From Demo to Deployment: Where Most Companies Get Stuck

**[00:00:00]** Hello everyone, my name is Armando Perez-Carreno, and welcome to the Web Talk Show. Today we are live and we're talking about why buying AI tools will not fix a broken workflow. Buying an AI tool is the easy part. Changing the workflow is a hard part. That's why so many AI projects look so incredible and demos then quietly disappear inside the company. Demo works because the demo is clean. Your business is not clean and I don't mean it as an insult. I'm being serious here. Real businesses are full of exceptions, hand-offs, missing information, old habits, weird edge cases, and people who know things because they have been around long enough to remember the last time everything broke. A demo typically does not have that. A demo has perfect sample data, a demo has clear input, a demo has polished output, a demo has a person moving through the exact path that the product team wants you to see. But then the tool enters the real business. That's a big difference. Demos seem amazing and I'm going to get into that a little bit more, but we really have to dive into the reality of things and why some of these projects end up failing when a demo looked so beautiful. The reason I'm doing this particular video is yesterday I was browsing around and I saw an ad, well, one of those paid collaborations. It was a it was a post, but it was paid by the company for this creator to publish or show. And so, it looks very interesting. It was it was for a one of those AI avatar generation tools, like HeyGen. HeyGen is really good, by the way. But it was another one and I I found it interesting. I saw it was really nice and I said, "Wow, let me go to their website." And I went to the website and it looked amazing. This was like one top-tier website. I'll tell you that. It was so clean, so polished. It made me feel like, "What? This is going to has to be an outstanding tool. I **[00:02:00]** might even consider it for a project or something." Then I go through the comments and I start seeing that the results weren't as advertised. So many people went through that same process of seeing, "Oh, it looks amazing, etc." and then they sign up and they pay for it and they start using it and it's like it matches nothing to what they're promising. And I've seen that happen in multiple different types of scenarios. So, again, comes to the demo part where you see something in a demo, it looks amazing, but once you start putting it into practice, that's where it was really interesting. So, going back to what we were discussing, when the tool enters a real business, that's where things get tricky. And that's why if, let's say you're a provider, then try to do a demo with actual real data so the customer actually sees what what it's going to do with real data. And if you're a company looking for a provider, then maybe consider asking the provider, "Hey, can I give you some data? Can we do some tests just to see if it's going to generate the type of outputs that that we're looking at, that you're promising?" Just to get a little better understanding, a piece of mind. Because nowadays, I can create a demo in 60 seconds. So, I could be talking right now and somebody can put the message in the chat and I could just create a demo and then talk about it and show it and it would look amazing. But it's not a real thing necessarily. So, while it's amazing that we can do that, because that way I can showcase an example that then we could then implement, that implementation takes a little bit longer cuz you have to think of all the edge cases, we have to think of all the edge cases. So, in reality, a request comes in through an email, maybe, in a business, right? Instead of the form. So, we have a form and we think everyone's going to fill out the form, but that's not true. Maybe they just send an email or a fax, right? They drop a letter. The customer gives half the information, they might not give all the information. The CRM might have an old record or multiple records for the same person. The project manager has newer context than what we have here because they have a separate message thread. Maybe they called on the phone. The **[00:04:00]** spreadsheet maybe has a column that no one trusts. It's there, you know which one, right? It's like it's there but everyone's like, "I don't know about that one." The person who knows the exception is on vacation. And then the AI tool is sitting there very confidently waiting for a clean process that does not exist. This is why buying AI tools will not fix a broken workflow. AI can make workflow faster. AI can make a clear process easier to run. AI can help people draft, summarize, classify, search, route, and decide with better context. But, if the workflow itself is unclear, AI will often just move the confusion around. Sometimes, it moves the confusion even faster, which is not quite the innovation we were hoping for. Gartner predicts that at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025. In this case, we're now in 2026 and we have seen that. You can look at a lot of these reports and it's happening. It's not that AI is bad. It is not that the AI won't work. It's that, like I said in my previous video, if you watch that and any of the our channels, you might be trying to fit AI into something that is currently broken. And when we go with clients, we always say, "Let's first figure out what a human would do in this scenario. Make sure we have a clear workflow and then you can put AI into it." And I'll talk a little bit more about that as well. Now, the reasons