Email. It's essential, obviously, but let's be honest, for most of us, it's just chaos, pure chaos. We talked last time about building the AI brain, the part that sorts through it all. Right, the classification layer. And that's huge, don't get me wrong. But thinking, that's only half of it. The real magic, the part that actually gives you back hours every week, that happens when your AI doesn't just read, but it acts.
We're basically giving the AI hands, you know, so it can manage your inbox all the time, 24 -7, sending replies, picking the right teams, clearing out all that junk. So that's the mission for this deep dive. We're getting into part two of that AI. Email Assistant Guide, focusing specifically on building out those action branches, the hands, as you said. Exactly, within the workflow. And we'll be mentioning NANN quite a bit. Just quickly for anyone unfamiliar, NANN is essentially a
no -code tool. It lets you build and automate these kinds of workflows, linking things like AI models to your Gmail. Yep, it's the glue. So the plan today is pretty straightforward. We're going to look at four key actions, kind of tiered by risk, actually. Low -risk auto -replies for common stuff. Then safe team notifications, especially for sensitive things like finance. After that, high priority draft creation where a human still needs to press send. And finally,
the easy one. Just cleaning up the inbox. Right. And we'll wrap up with how to actually turn it on the activation step, which is critical. And then some really cool ways to scale this up, including, you know, the costs involved. Sounds like a solid plan. Let's start with those auto replies. The customer support branch. This feels like where the biggest immediate time saving could be, right? Getting rid of repetitive answers. Oh, absolutely. Total elimination of that Q &A
grind. So once an email is classified as, say, a support request, the next step is you add an AI agent node. Okay, the agent node. That's where the thinking happens. Well, the writing happens. It needs to generate the reply content. So you configure it to see the original email's subject and body. It just pulls that info right from the Gmail trigger that kicked everything off. Makes sense. It needs the context. But how do we stop it sounding like, well, like a robot?
How do we make sure it represents the company properly? Ah, that's all down to the system prompt. This is absolutely crucial. Think of it less like a suggestion and more like the AI's hard -coded rulebook. It's job description, basically. You have to give it really specific rules like define the exact tone, be friendly, professional, empathetic, specify its expertise. You only know about the 2024 product guide, things like that. So you're building guardrails. What if it gets
asked something outside that guide? It can't just make something up for you. Exactly. That's maybe the most important guardrail. You put in a hard rule. If you do not know the answer, stop. Do not guess. Just tell the user, we've passed this to a human. Please reach out to support at company .com or whatever the real email is. And you even specify the sign off like AI fires AI assistant for total brand consistency. And technically, how does that reply actually get
sent? You'd pick a good conversational model. Something like Claude 3 .7 Sonnet is great for this. Then in the Gmail node itself, the key action is reply to a message. Why that specific action? Because it keeps the reply perfectly threaded within the original email conversation. It looks completely natural to the person who emailed you. Oh, and a pro tip. Disable the little N8N attribution footer in the settings. Keeps it looking professional. That system prompt really
does sound like the linchpin then. How critical is getting that right for making the AI act like a reliable company representative? Oh, it's everything. The prompt is the AI's script. It guarantees the brand voice and provides those essential guardrails. All right, let's climb that ladder of sensitivity you mentioned. Finance and billing emails definitely don't want AI auto -replying there. What's the safe play? Yeah, absolutely
not. The goal here is different. It's safe. professional human notification, instantly letting the right team, maybe finance at your company .com, know something needs their attention. But crucially, we do not want the AI generating any content here. So this branch actually skips the AI agent node entirely. Interesting. So you're deliberately not using the AI here, even though it's an AI workflow. Is it just to save a few pennies on
API calls? That's a tiny part of it, maybe. But the main reason, it's risk, compliance risk, accuracy risk. LLMs can hallucinate, right? You just cannot risk the AI misreading an account number or a payment amount or a due date. Finance stuff needs humanized, period. Accuracy and safety first. Right. So the potential downside of an AI mistake, a hallucination, is just way too high when money's involved. Makes sense. Totally. So technically it's simple. You just use a standard
Gmail send node. You're basically just forwarding the key info. You build a clean little report in the body of that email, the one going to your internal team. Use dynamic expressions to pull in, say, the sender's name, their email address, the original subject. Makes it super easy for the finance team to see what's up at a glance. So beyond saving a bit on API costs. The expert level thinking for avoiding AI generation in finance is really about safety and compliance.
