← Live demos Free live demo
AI builds an invoice and receipt extractor One short prompt to ChatGPT Codex, and it builds a Make scenario that reads receipts and invoices with OCR, checks them with a Make AI agent and stores the details. Then we add WhatsApp so staff can send a photo of a slip.
19 min ChatGPT Codex Make MCP Make AI agent OCR WhatsApp
Your browser does not support video playback. Back 10 sCaptions on 0.75× 1× 1.25× 1.5×
In this demo 00:00 What we’re building01:33 The prompt and a blank scenario03:40 While Codex builds: meet Maia05:59 Which model to use, webhooks vs schedules08:04 What it built: OCR and an AI agent09:45 Testing it11:00 It skipped a step: adding WhatsApp12:30 The AI does the hard mapping14:09 Knowledge bases and the glass box16:50 The WhatsApp scenario, and why it matters
Keyboard shortcuts When the video or its buttons have focus:
Space or K play or pauseJ back 10 secondsL forward 10 secondsC captions on or offTranscript 00:00 Let's build an invoice and receipt 00:02 extraction automation using ChatGPT, and 00:05 make automations. 00:06 so if you've got staff and you've got this 00:08 administrative burden of constantly 00:10 needing to manually insert your invoices 00:13 or receipt, receipts or slips, this 00:15 automation allows your staff to, use OCR, 00:18 and an AI to upload that into your records 00:21 so that you can use that for accounting or 00:22 what-whatana-analysis ever you want to use 00:25 it for. 00:25 we're gonna use ChatGPT Codex, and we're 00:27 gonna use the Make MCP, Model Context 00:29 Protocol. 00:30 So we're gonna start from scratch, and 00:32 we're gonna actually have ChatGPT build 00:34 this out for us entirely. 00:36 So follow along. 00:37 We're gonna move into ChatGPT first. 00:39 we're gonna quickly brief it with a small 00:42 prompt, and then we're gonna see what it 00:44 builds out for us in, in the Make 00:46 scenario, and then we can test it with 00:47 some generic information. 00:49 so we're gonna do a deep dive. 00:50 All these skills are available on the 00:52 course, on the Cogniz Build course, to 00:54 allow you to connect ChatGPT, to 00:56 understand Make and what you can do in 00:58 Make, and how you can build automations 01:00 that you can actually sell to other 01:01 companies. 01:02 So a company that's got a lot of slips, 01:05 petrol slips, manual slips, so, someone 01:07 that's in, in, the service industry, a 01:09 plumber, electrician, handyman, that buys 01:12 at different places, screws or, or nuts 01:14 here and pipes there and petrol slips. 01:17 This is ideal because you could extend 01:19 this. 01:19 We're not gonna do it in this ana this 01:20 video, but you can extend this so that you 01:22 can actually send from your WhatsApp or 01:24 from your staff's WhatsApp directly to the 01:27 system a photo of the receipt, the 01:29 invoice, the slip, or whatever you need to 01:31 do. 01:31 so let's get it started. 01:33 In ChatGPT, we're going to do a quick 01:35 prompt. 01:35 Now, you can copy some of the prompts, and 01:37 there's a lot of prompts you can download, 01:39 but we're gonna do a quick little brief. 01:41 you can use the voice mode, but I'm just 01:43 gonna type quickly here for reference. 01:45 so let us let's use our Make skills 01:50 and the Make MCP to build out an 01:53 automation that receives via a webhook for 01:57 now, and this can be WhatsApp, this can be 01:59 email, different means in which you can 02:01 receive information in your on your Make 02:03 automation. 02:04 receive via webhook, Receipt uploads or 02:09 invoices. 02:10 And we want to use the Make AI 02:14 agent, call it new. 02:16 it's a new one, but it's the only one that 02:17 they currently have. 02:18 And, AI tools OCR to extract the 02:23 content and add it to a Google 02:27 sheet and/or database. 02:30 please build out the entire scenario using 02:34 this blank scenario. 02:36 All right, so I've got a blank scenario 02:37 set up, so that we can actually quickly 02:40 copy, copy that. 02:42 then we can start from there so that we 02:45 can reference. 