include unclear business value, poor data quality, inadequate risk controls, and of course rising costs. In plain English, a lot of companies are building AI projects before the work is ready for AI. That's the heart of the problem. A proof of concept can show that AI can do something, like I mentioned earlier, and it's very easy to do. But, it does not prove that your company is ready to use it. Those are two entirely different things. The model **[00:06:00]** might be capable, and they're all very capable now, but the workflow might not be. The tool might be impressive, but the data might be very messy. The output might be good in isolation, and the team might still avoid it because it doesn't fit to how the work actually happens. Now, take a break there because this is what we've seen throughout the past 10-15 years with software in general, enterprise software mostly. When there's this top-down approach of "Hey, we bought this ERP system, and now everyone's going to use it." But, a ERP system then tries to fit everyone into this box that might not apply to that particular industry. Some industries are very standard, others are severely different in many many ways. And what happens is some people at the top use it, a lot of people in the middle might not use it, nobody at the bottom uses it because they're out in the field and it doesn't match it what their workflow really does. And so, what ends up happening there is they'll do it on paper first, and then at the end of the day, maybe they'll fill it out. But then, by that time, there could be a lot of mistakes. There might be inconsistencies. There might be things where they wrote the wrong thing, where they understood the wrong thing, they might forget about it, the paper might get lost, it might get filled in the next day. There's so many things that happen out in the field. So, this has been happening. And it's not an AI thing, it's just a software thing, it's just a people thing. So, before a company asks "Which AI tool we should buy?", I think a better question should be "What workflow are we trying to improve?" That one saves a lot of trouble because once you name the workflow, the conversation changes. You stop talking about AI in general and you start talking about work, sales follow-up, client onboarding, customer support triage, proposal creation, invoice review, weekly reporting, internal knowledge search, which is huge, scheduling, compliance checks, **[00:08:00]** document review. Each one has a different shape. Each one has different risks. Each one needs a different kind of solution. And some of them should not start with AI at all. Some should start with just a better form or a checklist or a cleaner database. And yes, you can have messy data and then you say AI to sort of help you clean some of that data for your cleaner database. That's another use case, right? I need to clean my data. And so maybe you use it for that. But you first have to clean your data, adjust your workflows before you think of an end-to-end workflow that moves data from one point to another, etc. Some should start with training. Some should start with a simple automation that does not lead a large language model anywhere near it. Like there's a lot of automations. I spoke about this at the National Lease event a few days ago. Sometimes you just need literally a wrench, right? Or a hammer. Sometimes you don't need AI. You just need a parser, a tool that just parses a document or a PDF or a CSV file and extracts fields and places it somewhere else. And there's no AI involved there. It's It's just a discrete set of steps that just work. So you don't always need AI. That's something to consider as well. And this is why workflow-first thinking matters. Tool-first thinking sounds like this. We bought this AI platform. What should we use it for? Workflow-first thinking sounds like this. This process wastes 6 hours a week. What is the simplest safe way to improve it? And that second question is much better. It is less shiny. It is also more likely to save money. So, I'll give you a few examples. Imagine a company wants to use AI for customer support. The tool demo shows an AI assistant answering customer questions. It looks great. It responds quickly. It uses a friendly tone. It knows the sample knowledge base. Everyone in the meeting is like, yes, this is what we need. But then the company tries to use it. And now the hard questions show up. **[00:10:00]** Where is the current support knowledge base? Is it up to date? Who are the answers? Which questions can AI answer by itself? Which one should be escalated? What happens when a customer is angry? What happens when a refund is involved? What happens when written policy and the real company have it are different? What happens when AI gives an answer that is technically correct but practically foolish? That last category by the way is where machines show real artistic range. The tool did not create those problems. The tool revealed them. The The broken part was there. A company had support knowledge scattered across documents like Slack, messages, old tickets, and people's memories. And AI did not fix that automatically. To make the support AI useful, the company first needs to turn support knowledge into something the system can trust. We talked about that in this previous video, legibility. We need the content, the knowledge, everything that's going on in the business to be legible to the AI first so that it can be used. Policies, standard operating procedures, approved answers, escalation rules, examples, tone guidance. That's really important as