Absolutely. Money requires human review. Notifications prioritize accuracy and safety over instant generated replies. Okay, now for that middle ground. High priority emails. Things that need a fast response, but definitely need a human to sign off before it goes out. How do we blend AI speed with human control here? Yeah, this is a really common need. The strategy is straightforward. Let the AI draft the response, but a human reviews potentially tweaks it and then hit send. So another AI agent
node. Yep. Another AI agent likely configured differently. Maybe you want a more formal or detailed tone for these high priority messages. The system prompt here is, again, absolutely key to getting that right. You know, it's funny, even doing this stuff regularly, I still wrestle with prompt drift sometimes trying to get that tone just right. It's amazing how one small change of the prompt can totally change. the AI's personality. Oh, tell me about it. It's incredibly sensitive.
That prompt basically sits the AI's entire character. We've seen people build wild system prompts just for testing. I remember one called Mr. Apocalypse. The goal was to make the AI brutally honest, even rude. Slash slightly. Wow. Shows how much control you have. But for a serious high priority response, you'd use a powerful model, maybe Claude Sonnet again, and a prompt that demands careful synthesis, maybe referencing specific internal policies if you have RG set up, which we can
talk about later. OK, so the AI writes this perfect context aware response. How does it get queued
up for the human review? it physically appear right the magic technical bit in the final gmail node for this branch you select the action create a draft simple enough but here's the detail that tricks people up constantly you must map the threaded from the original email trigger into the draft id field of that action hold on threaded not message id for folks may be less familiar with email apis why is that specific id so important what happens if you accidentally use message
id Good question. If you use MessageEyed, yeah, it'll probably create a draft. But often it's like disconnected. It just shows up as a brand new standalone email in your drafts folder, completely separate from the original conversation. But when you use the threaded... You're linking that AI -generated draft directly to the entire conversation. So when you open the original email, boom, the draft is sitting right there ready to go in context. You can just review it, maybe tweak a word, and
hit send. It keeps the flow perfect. So threaded is the key. Why is using the threaded so critical versus the message -eyed? Threaded links the draft to the whole conversation, making it ready for a quick context -aware send. Okay, let's hit the easiest one now, the promotional emails. Yeah. The junk mail branch. Yeah. This sounds less about complex actions and more about satisfaction. Chuckles. Yeah, this is the simplest but maybe the most satisfying part. Pure inbox cleanup,
reducing that cognitive load. Once an email gets classified as promotion, the action is super quick. Two steps. First, you maybe add a promotion label just for organization. Then immediately after, you use another Gmail node with the action mark a message as read. That's it. That's it. Just mark as read. It sounds almost too simple. Are there any hidden downsides? Does marking it read immediately cause any problems? Honestly, very few downsides for true promotional stuff.
The email still gets filed away by the label if you used one, so you can find it later if you really need to. The main goal here is psychological. How so? It stops that email from contributing to the unread count. You know, that little red badge that causes so much anxiety. Yeah. This just removes that low value interruption from your immediate attention. You basically set it and forget it. It tidies things up automatically.