02:45 But you don't have to start from a blank 02:46 scenario. 02:47 You can actually just ask ChatGPT to 02:48 connect to your Make, create a new 02:50 scenario, select your team. 02:51 It'll ask you, prompt you a couple of 02:53 questions. 02:53 So you just need to make sure. 02:55 I'm gonna copy this URL. 02:56 It's a brand-new, you can see, blank, 02:58 absolute blank, new scene, new scenario in 03:01 one of my Make teams. 03:03 And then we're going to paste that into 03:06 ChatGPT so that we can use it to start 03:08 there. 03:09 Now, it might ask a couple of questions, 03:11 and ideally Then you can, you can always 03:14 ask ChatGPT to ask you questions before it 03:17 does anything. 03:18 All right? 03:18 So if you want to, have it clarify any 03:21 specifics of what you wanna do. 03:22 Now you can see I've got full permission 03:24 on mine. 03:24 My ChatGPT Codex is, set up to work, 03:27 exceptionally well with me. 03:28 It knows what it's allowed to do and not 03:30 allowed to do, so I don't want it to ask 03:32 me too many questions all the time. 03:33 My work environment and my workspace on 03:35 this device is set up, securely so that I 03:38 can execute, what I when I want to 03:39 execute. 03:40 So we're gonna let it, let it, let it go, 03:42 and then we'll just need to refresh the 03:44 Make scenario, once ChatGPT, started, 03:47 including the modules. 03:49 so it's not, it's not live. 03:50 You can't see when, when it's busy. 03:52 You can't see it populate the, the live, 03:54 the live, modules when you, when you're 03:56 busy with it. 03:57 Just quickly switch back to the Make 03:59 scenario. 04:00 I can actually show you if we do, I'm not 04:03 gonna save this, but here you can go and 04:05 obviously select the webhook, which it can 04:08 start off from, or it can start off via, 04:10 an email that comes in. 04:12 There's different You can do an HT-HTTP 04:14 call. 04:14 you can have your WhatsApp in there as 04:16 well. 04:16 So here you can, you can view a lot of 04:18 the, the options that Make provides in 04:20 terms of what you can do with their AI 04:23 agents. 04:23 we're gonna refresh this and see if Codex 04:26 has started anything here. 04:28 And you can wait for its response, but 04:30 it'll build out because it's connected to 04:31 my Make skills. 04:32 let's quickly see. 04:34 Okay, so not yet. 04:35 So that's gonna start off the scenario. 04:38 So if we head back to ChatGPT, you can see 04:40 that it's using my Make skills, which is 04:42 installed, and it's going through via the 04:45 MCP. 04:45 so it's already confirmed the scenario is 04:47 blank, and we're gonna use the Make AI 04:49 provad provider connection. 04:51 Now, you can chat with, with ChatGPT while 04:53 it's busy as well, 'cause what we're gonna 04:55 do is... 04:56 Yes, we're using the Make AI module and 05:00 Make AI provider and the required OCR 05:04 extraction tools, for that agent. 05:07 Okay, so it's running now. 05:08 We'll just wait, wait a second, and then 05:11 I'll cut you, and we can, we can refresh 05:13 and see, see how it's building. 05:15 Now, when you're in Make, Make's got a, a 05:17 native AI agent in the scenario that you 05:20 can also use. 05:20 This is Maia. 05:21 Maia is still in, in beta, but, it's 05:23 available on all paid platforms. 05:25 Now, this consumes credits in your Make 05:26 account. 05:27 Right? 05:27 So you can, you can actually build, live 05:30 within Maia as well, so that you don't 05:32 have to use ChatGPT or Claude. 05:34 and then you can actually visually see it 05:36 examining modules, adding modules, and 05:39 populating those modules. 05:40 it'll ask you questions, clarifying 05:42 questions. 05:43 It can run the scenarios to test them as 05:45 well. 05:45 So it's quite handy, but it does consume 05:47 credits within your Make account. 05:49 So if you've already got a paid ChatGPT or 05:51 Claude, then, you can use, those with the 05:54 Model Context Protocol, the MCP, to 05:56 connect to, to your Make scenarios. 