WEBVTT 00:00:00.000 --> 00:00:04.960 Email. It's essential, obviously, but let's be 00:00:04.960 --> 00:00:07.379 honest, for most of us, it's just chaos, pure 00:00:07.379 --> 00:00:09.699 chaos. We talked last time about building the 00:00:09.699 --> 00:00:11.759 AI brain, the part that sorts through it all. 00:00:11.859 --> 00:00:13.800 Right, the classification layer. And that's huge, 00:00:13.919 --> 00:00:16.940 don't get me wrong. But thinking, that's only 00:00:16.940 --> 00:00:20.019 half of it. The real magic, the part that actually 00:00:20.019 --> 00:00:22.660 gives you back hours every week, that happens 00:00:22.660 --> 00:00:24.699 when your AI doesn't just read, but it acts. 00:00:25.850 --> 00:00:28.190 We're basically giving the AI hands, you know, 00:00:28.210 --> 00:00:30.870 so it can manage your inbox all the time, 24 00:00:30.870 --> 00:00:33.810 -7, sending replies, picking the right teams, 00:00:34.030 --> 00:00:36.509 clearing out all that junk. So that's the mission 00:00:36.509 --> 00:00:38.570 for this deep dive. We're getting into part two 00:00:38.570 --> 00:00:42.000 of that AI. Email Assistant Guide, focusing specifically 00:00:42.000 --> 00:00:44.939 on building out those action branches, the hands, 00:00:45.020 --> 00:00:46.960 as you said. Exactly, within the workflow. And 00:00:46.960 --> 00:00:49.200 we'll be mentioning NANN quite a bit. Just quickly 00:00:49.200 --> 00:00:51.500 for anyone unfamiliar, NANN is essentially a 00:00:51.500 --> 00:00:54.299 no -code tool. It lets you build and automate 00:00:54.299 --> 00:00:56.659 these kinds of workflows, linking things like 00:00:56.659 --> 00:00:59.039 AI models to your Gmail. Yep, it's the glue. 00:00:59.200 --> 00:01:01.219 So the plan today is pretty straightforward. 00:01:01.420 --> 00:01:03.820 We're going to look at four key actions, kind 00:01:03.820 --> 00:01:06.680 of tiered by risk, actually. Low -risk auto -replies 00:01:06.680 --> 00:01:10.109 for common stuff. Then safe team notifications, 00:01:10.250 --> 00:01:12.329 especially for sensitive things like finance. 00:01:12.890 --> 00:01:15.750 After that, high priority draft creation where 00:01:15.750 --> 00:01:18.709 a human still needs to press send. And finally, 00:01:18.810 --> 00:01:21.750 the easy one. Just cleaning up the inbox. Right. 00:01:21.849 --> 00:01:23.849 And we'll wrap up with how to actually turn it 00:01:23.849 --> 00:01:26.170 on the activation step, which is critical. And 00:01:26.170 --> 00:01:27.969 then some really cool ways to scale this up, 00:01:28.049 --> 00:01:30.409 including, you know, the costs involved. Sounds 00:01:30.409 --> 00:01:32.530 like a solid plan. Let's start with those auto 00:01:32.530 --> 00:01:35.650 replies. The customer support branch. This feels 00:01:35.650 --> 00:01:37.969 like where the biggest immediate time saving 00:01:37.969 --> 00:01:40.150 could be, right? Getting rid of repetitive answers. 00:01:40.349 --> 00:01:43.760 Oh, absolutely. Total elimination of that Q &A 00:01:43.760 --> 00:01:47.120 grind. So once an email is classified as, say, 00:01:47.200 --> 00:01:49.640 a support request, the next step is you add an 00:01:49.640 --> 00:01:52.400 AI agent node. Okay, the agent node. That's where 00:01:52.400 --> 00:01:54.260 the thinking happens. Well, the writing happens. 00:01:54.319 --> 00:01:56.299 It needs to generate the reply content. So you 00:01:56.299 --> 00:01:58.579 configure it to see the original email's subject 00:01:58.579 --> 00:02:01.120 and body. It just pulls that info right from 00:02:01.120 --> 00:02:02.640 the Gmail trigger that kicked everything off. 00:02:02.859 --> 00:02:06.099 Makes sense. It needs the context. But how do 00:02:06.099 --> 00:02:09.960 we stop it sounding like, well, like a robot? 00:02:10.300 --> 00:02:12.020 How do we make sure it represents the company 00:02:12.020 --> 00:02:14.360 properly? Ah, that's all down to the system prompt. 