So this sounds almost too simple. Are there any drawbacks to immediately marking promotional mail as read? Minimal risk, high satisfaction. The goal is removing low value interruptions from the main view. Right then. We've built out all four paths. The auto replies, the team notifications, the drafts, and the cleanup crew for promotions. We have this whole sophisticated workflow sitting there, but it's not actually doing anything yet,
is it? Exactly right. Everything we've described so far only runs if you go into NAN and manually click execute workflow, which isn't very automated. No. The final absolutely essential step is you go to the top right corner of the NANN workflow editor and you find the toggle switch that says inactive. You have to flip that switch to active. Yeah, the big green switch. That's the one. That tells the system, OK, now you're live. Start
checking Gmail automatically. Usually it checks every minute by default for new emails and run this process. OK, so it's active. It's checking every minute. How do we know it's working correctly? Like, how do we check if it's actually sending emails down the right paths? Good call. You need to monitor it, especially at first. Inside NEN, there's a tab called Executions. That's your command center, your audit log. Every time the workflow runs because a new email came in, it
logs an execution. You can click into each one and visually see the path it took. Did it go down the support branch? Did it correctly identify a finance email and just send the notification? It's how you verify and troubleshoot. When people first build something like this, what's the most common, maybe... slightly embarrassing mistake they make right after building it that stops it from working? Honestly, forgetting to flip
that switch. They build this amazing thing, spend hours on it, and then wonder why nothing's happening. It's just sitting there, inactive. So what is the most common, simple mistake people make right after building the workflow? Forgetting to switch the workflow toggle from inactive to active, it's a manual setup detail. Okay, so once you have that basic system humming along nicely, filtering emails, sending replies, creating drafts,
then you can start thinking bigger. How do we upgrade this from just an assistant to more like a chief of staff for your inbox? Scaling up. I like the sound of that. Where do we start? What are the big enhancements? Okay, three key things come to mind. First, build a proper logging system. Think of it like the black box recorder
for your workflow. How does that work? At the end of every single one of those action branches we just discussed, the reply, the notification, the draft, the cleanup, you add one final node, maybe a Google Sheets node or a database node if you prefer. And in that node, you log key details, the timestamp. which category the email fell into, maybe the sender. And if an AI response was generated, log the actual response text. This creates a permanent record and audit trail.
You can see exactly what the AI is doing over time. Invaluable for improvement. That sounds powerful for tracking. What's next? Yeah. How do we make the AI itself smarter? Right. Connect knowledge bases. This is where you implement RA. That's retrieval augmented generation. RA. We hear that term a lot. Break it down simply. What does RA let the AI do? Basically, RIG lets your AI look things up in your own documents
before it answers. Instead of just relying on its general knowledge from its training data, it can consult your specific internal help docs, product manuals, policy guides, whatever you give it. Ah. So it stops guessing about company -specific stuff. Precisely. It uses a vector store that's a special database, like Supabase or Pinecone, to hold your documents in a way the AI can quickly search. It finds the relevant passages from your info and uses that to construct
the answer. Massive quality improvement. Okay. Logging for monitoring, RAG for accuracy. What's the third? A -B testing models. Don't just pick one AI model and stick with it forever. Models evolve, prices change. Set up a simple splinter node early in the workflow. Send, say, 50 % of incoming emails to a cheaper, faster model, maybe like GPT 4 .1 Mini. Send the other 50 % to a more powerful, maybe slightly more expensive
model, like Claude 3 .7 Sonnet. Then... use that logging system we just talked about log which model handled which email and maybe even track the quality or if a draft needed editing over time you get real data on which model gives you the best bang for your buck for different types of emails whoa hang on if you combine that logging with the a b testing You could scale that across
a whole company. Use the logs to scientifically figure out the absolute best, most cost -effective model for every single different email task. That's genuinely data -driven efficiency right there. Exactly. Optimizing cost and quality based on real results. How quickly does connecting a knowledge base with our ag actually change the quality? Is it noticeable right away? Immediately. The agent shifts from potentially guessing to actually looking up facts from your own internal
docs. It's night and day for accuracy on specific topics. OK, but it's tech. Things will go wrong sometimes. Yeah. What are the usual problems people run into and how do they fix them quickly? Yeah, inevitably there's some turbulence. We usually see kind of a big four. First is misclassification emails going down the wrong path. The fix there is usually refining those category descriptions in your initial classifier. Add more specific keywords, maybe better examples. Okay, what else?