05:59 Let's refresh quickly and see if, ChatGPT 06:01 has started anything yet, or is it still, 06:03 still thinking. 06:04 Okay, so, so nothing there. 06:06 Let's quickly head back to ChatGPT to see 06:08 where we are. 06:08 okay, so it's actually running a couple of 06:10 things now. 06:11 I'm using, I'm using, ChatGPT Sol. 06:13 it's on extra high. 06:14 I mean, you can, you can switch to a 06:16 different model. 06:17 Obviously, Astra, is too intensive for 06:19 this, for this role. 06:20 So you always start off with something 06:22 like Sol or Tera, to build out the base of 06:24 the scenario. 06:25 When you encounter any issues that might 06:27 be complex, depending on the scenario that 06:29 you build, then yes, you can switch to 06:30 Astra, but the but note that it obviously 06:32 takes more, from your, your, your usage, 06:35 allowance on your account. 06:36 Okay, so it's actually editing the 06:38 scenario as we speak. 06:39 So we can, expand that, and then we can 06:41 see what it's busy doing. 06:42 let's quickly refresh again to see if 06:45 we've got anything happening in Make. 06:47 but it would it will be, it will be 06:48 populating, the scenario quite soon. 06:51 Okay, so it's already started with the 06:52 webhook. 06:53 So we'll refresh that shortly. 06:54 It might prompt us to save, to save the 06:57 scenario or discard changes, while it's 06:59 busy. 06:59 We'll, we'll just refresh and discard 07:01 changes while ChatGPT is busy building 07:03 this. 07:03 Now, the webhook is, very handy, I think 07:06 for, for a lot of people that are still 07:08 beginners, and you want to connect an 07:09 ex-external platform or external service 07:12 or external data source to a Make 07:15 scenario, then you can start with a 07:17 webhook which can, which can run 07:18 instantly. 07:19 So it's currently set to instantly, which 07:21 means as soon as data is sent to the 07:24 webhook, it triggers the automation. 07:26 But you can also schedule automations, 07:28 right, within Make. 07:29 So you can set a schedule every ten 07:32 minutes, every five minutes, every two 07:34 days, or every ten days, run this 07:36 automation for you. 07:37 And that depends on, on what you want to 07:39 achieve. 07:39 If it's research, or if it's an evaluation 07:42 of, some of your da-data sets, you can 07:44 have scheduled tasks. 07:46 You can do, for argument's sake, cold 07:48 outreaches or newsletters or blog posts or 07:51 social media posts with a schedule. 07:53 So where you can go and say nine AM every, 07:56 every morning, link up your LinkedIn 07:57 account or your Facebook, business page, 07:59 and then it can do an automation and, and 08:01 actually post content there for you. 08:04 I'm gonna quickly refresh and see if we've 08:05 got anything new. 08:06 We're gonna reload it. 08:07 We don't wanna save anything currently 08:09 while ChatGPT is busy populating this for 08:12 us. 08:12 and there we go. 08:13 Already, a lot of modules added. 08:15 So it's receiving We can see here it's 08:17 receiving the receipt, it's receipt or 08:19 invoice, via the webhook, and then it's 08:21 validating the upload. 08:22 if it's, if it's unreadable, it'll reject 08:24 it. 08:25 So it'll send information back to the 08:27 platform that sent it and say, "Listen, 08:29 you need to re-upload this or get, get a 08:32 better version." And then we've got an AI 08:34 agent component here or tool, which is an 08:37 OCR extraction, module, where it takes a 08:40 file that you upload, and then you can map 08:42 the data within the file page ranges, any 08:45 limits that you want to apply. 08:46 it'll extract the information, then it'll 08:48 combine different pages, so it'll iterate 08:50 through them and aggregate the extracted 08:53 content, and then it'll send it to a Make 08:55 AI agent. 08:56 Now, this is a very, very powerful tool. 08:58 you can add additional tools to any Make 09:02 AI agent. 09:02 So you could, for argument's sake, do web 09:04 search. 09:05 You can add another agent if you want 09:07 specific roles. 