00:02:14.620 --> 00:02:17.219 This is absolutely crucial. Think of it less 00:02:17.219 --> 00:02:20.039 like a suggestion and more like the AI's hard 00:02:20.039 --> 00:02:23.099 -coded rulebook. It's job description, basically. 00:02:23.379 --> 00:02:25.759 You have to give it really specific rules like 00:02:25.759 --> 00:02:29.280 define the exact tone, be friendly, professional, 00:02:29.620 --> 00:02:33.020 empathetic, specify its expertise. You only know 00:02:33.020 --> 00:02:35.759 about the 2024 product guide, things like that. 00:02:35.879 --> 00:02:37.659 So you're building guardrails. What if it gets 00:02:37.659 --> 00:02:40.310 asked something outside that guide? It can't 00:02:40.310 --> 00:02:43.129 just make something up for you. Exactly. That's 00:02:43.129 --> 00:02:44.810 maybe the most important guardrail. You put in 00:02:44.810 --> 00:02:46.969 a hard rule. If you do not know the answer, stop. 00:02:47.250 --> 00:02:49.870 Do not guess. Just tell the user, we've passed 00:02:49.870 --> 00:02:51.710 this to a human. Please reach out to support 00:02:51.710 --> 00:02:54.129 at company .com or whatever the real email is. 00:02:54.250 --> 00:02:57.110 And you even specify the sign off like AI fires 00:02:57.110 --> 00:03:00.750 AI assistant for total brand consistency. And 00:03:00.750 --> 00:03:02.750 technically, how does that reply actually get 00:03:02.750 --> 00:03:04.889 sent? You'd pick a good conversational model. 00:03:05.370 --> 00:03:08.389 Something like Claude 3 .7 Sonnet is great for 00:03:08.389 --> 00:03:11.370 this. Then in the Gmail node itself, the key 00:03:11.370 --> 00:03:15.189 action is reply to a message. Why that specific 00:03:15.189 --> 00:03:17.550 action? Because it keeps the reply perfectly 00:03:17.550 --> 00:03:19.430 threaded within the original email conversation. 00:03:19.889 --> 00:03:22.210 It looks completely natural to the person who 00:03:22.210 --> 00:03:25.129 emailed you. Oh, and a pro tip. Disable the little 00:03:25.129 --> 00:03:28.830 N8N attribution footer in the settings. Keeps 00:03:28.830 --> 00:03:31.169 it looking professional. That system prompt really 00:03:31.169 --> 00:03:33.659 does sound like the linchpin then. How critical 00:03:33.659 --> 00:03:36.280 is getting that right for making the AI act like 00:03:36.280 --> 00:03:39.139 a reliable company representative? Oh, it's everything. 00:03:39.319 --> 00:03:42.080 The prompt is the AI's script. It guarantees 00:03:42.080 --> 00:03:44.080 the brand voice and provides those essential 00:03:44.080 --> 00:03:46.120 guardrails. All right, let's climb that ladder 00:03:46.120 --> 00:03:48.400 of sensitivity you mentioned. Finance and billing 00:03:48.400 --> 00:03:50.960 emails definitely don't want AI auto -replying 00:03:50.960 --> 00:03:53.080 there. What's the safe play? Yeah, absolutely 00:03:53.080 --> 00:03:56.560 not. The goal here is different. It's safe. professional 00:03:56.560 --> 00:03:59.259 human notification, instantly letting the right 00:03:59.259 --> 00:04:02.199 team, maybe finance at your company .com, know 00:04:02.199 --> 00:04:04.860 something needs their attention. But crucially, 00:04:04.919 --> 00:04:07.419 we do not want the AI generating any content 00:04:07.419 --> 00:04:11.500 here. So this branch actually skips the AI agent 00:04:11.500 --> 00:04:13.900 node entirely. Interesting. So you're deliberately 00:04:13.900 --> 00:04:17.139 not using the AI here, even though it's an AI 00:04:17.139 --> 00:04:20.500 workflow. Is it just to save a few pennies on 00:04:20.500 --> 00:04:23.800 API calls? That's a tiny part of it, maybe. But 00:04:23.800 --> 00:04:26.819 the main reason, it's risk, compliance risk, 00:04:27.060 --> 00:04:30.959 accuracy risk. LLMs can hallucinate, right? You 00:04:30.959 --> 00:04:33.639 just cannot risk the AI misreading an account 00:04:33.639 --> 00:04:36.279 number or a payment amount or a due date. Finance 00:04:36.279 --> 00:04:39.199 stuff needs humanized, period. Accuracy and safety 00:04:39.199 --> 00:04:41.819 first. Right. So the potential downside of an 00:04:41.819 --> 00:04:44.339 AI mistake, a hallucination, is just way too 00:04:44.339 --> 00:04:46.670 high when money's involved. Makes sense. Totally. 