well cuz you probably had this experience before where you have you ask ChatGPT or Claude to write something that just sounds like everything else on the internet because they the model just behaves in a similar way. But you can give it voice and tone guidance so that it sounds much different. What are sensitive topics? Human review points. That is workflow work. I was doing just yesterday something about this where we were building a voice AI. And you might be 90% 95% of the way there and you think this is excellent. We've got everything within the knowledge of it. We've got the prompts. We've got the right structure. We've got the tools. We have We're talking to the schedule. We're talking to the other system. Everything looks fine and then the But then there's little nuances that pop up at the end when everything's already good. You've already went through the all little bumps. At the **[00:12:00]** end, by the way, we forgot that this particular scenario happens or that you can have two procedures at the same time. Let's say it's a medical thing. You are going through the clear path of someone just needs one procedure. But what about me either coming in and they actually need two or three or four procedures and they're going to do them in the same session. Do you need to then stack them up? And do you need to then consider that you're going to use more time? And so the agent has to know that now you're going to have to consider you need multiple sequential slots back-to-back and offer that instead of just one because it won't work any other way. But that didn't come up until the very end. So that happens and that's the type of things that you need to consider at the beginning and figure out and really nail down what the workflow looks like before doing all the AI stuff. Now let's take sales. A company buys an AI sales tool because it promises better outreach. And I've seen a lot of those and they're Trust me, they're they're very enticing. You see them and it's all automated and it researches your ICP for you and it generates the emails and it sounds like magic. But if the company has no clear qualification rules, the founder only knows a good lead when they see one but it's not written down, sales team uses a different language depending on who's writing, the CRM has incomplete fields, the best objections and replies are buried in old email threads and no one knows which ones are the ones that have worked. So what happens? The AI will write sales emails and they'll be fine and polite and professional, but I mean, it's nothing more than what an intern would write like just getting started. They won't sound like the company. They don't qualify people properly. They won't reflect the founder's judgment. They don't know which prospects are worth chasing. It might make a mistake that no real person would do and I've seen a lot of those where I get I get a lot of cold email and typically you just delete it. Sometimes you might get a good one and **[00:14:00]** that's okay and that's why it exists and and if someone targets me directly and knows about my business and that's the way I got that cold email and they actually did the research, then I'm not I'm not mad I received that cold email. It's like, oh, maybe this will help or maybe it won't. I don't know. But it's worth a conversation, maybe. But if I get an email where first of all, they don't have my name right, they have another person in the company's name in my email, then that tells me it's automated immediately. And then, aside from that, they they might be a competitor and they're they're offering my same services to me because they didn't even research a little bit about what we do. So, those are red flags where you just see it immediately, you delete it. Like, there's no chance you're going to give them a conversation because they they just blast they they did the least possible effort. So, if you can decrease some of those by having proper procedures, etc., then even if you use automated tools, you can make it sound like it's the real deal. So again, the tool is not the main problem. The workflow is not ready. So, if that same company maps the workflow first, the AI project makes it's going to get much stronger. They define what makes a qualified lead, they collect examples of strong replies, they write down the common objections, they decide which leads need further review, they clean up the CRM fields that matter, they create follow-up rules. Now, the AI has something useful to work with. The same model can produce a much better result because the workflow around it is clearer. Better workflow, better AI. Weak workflow, weak AI. There are exceptions, right? I mean, sometimes you just get something magical, but it's not enough to bet a lot of money into. So, how do you know whether a workflow is ready for AI? Well, use this readiness test. First, can you describe the workflow in plain English? If you cannot explain how the work moves from start to finish, you do not automate it yet. First, map it. I'll do a pause here. Sometimes, we don't know how to map. People will tell you, "Map the **[00:16:00]** workflow." Well, here's a nice use for AI, and this is not like actual deployment AI. This is just common sense AI. If you don't know how to map a workflow, then grab Claude Code or or any other co-work, if you want, or Code X, and explain what you're trying to do. Say, "I have this workflow. I'm trying to make heads and tails of it and understand what it really does, or figure out how we actually do it. Please help me put what I have in my mind down." And so, you can have this