09:08 you can add a knowledge base about your 09:10 company, information that it should know. 09:12 You can add separate Make scenarios to an 09:15 AI agent for it to orchestrate that, to 09:18 manage that, and to even run that. 09:20 and then you can also include a lot of 09:22 your standard Make, Make modules and 09:24 connections. 09:25 You can add your Gmail, Airtable, 09:27 Telegram, Facebook, you name it. 09:29 There's over three thousand different apps 09:31 that you can connect. 09:32 A lot of them, obviously, you need to have 09:33 an account with that service provider, 09:35 Zoho, Xero, whatever you wanna do. 09:38 and you can add those tools to a s- Make 09:40 AI agent to execute, or understand, 09:42 whatever you need from, from those 09:44 scenarios. 09:45 So we'll quickly refresh this, and we're 09:47 gonna head back to, to ChatGPT to see 09:49 what, if there's any progress. 09:51 Okay, so it's already added additional 09:53 tools here, you can see. 09:54 It added, two tools to the AI agent. 09:57 It's parsing, and extraction, 09:59 ex-extracting the, the extracted content 10:02 and then, parsing the JSON. 10:04 It's adding it to a data store, and then 10:05 it's sending information back to, to the 10:08 platform that sent the information in the 10:09 first place. 10:10 So let's quickly head back to ChatGPT and 10:12 see where we are with this. 10:13 Okay. 10:14 so it's running through the flow, it's 10:16 saved and supply scenario. 10:17 it'll test the multi-part upload and check 10:19 the stored record. 10:20 So it's gonna use generic information. 10:22 and then you can, you can feed it. 10:24 now you can expand on this chat with 10:25 ChatGPT. 10:26 this is all, again, part of the courses 10:28 that we're offering on automation, on 10:30 website development, platform development, 10:32 understanding AI, AI and automations as an 10:36 absolute beginner on build.cogniz.online 10:39 You can also sign up for a free Make 10:40 account. 10:41 You get a thousand credits per month. 10:42 And then you can use your ChatGPT or your 10:44 Codex to build an automation that you can 10:46 actually sell to customers or use 10:48 internally to optimize, your operating 10:51 procedures within your business, and get 10:53 rid, get rid of, of administrative burdens 10:55 that take up a lot of time, but not then 10:57 that doesn't necessarily hold a lot of 10:59 value. 11:00 Okay, let's see where we are. 11:02 so Make has gone or ChatGPT's gone and 11:06 tested, with some generic data, our 11:08 workflow. 11:09 I see it didn't include the, Google Sheets 11:12 like I initially asked. 11:13 instead, it used a webhook to return the 11:17 information. 11:17 So that's obviously that what you can, 11:19 what you can update. 11:20 Let's quickly refresh. 11:21 let's refresh the Make scenario so we can 11:24 see what, what we got here. 11:26 See if it maybe included the Excel sheet 11:28 after it did the initial testing. 11:30 Or maybe it included a separate scenario. 11:32 No, so only still one scenario, and we've 11:35 got a data store, but we do not have a, a 11:37 Google sheet. 11:38 So you can now go and prompt ChatGPT to 11:40 correct this. 11:41 but yeah, in, in normal platforms, if you 11:42 don't wanna use a Google Sheet, and you've 11:44 got... 11:44 you're advising a company that wants a 11:46 system, a platform where they can manage 11:48 this, then you can return via webhook to 11:50 your platform that you host for the 11:52 company or the service that you provide to 11:54 them, or to any other connections. 11:56 You can also use HTTP, and then you can 11:58 now extend this with a WhatsApp line. 12:01 So, let's do a part two while we're at it. 12:04 let's ask ChatGPT to modify this so that 12:06 we can also accept, messages from WhatsApp 12:09 and run it through the same, OCR 12:11 extraction and AI agent, process. 12:13 So we let's quickly head over back to 12:15 ChatGPT. 