Generic responses. If the AI's replies sound bland or unhelpful, that points straight back to the system prompt for that AI agent. You need to inject more detail, more personality, more specific company context. Make the prompt richer. We talked about drafts earlier. What if the drafts aren't linking properly to the original email thread? 99 % of the time, that's the threaded issue again. Go back to that create draft action
in the Gmail node. Double, triple check that you've correctly mapped the threaded from the trigger into the draft ID field, not the message ID. That's almost always the culprit. Got it. Threaded. And the fourth common issue. Credentials errors or the workflow just stopping. This could be simple, like needing to reauthenticate your Google connection in N8NN. Or it could be you've run out of API credits with your AI provider,
like OpenRouter or OpenAI. Always check your billing dashboard if things just stop working. So if a user's workflow just dies unexpectedly, what's the very first thing they should check before diving into complex debugging? Honestly, confirm that active toggle switch is still green. It sounds silly, but sometimes things can get accidentally deactivated. It's the simplest possible fix for what looks like a dead workflow. This all sounds incredibly powerful. A 2047 AI assistant
managing email must cost a fortune, right? Let's talk ROI. What's the actual cost here? You'd be surprised. It's actually very accessible. Let's break it down. You've got your base platform cost, so the NAN cloud plan you'd likely need is around, say, $25 a month. That's fixed. Okay. Then you have the variable cost. the AI API calls. This is pay as you go. But the key is, remember how we use cheaper, faster models like GPT 4 .1 mini for that initial high volume classification
step? Right. Most emails just get sorted. Exactly. That keeps the cost way down. So even if you're processing, let's say, 100 emails every single day. Yeah. Give me a real world number for that kind of volume. With this kind of optimized setup, you're probably looking at only about... 50 cents to maybe a dollar per day in total API fees. Seriously, that low? Yeah. So add that up for a month, maybe $15 to $30 max for the AI usage.
Your total all -in cost per month for this whole system, realistically, somewhere around $40 to $55. Okay, hold on. $55 a month sounds amazing, almost too good to be true, but that's just the software and API cost. What about the human time? Isn't there a hidden cost in setting this up,
tweaking the prompts? fixing things when they break that's a really fair point yes there's an initial time investment you got to build it right maybe budget i don't know three to five hours for the initial setup and the first week of watching it tuning it okay but once it's stable it really does mostly run itself maybe occasional tweaks so you weigh that initial time cost and the 55 a month against the time it saves you if this thing genuinely handles tasks that used
to take you five maybe 10, even more hours every single week. Yeah. That $55 looks pretty insignificant pretty quickly. The ROI is usually exceptional, often within the first month or two. And that minimal running cost, is that mostly because we're being smart about using cheaper AI for the bulk classification work? Yes, absolutely. Optimizing model choice by task is the key. It keeps those API costs extremely low while still using powerful models where you really need quality.
Wow. OK, so we've really seen this technical guy transform, haven't we? It's gone from just sorting emails into a genuinely operational system, one that sends replies, preps drafts for you, alerts the right people and just cleans up the clutter. Yeah, it's about building structure. This AI assistant, when set up right, it delivers organization, sure, but also like genuine peace
of mind. And yeah, critically, it really does give you back those five or 10 hours every week, the gap between email chaos and having this like. personal 247 chief of staff. It's just a bit of setup time and that's surprisingly small monthly cost. So for everyone listening, maybe take a minute after this to think about your own email habits. Where are those repetitive tasks? Where does all the time go? Start thinking about which action branches you'd build first. And that brings
us to our final thought. A bit provocative, maybe. If you actually manage to delegate 10 hours of email work every single week using a system like this, what's the one big, important, high -value project you'd finally have the bandwidth, the focus, the time to really dedicate yourself to? What becomes possible then? At UTR, we'll music.