00:04:46.790 --> 00:04:48.689 So technically it's simple. You just use a standard 00:04:48.689 --> 00:04:50.670 Gmail send node. You're basically just forwarding 00:04:50.670 --> 00:04:52.810 the key info. You build a clean little report 00:04:52.810 --> 00:04:55.050 in the body of that email, the one going to your 00:04:55.050 --> 00:04:57.930 internal team. Use dynamic expressions to pull 00:04:57.930 --> 00:05:00.170 in, say, the sender's name, their email address, 00:05:00.230 --> 00:05:02.589 the original subject. Makes it super easy for 00:05:02.589 --> 00:05:04.370 the finance team to see what's up at a glance. 00:05:04.670 --> 00:05:08.449 So beyond saving a bit on API costs. The expert 00:05:08.449 --> 00:05:10.769 level thinking for avoiding AI generation in 00:05:10.769 --> 00:05:13.110 finance is really about safety and compliance. 00:05:13.509 --> 00:05:17.250 Absolutely. Money requires human review. Notifications 00:05:17.250 --> 00:05:20.629 prioritize accuracy and safety over instant generated 00:05:20.629 --> 00:05:24.529 replies. Okay, now for that middle ground. High 00:05:24.529 --> 00:05:27.670 priority emails. Things that need a fast response, 00:05:27.829 --> 00:05:29.930 but definitely need a human to sign off before 00:05:29.930 --> 00:05:33.089 it goes out. How do we blend AI speed with human 00:05:33.089 --> 00:05:35.379 control here? Yeah, this is a really common need. 00:05:35.480 --> 00:05:38.060 The strategy is straightforward. Let the AI draft 00:05:38.060 --> 00:05:40.740 the response, but a human reviews potentially 00:05:40.740 --> 00:05:43.579 tweaks it and then hit send. So another AI agent 00:05:43.579 --> 00:05:46.100 node. Yep. Another AI agent likely configured 00:05:46.100 --> 00:05:48.639 differently. Maybe you want a more formal or 00:05:48.639 --> 00:05:50.759 detailed tone for these high priority messages. 00:05:51.000 --> 00:05:52.819 The system prompt here is, again, absolutely 00:05:52.819 --> 00:05:54.620 key to getting that right. You know, it's funny, 00:05:54.699 --> 00:05:56.740 even doing this stuff regularly, I still wrestle 00:05:56.740 --> 00:05:59.120 with prompt drift sometimes trying to get that 00:05:59.120 --> 00:06:01.620 tone just right. It's amazing how one small change 00:06:01.620 --> 00:06:04.110 of the prompt can totally change. the AI's personality. 00:06:04.470 --> 00:06:07.129 Oh, tell me about it. It's incredibly sensitive. 00:06:07.389 --> 00:06:10.970 That prompt basically sits the AI's entire character. 00:06:11.350 --> 00:06:14.329 We've seen people build wild system prompts just 00:06:14.329 --> 00:06:16.819 for testing. I remember one called Mr. Apocalypse. 00:06:17.279 --> 00:06:19.860 The goal was to make the AI brutally honest, 00:06:20.100 --> 00:06:24.519 even rude. Slash slightly. Wow. Shows how much 00:06:24.519 --> 00:06:27.439 control you have. But for a serious high priority 00:06:27.439 --> 00:06:29.720 response, you'd use a powerful model, maybe Claude 00:06:29.720 --> 00:06:31.959 Sonnet again, and a prompt that demands careful 00:06:31.959 --> 00:06:34.279 synthesis, maybe referencing specific internal 00:06:34.279 --> 00:06:36.920 policies if you have RG set up, which we can 00:06:36.920 --> 00:06:39.699 talk about later. OK, so the AI writes this perfect 00:06:39.699 --> 00:06:42.980 context aware response. How does it get queued 00:06:42.980 --> 00:06:45.379 up for the human review? it physically appear 00:06:45.379 --> 00:06:48.360 right the magic technical bit in the final gmail 00:06:48.360 --> 00:06:50.600 node for this branch you select the action create 00:06:50.600 --> 00:06:53.379 a draft simple enough but here's the detail that 00:06:53.379 --> 00:06:56.680 tricks people up constantly you must map the 00:06:56.680 --> 00:06:59.060 threaded from the original email trigger into 00:06:59.060 --> 00:07:02.360 the draft id field