12:16 so, oh, they'd asked, can we connect 12:18 Google Sheets to have this? 12:19 No, we're gonna skip that for now. 12:20 let's add a WhatsApp intake so staff can 12:24 send documents and photos via WhatsApp to, 12:28 to the same flow. 12:30 Now, this will take a while as well, so 12:31 it's going to actively modify that 12:33 scenario, or it might elect to build a 12:36 separate scenario that feeds into that 12:38 same webhook. 12:38 now you can be much more explicit in terms 12:41 of how you want this orchestrated, how you 12:44 want this delivered, and obviously how you 12:45 deal with errors as well. 12:47 so this is a simple, a simple scenario. 12:49 You need to brief, ChatGPT with exactly 12:52 the flow that you envision. 12:53 Now, you don't have to understand every 12:55 module in Make and how to even set up 12:57 those modules. 12:58 and I'll actually show you quickly if we, 13:00 if we had to Make, where there can be 13:03 complexities. 13:03 so in, in setting up the, the data store, 13:07 or in parsing the JSON. 13:09 Okay, so this is now a simple one, 13:11 sim-simple version. 13:12 It's all, run by this AI agent itself. 13:14 but it ca-can become quite, daunting for 13:17 beginners and anyone that doesn't really 13:19 code, any vibe coder, to, to populate some 13:22 of these modules, especially when you do 13:23 da-data validation, and when you need to 13:25 do mapping of certain variables and 13:28 strings and Booleans into all of the 13:31 different modules for it to work. 13:32 But ChatGPT Codex and Claude actually 13:34 understands that quite well. 13:36 Let's see. 13:36 We've got a message from ChatGPT. 13:38 the existing Cogniz WhatsApp number also 13:40 handles customer inquiries, so that's, 13:41 that's the business number that we've got 13:42 on there. 13:43 let's, let's just do a mock-up. 13:44 just do a mock for now. 13:47 this is for a video demonstration. 13:50 Demonstration. 13:51 so just add the module in how it would 13:54 look. 13:54 All right, so now it's gonna probably add 13:56 that module for us just to show how that 13:58 works, or then do that, different 14:00 scenario. 14:01 Quickly refresh our Make. 14:02 And then I think what we'll do is we'll, 14:04 we'll give it a, a, a minute or two to 14:06 populate that additional module in there. 14:09 Now, obviously, you can go and add custom 14:10 modules anytime you want, So you can add 14:12 more tools. 14:13 just remember, if you modify this while 14:15 ChatGPT is modifying it, the last one to 14:18 save will have the version, and the other 14:20 one won't. 14:21 So, yeah, just if, if ChatGPT or, or 14:23 Claude is working on it, just let them do 14:25 their job first. 14:26 and then you can go, and you can add a 14:28 knowledge base if you want to. 14:30 So let's say, you're building an 14:31 automation where, the information needs to 14:34 come from your company sources. 14:36 That might be your brand assets. 14:37 that might be your, procedures that you 14:40 follow when you engage customers, the 14:42 narratives that you use. 14:43 it might be specific, product information 14:46 sheets. 14:46 you can add these to this AI agent, so 14:49 it's got the knowledge base. 14:50 So you can upload files, PDFs, doc, text, 14:53 and JSON files, and then it'll actually 14:56 have more context on, on what it should 14:59 consider. 15:00 So it's not necessarily just what it, what 15:02 it conjures up. 15:03 you can also, like I mentioned earlier, 15:05 you can add web search. 15:06 So if you wanted to go and search a 15:07 specific website, whether you're building 15:09 a scraper or competitive, analysis agent 15:12 or a market research agent, you can have 15:14 it go and search the web and add your 15:17 knowledge base if you, if you want to do 15:19 if, if you want to provide more context to 15:21 your AI agent. 15:22 Let's quickly refresh. 15:23 And while we wait for that, now imagine 15:25 the possibilities, of what you can 15:27 actually do because ChatGPT and Claude can 15:31 connect to one of these scenarios. 15:33 So you can actually extend your ChatGPT or 15:36 your Claude with a Make automation, to 15:39 connect to tools that might not be 15:40 available in ChatGPT or Claude code. 15:42 you can have Claude Code and ChatGPT 15:45 communicate with each other through a Make 15:47 scenario if you want to. 15:48 You can add a Make AI agent to serve 15:50 spec-specific roles or webhooks, or data 15:53 stores. 