interaction with the LLM where you tell it this is step A, step B, step C. Please interrupt me at any moment if you have questions about maybe things that I might be missing. And so, the LLM is going to be very good in that interaction at asking you, "But what happens if this other thing happens, or what happens if this step is missing?" And so, you can go through that back and forth like you would do with a person. And obviously, if you have a person to do this with, do it with a person. But, this way, you would start to get those questions that you might not think of because you have it so ingrained in your mind because you do it every day. And then, you can ask it to, "Okay, great. Now, can you help me map it in Excalidraw or or another tool that's out there?" So, now these LLMs have connectors that allow you to connect to your typical tools. And so, that way, it's amazing because it can build out a whole workflow diagram. Maybe you know don't know the nomenclature and the types of what a circle means and what a square and what a rounded oval means and what the arrows mean and all that, but the LLMs do. Anything that's coded, they're very good at doing. So, once you explain the flow, then it can create a diagram for you and that will be clear when you're handing off to a development team and a implementation company, et cetera. You'll have something ready in your hands that can actually help and move the needle. Quick tip there that you may want to use. Second, does the workflow have a clear owner? If no one **[00:18:00]** owns the process, no one will own the AI output. That creates confusion fast. Third, are the inputs clear? What information starts the process? Where does it come from? Is it reliable? Is it structured? Does someone need to clean it before AI can use it? Fourth, what are the outputs and are they clear? What should the process produce? Does it generate an email? Does it generate a task somewhere? Does it create a report? A summary? A decision? A recommendation? If the output is vague, the system will be very hard to judge. You you won't be able to measure if it's doing the right thing. And I remember when Andrej Karpathy created this sort of self-iterating self-improvement loop uh with Auto Research. It sounds great and it works amazing, but you do need to do it with a measurable output. So, if your system is not generating a measurable output, then there's no way to run those types of amazing tools that already exist. But, if you do have a measurable output, then you can have the system or an agent iterate on itself and make itself better. Fifth, do you know what good looks like? This is one of the most overlooked questions. If a human cannot explain what a good result looks like, AI will not magically know. You need examples, good examples, bad examples, edge cases, terrible examples, review criteria, etc. This happened in a project a while back ago. It was they needed a lot of things, a lot of different outputs, but the problem is there were so many that they didn't have examples for all of them. So, you need to really quickly stop that process and and say, "Okay, let's step back. Let's either do an MVP or let's do a small segment of or subset of the things that we're going to do and make sure we have very clear examples of inputs, outputs, good examples, bad examples, etc. for each one. It's better to do five and have everything and then do the sixth **[00:20:00]** and have everything and the seventh and the eighth than to say we need 30, but then there's no way you're going to get all the examples and so you're just going to get trash data out. There's no way you're going to be able to make it better. So, that's another very important thing because especially if you start small with five and then you make the sixth and the seventh and the system can also review those and get a better idea, a better grasp of how they relate against each other and things like that to self-improve, which is very good. Now, do you know what needs human approval? Some work can be automatic. Some work should be drafted for review. Some work should be escalated. Some work should never be handled by AI. That boundary should be clear before you deploy. Seventh, can you measure improvement? If the workflow gets better, how will you know? Is it going to be time saved, faster response, fewer errors, more complete records, better hand-offs, less rework, higher customer satisfaction? A good AI project needs a scorecard because without one you're just judging by vibes and vibes are a very poor operating system. Now, let's turn this into a practical example. Suppose you run a professional services business, and I know many of you do, and it's actually very good space to get in right now. Client onboarding, unfortunately, in a service business is very messy. After a deal closes, the team has a task to collect all the information, set up folders, create tasks, schedule kickoff, ask for missing documents, and brief the delivery team. Right now, it depends on a project manager remembering all the steps. Now, someone suggests buying an AI project management tool, and maybe that tool helps, but most likely it does not. So, before buying anything, map the onboarding workflow. Step one, deal marked closed. Step two, onboarding email sent. Step three, intake form completed. Step four, missing information checked. Step five, project folder created. Step six, task list created. Step seven, kickoff scheduled. **[00:22:00]** Step eight, internal brief prepared. Step nine, client receives next steps. Now, look at each step and see, okay, which of these is