of that action hold on threaded 00:07:02.360 --> 00:07:04.899 not message id for folks may be less familiar 00:07:04.899 --> 00:07:08.019 with email apis why is that specific id so important 00:07:08.019 --> 00:07:10.139 what happens if you accidentally use message 00:07:10.139 --> 00:07:12.870 id Good question. If you use MessageEyed, yeah, 00:07:12.910 --> 00:07:15.230 it'll probably create a draft. But often it's 00:07:15.230 --> 00:07:18.730 like disconnected. It just shows up as a brand 00:07:18.730 --> 00:07:21.750 new standalone email in your drafts folder, completely 00:07:21.750 --> 00:07:24.069 separate from the original conversation. But 00:07:24.069 --> 00:07:27.050 when you use the threaded... You're linking that 00:07:27.050 --> 00:07:30.189 AI -generated draft directly to the entire conversation. 00:07:30.329 --> 00:07:32.589 So when you open the original email, boom, the 00:07:32.589 --> 00:07:34.829 draft is sitting right there ready to go in context. 00:07:35.110 --> 00:07:37.750 You can just review it, maybe tweak a word, and 00:07:37.750 --> 00:07:40.050 hit send. It keeps the flow perfect. So threaded 00:07:40.050 --> 00:07:43.170 is the key. Why is using the threaded so critical 00:07:43.170 --> 00:07:45.709 versus the message -eyed? Threaded links the 00:07:45.709 --> 00:07:47.730 draft to the whole conversation, making it ready 00:07:47.730 --> 00:07:50.370 for a quick context -aware send. Okay, let's 00:07:50.370 --> 00:07:53.430 hit the easiest one now, the promotional emails. 00:07:53.629 --> 00:07:56.189 Yeah. The junk mail branch. Yeah. This sounds 00:07:56.189 --> 00:07:59.350 less about complex actions and more about satisfaction. 00:07:59.709 --> 00:08:02.029 Chuckles. Yeah, this is the simplest but maybe 00:08:02.029 --> 00:08:04.930 the most satisfying part. Pure inbox cleanup, 00:08:05.129 --> 00:08:08.069 reducing that cognitive load. Once an email gets 00:08:08.069 --> 00:08:10.250 classified as promotion, the action is super 00:08:10.250 --> 00:08:13.629 quick. Two steps. First, you maybe add a promotion 00:08:13.629 --> 00:08:16.329 label just for organization. Then immediately 00:08:16.329 --> 00:08:19.350 after, you use another Gmail node with the action 00:08:19.350 --> 00:08:22.589 mark a message as read. That's it. That's it. 00:08:22.589 --> 00:08:24.689 Just mark as read. It sounds almost too simple. 00:08:24.850 --> 00:08:27.649 Are there any hidden downsides? Does marking 00:08:27.649 --> 00:08:30.310 it read immediately cause any problems? Honestly, 00:08:30.430 --> 00:08:32.990 very few downsides for true promotional stuff. 00:08:33.250 --> 00:08:35.570 The email still gets filed away by the label 00:08:35.570 --> 00:08:37.669 if you used one, so you can find it later if 00:08:37.669 --> 00:08:41.299 you really need to. The main goal here is psychological. 00:08:41.820 --> 00:08:44.399 How so? It stops that email from contributing 00:08:44.399 --> 00:08:46.360 to the unread count. You know, that little red 00:08:46.360 --> 00:08:49.139 badge that causes so much anxiety. Yeah. This 00:08:49.139 --> 00:08:51.639 just removes that low value interruption from 00:08:51.639 --> 00:08:53.860 your immediate attention. You basically set it 00:08:53.860 --> 00:08:55.720 and forget it. It tidies things up automatically. 00:08:56.200 --> 00:08:58.700 So this sounds almost too simple. Are there any 00:08:58.700 --> 00:09:01.100 drawbacks to immediately marking promotional 00:09:01.100 --> 00:09:04.080 mail as read? Minimal risk, high satisfaction. 00:09:04.379 --> 00:09:07.460 The goal is removing low value interruptions 00:09:07.460 --> 00:09:10.600 from the main view. Right then. We've built out 00:09:10.600 --> 00:09:15.139 all four paths. The auto replies, the team notifications, 00:09:15.539 --> 00:09:18.100 the drafts, and the cleanup crew for promotions. 