmail. Your own rule, next, works the same way.
Your turn: a rule of your own
In WS2.3 - Triage New Mail, a second branch from check_triage
Codecheck_triage
1Ifif_urgent
true2Postgrespost_star
if_urgent asks: does the subject say "urgent"? If so, post_star gives the mail a star in Gmail.
In the workbook: block 5, Your own rule: how to add the branch, and the settings to copy.
Publish WS2.3 again, then send “Urgent: call me back” from your SKKU account:
★YouUrgent: call me backyour rule
✓ Done: your own rule, running in a real system. A rule decides here, not the model: the same every time.
6
Why we put the human in the loop in WS2.4
Often, a little adjustment is all you need
The mail AI assistant's draft named the right person, as Kaminski did.
Incoming mail · From: John A. Cote · 11 May 2001 · Subject: UT Recruiting
Vince, Czesc,
Our group would like to recruit a new hire / intern from UT's energy studies program. Kristin thought you might have a contact in this program. (Any info would be helpful!) Dziekuje.
Kaminski's actual response · 15 May 2001
John,
You can contact Prof. Ehud Ronn.
Ehud I. Ronn, Professor of Finance, Director, Center for Energy Finance Education and Research, University of Texas at Austin · phone · mail
Vince
Draft of the mail AI assistant · the Langflow flow Draft Reply
John,
Czesc. Good to hear from you.
I do not have a contact in the energy studies program itself. My contacts at UT Austin are Ehud Ronn and Sheridan Titman … I will ask them about the program …
Vince
This draft saved some time. Possible extension: another loop that incorporates your adjustments.
Yanbo's five trust components, in your mail AI assistant
Component
Principle behind it
In your mail AI assistant
Input validation
What a rule decides
Before a model sees a mail, WS2.3 masks card and bank numbers. A mail that tries to give the mail AI assistant orders goes to a person. WS2.4 masks again before Draft Reply.
Process guardrails
What each part may do
The Langflow flows may only read the mailbox. Draft Reply never sees a confidential mail: the database hides it. Only your OK lets a mail go out.
Output filter
Make a contract about what goes in, what comes out
Each Langflow flow checks its own answer by fixed rules. Check Triage: one of the four actions. Check Draft: a greeting, no gap like [Name], addresses and numbers only from the mails. n8n acts only on an answer in order. You see Check Draft's result on the approval form.
Error handling
Where a person decides
n8n calls Langflow up to twice. Triage's answer not in order: a person looks at the mail. No draft, or no one to send it to: nothing is proposed, and the problem is noted.
Audit trail
One place for the data
n8n's Executions: every run that decides, node by node. The table outbox: each draft with its status (proposed, approved, sent), who approved it, and why.
The middle column: the principle behind each component. All six principles: in the appendix.
As with Yanbo, each Langflow flow checks its own answer. n8n checks the input and decides what happens.
What you built today
A mail AI assistant that answers your mail, after your OK
A better prompt for your Triage flow, measured in three rounds
Your own rule: an urgent mail gets a star in Gmail
The big idea: the model proposes, a rule checks, a person approves.
Appendix: six principles to align your team around
You saw each one today. Write down your team's own, before anyone builds.
About the pieces
1 · One task per piece
The Langflow flow Triage only sorts: easy to test (WS2.1) and to reuse.
2 · Make a contract about what goes in, what comes out
The Langflow flow Triage: fixed fields. n8n reads exactly these.
3 · One place for the data
The mails and the search in PostgresDB. n8n writes, Langflow reads.
About control
4 · What a rule decides
Not the model: a confidential mail gets a label, never a draft.
5 · What each part may do
Draft Reply cannot send. Only your OK lets a mail go out.
6 · Where a person decides
Nothing leaves without your OK. WS2.4 waits on a form.
For your project: the same six in ws2_rules.html in your download, with your team's spec at its end. A team that agrees on them first spends its time on the project, not on making its pieces fit.
Appendix: one contract that both tools can read
n8n and Langflow do not tell each other what a flow needs and gives. The standard way: write it down once, in a file, and check both sides against it.
JSON Schema
Describes one JSON object: which fields, which type, which values. One file per flow, and Langflow, n8n and the tests all check against it. Yanbo's "structured output with a schema" is the same idea.
MCP (Model Context Protocol)
For AI agents that choose their tools live, while they work. Each tool comes with a name, what it does, and a JSON Schema for in and out. Natural for us: a Langflow agent calls n8n workflows as tools. A fixed step, like n8n calling Triage, needs no MCP.
All four flows: contracts/flows.openapi.yaml (OpenAPI with JSON Schemas). Not needed in the course. Using a coding agent anyway, such as Claude Code? Ask it to write yours: the prompt is in the workbook.
Appendix: WS2.0 from your homework, node by node
n8n workflow WS2.0 - Embed Mails: makes new mails searchable, every minute
1Schedule Triggerschedule_every_minutestarts it every minute
2Postgresget_mails_to_embedthe mails not yet searchable
3Codetransform_chunks_in_batchesone text per mail, in groups of 32
4HTTP Request → OpenRouterfetch_embeddings_bge_m3one request per group
5Codetransform_vectors_to_chunkseach embedding back to its mail
6Postgrespost_chunkssaves them in PostgresDB
Node 3: why one text per mail
subject, sender, date and the mail's own words
the quoted older mails cut off: they are stored as mails of their own
Node 3: why groups of 32
node 4 sends one request to the model for 32 mails, not 32 requests
fewer and faster calls; a failed call repeats only its group