manual? Which step is repetitive? Which one depends on judgment? Which one gets often delayed? Which step could be automated without risk? Which step could AI draft for review? You might find that you don't need a new AI platform. You might just need your CRM, forms, calendar, project management tool, and document templates connected properly, and that might not need AI. You might need AI only for the internal brief and first draft of the client email. That's a small like very, very much smaller project, and it's also more likely to work. And this is one of the big lessons for small businesses. The best AI project is often not the biggest one. It is the one that fits the workflow, and the workflow that fits the business is easier to adopt. Adoption matters more than people admit. A tool that nobody uses has a very consistent ROI, zero. Sometimes less than zero if it also creates meetings. And the team has to trust the system. They have to understand what's changed. They have to know when to rely on it and when to override it. They have to see that it removes work instead of adding another place to check. Because sometimes, again, we put AI in and it's just another thing we have to be checking and eventually we stop using it because it's impractical. That is why change management is not some enterprise buzzword. It is very practical. If your team does not change behavior, the project did not deploy. It only launched. There's a difference. Launch means the system exists. Deployment means the system is now part of how the works gets done. Many AI projects launch. There's There's too many to count. Fewer deploy. So, when you look at your business, do not ask, "What AI tool should we add?" Ask, "What behavior do we need to change? Should the sales team stop writing every follow-up from scratch? Should support stop answering **[00:24:00]** the same question manually? Should operations stop building the same report every week? Should managers stop chasing status updates in the chat? Should the founder stop being the only person who can approve routine decisions?" That is where the value hides. And once you know the behavior change, you can design a system around it. Here's another useful distinction. There are three kinds of AI tool purchases. The first is personal productivity. This is when individuals use AI to write, summarize, and brainstorm, research, or analyze. This is useful. It helps people work faster, but it is often the informal part. The second is team workflow. This is where AI becomes part of the shared process. The team agrees on how it works. There are rules. There's an owner. There's a repeatable output. This is where small businesses should focus. Third is operating system change. This is where the AI helps the company monitor workflows, learn from failures, update knowledge, and improve over time. That is powerful, but it usually comes after the team has learned to deploy simpler workflows. Most companies skip from personal productivity dreams straight to operating system fantasies. They miss the middle entirely, and the middle is where the real adoption happens. A lead process that works better, an onboarding process that is easier, a support process that is faster, a reporting process that saves our um a knowledge process that stops repeated questions. That's not glamorous or sexy, it's just useful. Useful wins. Now, if you're evaluating an AI tool right now, here are the questions I would ask before signing anything. Which specific workflow will this improve? What part of that workflow does it handle? What data does it need? Where does that data live today? Who will own the tool internally? What will the team stop doing if it works? What new behavior will the team need to adopt? What happens when the AI **[00:26:00]** is wrong? Can a human review high-risk output? How do we measure success after 30 days? What would make us cancel this tool? And the last question is useful. If you do not know what failure looks like, you probably do not know what success looks like, either. You should ask whether you need a tool, an automation, or a service. A tool gives you capability, an automation connects the steps, and a service helps design and implement the change. A lot of small businesses buy tools when what they actually need is implementation. That is why the forward deployed model is becoming important. Large AI companies are realizing that model access is not enough. Everyone has access, but still the ball isn't moving far enough as it should be. So, that's why um Anthropic, OpenAI recently created these huge companies precisely for deployment. They're building deployment teams because customers need help turning AI into real work. Small and medium businesses need the same kind of thinking just at a scale that fits. You probably do not need a massive enterprise team embedded for months, but the same applies. So, if they're doing it up that scale, you need to do it at the small and medium business scale as well. It's just a little different. You need someone to help you map the workflow, pick the right first project, build or configure the system, train the team, and measure the results. That is deployment. It's very different from buying a subscription. What I've seen with software in general, and a lot of companies that we work with, they see a demo like we were talking at the beginning, and they fall in love with the idea of it, but then once implementation once the sale goes through because the sales people are really good at what they're doing. But once the once the sales come through, then onboarding is supposed to