00:09:18.379 --> 00:09:20.700 We have this whole sophisticated workflow sitting 00:09:20.700 --> 00:09:23.360 there, but it's not actually doing anything yet, 00:09:23.399 --> 00:09:25.620 is it? Exactly right. Everything we've described 00:09:25.620 --> 00:09:28.500 so far only runs if you go into NAN and manually 00:09:28.500 --> 00:09:31.519 click execute workflow, which isn't very automated. 00:09:31.759 --> 00:09:34.440 No. The final absolutely essential step is you 00:09:34.440 --> 00:09:36.779 go to the top right corner of the NANN workflow 00:09:36.779 --> 00:09:39.679 editor and you find the toggle switch that says 00:09:39.679 --> 00:09:42.620 inactive. You have to flip that switch to active. 00:09:42.860 --> 00:09:45.679 Yeah, the big green switch. That's the one. That 00:09:45.679 --> 00:09:47.419 tells the system, OK, now you're live. Start 00:09:47.419 --> 00:09:49.480 checking Gmail automatically. Usually it checks 00:09:49.480 --> 00:09:52.399 every minute by default for new emails and run 00:09:52.399 --> 00:09:54.240 this process. OK, so it's active. It's checking 00:09:54.240 --> 00:09:56.419 every minute. How do we know it's working correctly? 00:09:56.639 --> 00:09:58.399 Like, how do we check if it's actually sending 00:09:58.399 --> 00:10:01.190 emails down the right paths? Good call. You need 00:10:01.190 --> 00:10:04.370 to monitor it, especially at first. Inside NEN, 00:10:04.529 --> 00:10:07.009 there's a tab called Executions. That's your 00:10:07.009 --> 00:10:09.470 command center, your audit log. Every time the 00:10:09.470 --> 00:10:11.590 workflow runs because a new email came in, it 00:10:11.590 --> 00:10:13.970 logs an execution. You can click into each one 00:10:13.970 --> 00:10:16.409 and visually see the path it took. Did it go 00:10:16.409 --> 00:10:18.889 down the support branch? Did it correctly identify 00:10:18.889 --> 00:10:21.110 a finance email and just send the notification? 00:10:21.470 --> 00:10:24.009 It's how you verify and troubleshoot. When people 00:10:24.009 --> 00:10:25.929 first build something like this, what's the most 00:10:25.929 --> 00:10:28.559 common, maybe... slightly embarrassing mistake 00:10:28.559 --> 00:10:31.240 they make right after building it that stops 00:10:31.240 --> 00:10:33.740 it from working? Honestly, forgetting to flip 00:10:33.740 --> 00:10:36.340 that switch. They build this amazing thing, spend 00:10:36.340 --> 00:10:38.960 hours on it, and then wonder why nothing's happening. 00:10:39.080 --> 00:10:41.740 It's just sitting there, inactive. So what is 00:10:41.740 --> 00:10:44.399 the most common, simple mistake people make right 00:10:44.399 --> 00:10:46.590 after building the workflow? Forgetting to switch 00:10:46.590 --> 00:10:48.750 the workflow toggle from inactive to active, 00:10:49.029 --> 00:10:51.870 it's a manual setup detail. Okay, so once you 00:10:51.870 --> 00:10:54.070 have that basic system humming along nicely, 00:10:54.210 --> 00:10:57.269 filtering emails, sending replies, creating drafts, 00:10:57.289 --> 00:10:59.529 then you can start thinking bigger. How do we 00:10:59.529 --> 00:11:02.549 upgrade this from just an assistant to more like 00:11:02.549 --> 00:11:06.789 a chief of staff for your inbox? Scaling up. 00:11:06.909 --> 00:11:08.529 I like the sound of that. Where do we start? 00:11:08.710 --> 00:11:11.029 What are the big enhancements? Okay, three key 00:11:11.029 --> 00:11:14.090 things come to mind. First, build a proper logging 00:11:14.090 --> 00:11:17.259 system. Think of it like the black box recorder 00:11:17.259 --> 00:11:19.779 for your workflow. How does that work? At the 00:11:19.779 --> 00:11:21.740 end of every single one of those action branches 00:11:21.740 --> 00:11:24.139 we just discussed, the reply, the notification, 00:11:24.480 --> 00:11:27.360 the draft, the cleanup, you add one final node, 00:11:27.559 --> 00:11:29.980 maybe a Google Sheets node or a database node 00:11:29.980 --> 00:11:32.860 if you prefer. And in that node, you log key 00:11:32.860 --> 00:11:36.799 details, the timestamp. which category the email 00:11:36.799 --> 00:11:39.679 fell into, maybe the sender. And if an AI response 00:11:39.679 --> 00:11:42.299 was generated, log the actual response text. 00:11:42.480 --> 00:11:44.940 This creates a permanent record and audit trail. 