15:54 So you can be very creative because what 15:56 Make offers you is a visual canvas where 15:58 you can see what the actions are, what the 16:01 processes are, and how they flow, so that 16:04 you can modify a step in between this 16:06 massive orchestration. 16:08 So Make calls it the glass box of AI, and 16:11 that's typically where a lot of people 16:12 fail and, and, and struggle. 16:14 You don't know you don't always know what 16:15 ChatGPT is doing. 16:17 You don't know what Claude's doing. 16:18 it's just running this, and it's bashing 16:20 this, and it's grepping this. 16:21 It's saving that, and then it's running 16:23 into PowerShell. 16:24 but you never really know where what is 16:26 and, and how to control that. 16:27 But with a Make scenario, you can mix 16:30 deterministic flows with agentic flows. 16:33 You can control the portions and the, the 16:35 elements that are critical and high risk, 16:38 and then you can integrate AI automation 16:41 and, and AI, into that workflows to 16:42 provide context, to provide actions, and 16:45 eventually an agentic orchestration that 16:47 can serve a, a specific purpose. 16:50 let's quickly head over to ChatGPT. 16:51 It asked us a question, I think. 16:53 No, it didn't. 16:53 built a separate WhatsApp, intake video. 16:56 Oh, it actually built us a, a video demo. 16:58 Right. 16:58 So that's why you need to be, very 17:00 specific. 17:01 So let's have it open up. 17:03 Okay, so it wants me to sign in. 17:04 I'm not gonna sign in there. 17:05 So we're just gonna copy this link, and 17:07 then we're gonna head over to, to Make, 17:09 and then we'll paste it there and view it 17:12 there. 17:13 Okay, let's quickly re-reload. 17:14 So this is the one that it's built, and it 17:16 might have added a sub-scenario to the AI 17:19 agent. 17:19 and this is the separate scenario that I 17:21 built. 17:21 Okay, so it built a duplicate, so it 17:23 didn't modify our, our initial, initial 17:26 one that we did from the webhook. 17:28 It actually built a separate one. 17:29 So this one, instant. 17:31 When a WhatsApp arrives to your connected 17:33 number, you'll see it's got a tab because 17:34 it's not, connected. 17:36 it'll then read the incoming message. 17:37 And I mean, you can do voice notes, you 17:39 can do documents, you can do text 17:41 messages, you can do images, you can do 17:42 video if you want to. 17:44 You can connect this to ElevenLabs and 17:46 transcribe the audio and the video, and 17:48 then feed that back in here as well. 17:49 download a photo document, and then it 17:52 runs through that similar process, 17:53 extracting the information, understanding 17:55 the information, and then executing on 17:57 your request to extract the information on 18:02 the receipt, slip, invoice, whatever 18:04 document that might be, adding it to a 18:06 database, and then sending it back to you, 18:08 or sending it to your platform or your 18:10 service provider. 18:11 So, very, very handy. 18:13 That is a business opportunity in itself. 18:15 you can go to your local plumber, 18:17 electrician, manufacturer, mechanic, 18:20 supply company, anyone that deals with a 18:22 lot of invoices coming in and out. 18:23 you can go and build them a WhatsApp 18:25 automation or a platform where they can 18:27 feed their documents, have it extracted, 18:30 collated, orchestrated, unified, and then 18:33 present it to them so that they can easily 18:35 see what's going on in their business. 18:36 And this is a simple entry-level, 18:38 demonstration of how you can use ChatGPT 18:41 natively, conversationally to build you an 18:44 automation that actually does real work. Want to build this yourself? The Cogniz course teaches each piece step by step: connecting ChatGPT or Claude to Make, reading a scenario, and checking what the AI built before you rely on it.
Try a free step → See membership