happen, but it doesn't really happen, or people don't really follow it, or it's not really appropriate, and they end up not using it. And this happens a lot with with software as a service and with **[00:28:00]** on-premise software, etc. And the same thing happens and will happen with AI. If you just go and buy a tool and say, "Hey everyone, we got licenses for Gemini and ChatGPT and Claude or whatever." Yes, it might it might help some people do their job faster and better because of what the regular tooling, but if you really want to do it in a team effort kind of way, then you very likely will need someone to assist with that procedure. Or follow the steps from this video to ask the right questions. You don't necessarily need someone from the outside to ask those questions. Sometimes it helps to bring someone in just as additional support so that you make sure those questions get asked to the right people. So, here's a practical exercise for today. Pick one workflow that feels more annoying than it should. Do not pick the whole company. Do not pick operations. Just pick one loop. For example, new lead to first follow-up, or signed contract to kickoff, or support ticket to resolution, or invoice received to approved, or meeting ended to task assigned. Write the workflow at the top of the page, then answer these questions. Where does it start? Where does it end? Who owns it? What tools does it touch? What information is needed? What gets copied? What gets rewritten? Where do people wait? Where do mistakes happen? What does a good outcome look like? What should AI never do here? What could AI safely draft, summarize, route, check, or recommend? What could be a simple automation handled without AI? What would save the team the most time in the next 2 weeks? And I know there were a lot of questions, so if you want the list, just send us an email through our website or fill out one of our quizzes. And then when you get your report or whatever, I'll I'll be sure to send you a list of all that questions. I'll explain a little bit more about that, but now that I'm thinking of it, these sound like a lot of questions. They're very simple questions that you can definitely answer. We might just put a form with **[00:30:00]** them on the site, and then that way you can answer them there and get your personalized report. How does that sound? I think that that could be a good idea. I could could spin that off pretty quickly. If you do that and you do it honestly, you will learn more than you would from watching 10 more AI tool demos. And I like demos. Like I said, we do demos all the time. They're fun. They're just not necessarily strategic. Strategy starts with the work. And this is also how you avoid random AI adoption. Random AI adoption is when everyone experiments on their own and nothing compounds. One person has a good prompt, another person has a paid tool, someone else has a clever workflow, but the company doesn't learn from it. No skills are generated, and I'll talk about skills in another video. There's no shared playbook, no shared standard, no shared measurement, no internal owner. Workflow-first adoption is different. You pick a process, you improve it, you're you document what worked, you train the team, you measure the impact, then you use that learning on the next workflow, and that compounds. The first project teaches the second project, the second project will teach the third project, and over time the company gets better at deploying AI. That is a capability that you actually want. It's not just one tool or one automation, it's the ability to keep finding, improving, and owning the right workflows. That is much harder for competitors to copy because tools can be copied. Workflows are specific. Your customer expectations are specific. Your internal knowledge is specific. Your team habits are specific. Your exceptions are specific. That's why workflow matters. The tool should serve the workflow, not the other way around. So, the next time someone tells you that their AI tool can transform your business, don't dismiss it. Just slow the conversation down. Ask which workflow it improves. Ask what data it needs. Ask who owns the result. Ask how the team will use it every week. Ask what happens when it is wrong. Ask how **[00:32:00]** you will measure the improvement. And if those answers are clear, you may have a good opportunity. If those answers are vague, then you're probably looking at another shiny object. And small businesses do not need more shiny objects. They need working systems. They need fewer drop balls. They need faster follow-up. They need cleaner hand-offs. They need less copy-paste work. They need more of the company's knowledge available to the people doing the work, especially when there's turnover. And AI can help with a lot of that, but only when it's deployed into a workflow that is ready for it. So, here's a takeaway. Do not buy AI to fix confusion. Use workflow mapping to find the confusion, then decide where AI belongs, if it does at all. That is a much better order. And if you want help doing that, check on our website. There is probably by the time we're doing this, a deployment gap scorecard. And if that's not there, there there should be an AI automation opportunity map. It's a little quiz, three or five questions, etc., that you just go through, and it will help you figure out what your next steps should be. I think that's pretty much covers it. So, before you buy any other AI tool, find the work that needs to change. That is where the value is. Thanks everyone for joining, and we'll see you in the next one.