00:11:45.220 --> 00:11:47.299 You can see exactly what the AI is doing over 00:11:47.299 --> 00:11:49.679 time. Invaluable for improvement. That sounds 00:11:49.679 --> 00:11:52.000 powerful for tracking. What's next? Yeah. How 00:11:52.000 --> 00:11:54.220 do we make the AI itself smarter? Right. Connect 00:11:54.220 --> 00:11:56.080 knowledge bases. This is where you implement 00:11:56.080 --> 00:11:59.159 RA. That's retrieval augmented generation. RA. 00:11:59.379 --> 00:12:01.679 We hear that term a lot. Break it down simply. 00:12:01.779 --> 00:12:05.360 What does RA let the AI do? Basically, RIG lets 00:12:05.360 --> 00:12:07.480 your AI look things up in your own documents 00:12:07.480 --> 00:12:09.960 before it answers. Instead of just relying on 00:12:09.960 --> 00:12:11.659 its general knowledge from its training data, 00:12:11.860 --> 00:12:14.360 it can consult your specific internal help docs, 00:12:14.600 --> 00:12:17.039 product manuals, policy guides, whatever you 00:12:17.039 --> 00:12:20.110 give it. Ah. So it stops guessing about company 00:12:20.110 --> 00:12:22.610 -specific stuff. Precisely. It uses a vector 00:12:22.610 --> 00:12:24.690 store that's a special database, like Supabase 00:12:24.690 --> 00:12:26.970 or Pinecone, to hold your documents in a way 00:12:26.970 --> 00:12:29.570 the AI can quickly search. It finds the relevant 00:12:29.570 --> 00:12:31.970 passages from your info and uses that to construct 00:12:31.970 --> 00:12:35.049 the answer. Massive quality improvement. Okay. 00:12:35.149 --> 00:12:38.370 Logging for monitoring, RAG for accuracy. What's 00:12:38.370 --> 00:12:41.590 the third? A -B testing models. Don't just pick 00:12:41.590 --> 00:12:44.289 one AI model and stick with it forever. Models 00:12:44.289 --> 00:12:47.169 evolve, prices change. Set up a simple splinter 00:12:47.169 --> 00:12:50.970 node early in the workflow. Send, say, 50 % of 00:12:50.970 --> 00:12:53.309 incoming emails to a cheaper, faster model, maybe 00:12:53.309 --> 00:12:57.049 like GPT 4 .1 Mini. Send the other 50 % to a 00:12:57.049 --> 00:12:58.830 more powerful, maybe slightly more expensive 00:12:58.830 --> 00:13:02.909 model, like Claude 3 .7 Sonnet. Then... use that 00:13:02.909 --> 00:13:05.049 logging system we just talked about log which 00:13:05.049 --> 00:13:07.889 model handled which email and maybe even track 00:13:07.889 --> 00:13:11.250 the quality or if a draft needed editing over 00:13:11.250 --> 00:13:14.049 time you get real data on which model gives you 00:13:14.049 --> 00:13:16.029 the best bang for your buck for different types 00:13:16.029 --> 00:13:19.190 of emails whoa hang on if you combine that logging 00:13:19.190 --> 00:13:22.009 with the a b testing You could scale that across 00:13:22.009 --> 00:13:24.409 a whole company. Use the logs to scientifically 00:13:24.409 --> 00:13:27.409 figure out the absolute best, most cost -effective 00:13:27.409 --> 00:13:29.769 model for every single different email task. 00:13:30.049 --> 00:13:33.090 That's genuinely data -driven efficiency right 00:13:33.090 --> 00:13:36.210 there. Exactly. Optimizing cost and quality based 00:13:36.210 --> 00:13:38.389 on real results. How quickly does connecting 00:13:38.389 --> 00:13:41.210 a knowledge base with our ag actually change 00:13:41.210 --> 00:13:44.159 the quality? Is it noticeable right away? Immediately. 00:13:44.279 --> 00:13:46.960 The agent shifts from potentially guessing to 00:13:46.960 --> 00:13:49.039 actually looking up facts from your own internal 00:13:49.039 --> 00:13:51.600 docs. It's night and day for accuracy on specific 00:13:51.600 --> 00:13:54.139 topics. OK, but it's tech. Things will go wrong 00:13:54.139 --> 00:13:56.200 sometimes. Yeah. What are the usual problems 00:13:56.200 --> 00:13:58.259 people run into and how do they fix them quickly? 00:13:58.649 --> 00:14:00.730 Yeah, inevitably there's some turbulence. We 00:14:00.730 --> 00:14:03.710 usually see kind of a big four. First is misclassification 00:14:03.710 --> 00:14:06.230 emails going down the wrong path. The fix there 00:14:06.230 --> 00:14:08.509 is usually refining those category descriptions 00:14:08.509 --> 00:14:11.110 in your initial classifier. Add more specific 00:14:11.110 --> 00:14:13.490 keywords, maybe better examples. Okay, what else? 00:14:13.750 --> 00:14:17.210 Generic responses. If the AI's replies sound 00:14:17.210 --> 00:14:20.350 bland or unhelpful, that points straight back 00:14:20.350 --> 00:14:23.250 to the system prompt for that AI agent. You need 00:14:23.250 --> 00:14:25.830 to inject more detail, more personality, more 00:14:25.830 --> 00:14:29.470 specific company context. Make the prompt richer. 00:14:29.769 --> 00:14:32.230 We talked about drafts earlier. What if the drafts 00:14:32.230 --> 00:14:34.690 aren't linking properly to the original email 00:14:34.690 --> 00:14:38.230 thread? 99 % of the time, that's the threaded 00:14:38.230 --> 00:14:41.009 issue again. Go back to that create draft action 00:14:41.009 --> 00:14:43.620 in the Gmail node. Double, triple check that 00:14:43.620 --> 00:14:45.860 you've correctly mapped the threaded from the 00:14:45.860 --> 00:14:48.320 trigger into the draft ID field, not the message 00:14:48.320 --> 00:14:50.240 ID. That's almost always the culprit. Got it. 00:14:50.360 --> 00:14:53.879 Threaded. And the fourth common issue. Credentials 00:14:53.879 --> 00:14:56.360 errors or the workflow just stopping. This could 00:14:56.360 --> 00:14:58.580 be simple, like needing to reauthenticate your 00:14:58.580 --> 00:15:01.919 Google connection in N8NN. Or it could be you've 00:15:01.919 --> 00:15:04.340 run out of API credits with your AI provider, 00:15:04.539 --> 00:15:07.379 like OpenRouter or OpenAI. Always check your 00:15:07.379 --> 00:15:09.340 billing dashboard if things just stop working. 00:15:09.799 --> 00:15:12.360 So if a user's workflow just dies unexpectedly, 00:15:12.879 --> 00:15:14.679 what's the very first thing they should check 00:15:14.679 --> 00:15:17.419 before diving into complex debugging? Honestly, 00:15:17.700 --> 00:15:20.559 confirm that active toggle switch is still green. 00:15:20.740 --> 00:15:23.139 It sounds silly, but sometimes things can get 00:15:23.139 --> 00:15:26.059 accidentally deactivated. It's the simplest possible 00:15:26.059 --> 00:15:28.980 fix for what looks like a dead workflow. This 00:15:28.980 --> 00:15:33.340 all sounds incredibly powerful. A 2047 AI assistant 00:15:33.340 --> 00:15:37.509 managing email must cost a fortune, right? Let's 00:15:37.509 --> 00:15:39.809 talk ROI. What's the actual cost here? You'd 00:15:39.809 --> 00:15:41.509 be surprised. It's actually very accessible. 00:15:41.730 --> 00:15:43.830 Let's break it down. You've got your base platform 00:15:43.830 --> 00:15:46.529 cost, so the NAN cloud plan you'd likely need 00:15:46.529 --> 00:15:49.110 is around, say, $25 a month. That's fixed. Okay. 00:15:49.210 --> 00:15:52.759 Then you have the variable cost. the AI API calls. 00:15:53.019 --> 00:15:55.899 This is pay as you go. But the key is, remember 00:15:55.899 --> 00:15:58.460 how we use cheaper, faster models like GPT 4 00:15:58.460 --> 00:16:00.840 .1 mini for that initial high volume classification 00:16:00.840 --> 00:16:03.500 step? Right. Most emails just get sorted. Exactly. 00:16:03.779 --> 00:16:06.440 That keeps the cost way down. So even if you're 00:16:06.440 --> 00:16:08.419 processing, let's say, 100 emails every single 00:16:08.419 --> 00:16:10.440 day. Yeah. Give me a real world number for that 00:16:10.440 --> 00:16:12.960 kind of volume. With this kind of optimized setup, 00:16:13.080 --> 00:16:15.500 you're probably looking at only about... 50 cents 00:16:15.500 --> 00:16:18.039 to maybe a dollar per day in total API fees. 00:16:18.279 --> 00:16:21.000 Seriously, that low? Yeah. So add that up for 00:16:21.000 --> 00:16:24.299 a month, maybe $15 to $30 max for the AI usage. 00:16:24.539 --> 00:16:27.340 Your total all -in cost per month for this whole 00:16:27.340 --> 00:16:30.100 system, realistically, somewhere around $40 to 00:16:30.100 --> 00:16:33.759 $55. Okay, hold on. $55 a month sounds amazing, 00:16:33.820 --> 00:16:36.200 almost too good to be true, but that's just the 00:16:36.200 --> 00:16:38.820 software and API cost. What about the human time? 00:16:39.259 --> 00:16:42.759 Isn't there a hidden cost in setting this up, 00:16:42.799 --> 00:16:45.289 tweaking the prompts? fixing things when they 00:16:45.289 --> 00:16:47.549 break that's a really fair point yes there's 00:16:47.549 --> 00:16:49.049 an initial time investment you got to build it 00:16:49.049 --> 00:16:51.049 right maybe budget i don't know three to five 00:16:51.049 --> 00:16:52.990 hours for the initial setup and the first week 00:16:52.990 --> 00:16:55.330 of watching it tuning it okay but once it's stable 00:16:55.330 --> 00:16:57.389 it really does mostly run itself maybe occasional 00:16:57.389 --> 00:17:00.169 tweaks so you weigh that initial time cost and 00:17:00.169 --> 00:17:02.450 the 55 a month against the time it saves you 00:17:02.450 --> 00:17:05.490 if this thing genuinely handles tasks that used 00:17:05.490 --> 00:17:09.250 to take you five maybe 10, even more hours every 00:17:09.250 --> 00:17:12.089 single week. Yeah. That $55 looks pretty insignificant 00:17:12.089 --> 00:17:14.910 pretty quickly. The ROI is usually exceptional, 00:17:15.150 --> 00:17:17.170 often within the first month or two. And that 00:17:17.170 --> 00:17:19.450 minimal running cost, is that mostly because 00:17:19.450 --> 00:17:22.549 we're being smart about using cheaper AI for 00:17:22.549 --> 00:17:25.150 the bulk classification work? Yes, absolutely. 00:17:25.609 --> 00:17:28.349 Optimizing model choice by task is the key. It 00:17:28.349 --> 00:17:31.569 keeps those API costs extremely low while still 00:17:31.569 --> 00:17:33.809 using powerful models where you really need quality. 00:17:34.150 --> 00:17:37.319 Wow. OK, so we've really seen this technical 00:17:37.319 --> 00:17:39.559 guy transform, haven't we? It's gone from just 00:17:39.559 --> 00:17:42.740 sorting emails into a genuinely operational system, 00:17:42.859 --> 00:17:46.160 one that sends replies, preps drafts for you, 00:17:46.279 --> 00:17:48.619 alerts the right people and just cleans up the 00:17:48.619 --> 00:17:50.279 clutter. Yeah, it's about building structure. 00:17:50.420 --> 00:17:53.140 This AI assistant, when set up right, it delivers 00:17:53.140 --> 00:17:56.059 organization, sure, but also like genuine peace 00:17:56.059 --> 00:17:58.420 of mind. And yeah, critically, it really does 00:17:58.420 --> 00:18:00.500 give you back those five or 10 hours every week, 00:18:00.519 --> 00:18:03.220 the gap between email chaos and having this like. 00:18:03.549 --> 00:18:06.009 personal 247 chief of staff. It's just a bit 00:18:06.009 --> 00:18:08.009 of setup time and that's surprisingly small monthly 00:18:08.009 --> 00:18:10.230 cost. So for everyone listening, maybe take a 00:18:10.230 --> 00:18:12.029 minute after this to think about your own email 00:18:12.029 --> 00:18:14.309 habits. Where are those repetitive tasks? Where 00:18:14.309 --> 00:18:15.869 does all the time go? Start thinking about which 00:18:15.869 --> 00:18:18.309 action branches you'd build first. And that brings 00:18:18.309 --> 00:18:20.789 us to our final thought. A bit provocative, maybe. 00:18:21.109 --> 00:18:23.670 If you actually manage to delegate 10 hours of 00:18:23.670 --> 00:18:26.109 email work every single week using a system like 00:18:26.109 --> 00:18:29.029 this, what's the one big, important, high -value 00:18:29.029 --> 00:18:31.470 project you'd finally have the bandwidth, the 00:18:31.470 --> 00:18:33.849 focus, the time to really dedicate yourself to? 00:18:33.950 --> 00:18:36.750 What becomes possible then? At UTR, we'll music.