How to Use AI on Files You're Not Allowed to Upload

This week I used AI on a file I would This week I used AI on a file I would never ever upload cuz it was way too never ever upload cuz it was way too never ever upload cuz it was way too sensitive and the model never saw the sensitive and the model never saw the sensitive and the model never saw the original file. It still found the three original file. It still found the three original file. It still found the three assumptions most likely to break my assumptions most likely to break my assumptions most likely to break my launch after I removed all the sensitive launch after I removed all the sensitive launch after I removed all the sensitive stuff. The customer name, the home stuff. The customer name, the home stuff. The customer name, the home address, private medical note that was address, private medical note that was address, private medical note that was in there, an API key, and unreleased in there, an API key, and unreleased in there, an API key, and unreleased price. The model needed an operating price. The model needed an operating price. The model needed an operating plan. It sure didn't need any of that plan. It sure didn't need any of that plan. It sure didn't need any of that PII stuff. And the original, it stayed PII stuff. And the original, it stayed PII stuff. And the original, it stayed on my computer. And the useful work on my computer. And the useful work on my computer. And the useful work still got done by a frontier model. And still got done by a frontier model. And still got done by a frontier model. And that is the world I want to live in and that is the world I want to live in and that is the world I want to live in and I couldn't live in it, so I built it I couldn't live in it, so I built it I couldn't live in it, so I built it myself. Most privacy advice stops too myself. Most privacy advice stops too myself. Most privacy advice stops too soon. It says, you know, don't paste soon. It says, you know, don't paste soon. It says, you know, don't paste your sensitive information into AI. your sensitive information into AI. your sensitive information into AI. Raise your hand if you seen slides like Raise your hand if you seen slides like Raise your hand if you seen slides like that. Great. I agree. Now what? The that. Great. I agree. Now what? The that. Great. I agree. Now what? The contract still needs a risk review, contract still needs a risk review, contract still needs a risk review, right? The performance review still right? The performance review still right? The performance review still needs to be written and yes, you can needs to be written and yes, you can needs to be written and yes, you can talk into WhisperFlow and try and get a talk into WhisperFlow and try and get a talk into WhisperFlow and try and get a lot of your thoughts out, but it needs lot of your thoughts out, but it needs lot of your thoughts out, but it needs to be organized. Sensitive work does not to be organized. Sensitive work does not to be organized. Sensitive work does not stop being work. If the safe path means stop being work. If the safe path means stop being work. If the safe path means doing all of the cleanup by hand or doing all of the cleanup by hand or doing all of the cleanup by hand or giving up on AI as a useful tool, the giving up on AI as a useful tool, the giving up on AI as a useful tool, the warning hasn't solved the problem. It's warning hasn't solved the problem. It's warning hasn't solved the problem. It's just handed it back to us, right? That's just handed it back to us, right? That's just handed it back to us, right? That's why I built Airlock. This pricing plan why I built Airlock. This pricing plan why I built Airlock. This pricing plan that I'm going to show you here is that I'm going to show you here is that I'm going to show you here is synthetic. I made it for the camera, but synthetic. I made it for the camera, but synthetic. I made it for the camera, but the problem is not. Let me show you the problem is not. Let me show you the problem is not. Let me show you exactly what I kept, what I removed, and exactly what I kept, what I removed, and exactly what I kept, what I removed, and why. I'm choosing the pricing plan here.

why. I'm choosing the pricing plan here. why. I'm choosing the pricing plan here. Before I do anything else, Airlock asks Before I do anything else, Airlock asks Before I do anything else, Airlock asks me to define my protected terms. And me to define my protected terms. And me to define my protected terms. And this is where I can enter a customer this is where I can enter a customer this is where I can enter a customer name, a project name, an internal name, a project name, an internal name, a project name, an internal product code name, or any other ordinary product code name, or any other ordinary product code name, or any other ordinary looking phrase that means something looking phrase that means something looking phrase that means something specific and confidential inside my specific and confidential inside my specific and confidential inside my company. Now, why do I have to do that? company. Now, why do I have to do that? company. Now, why do I have to do that? Because a string like Project Lantern, Because a string like Project Lantern, Because a string like Project Lantern, that doesn't look private to a machine, that doesn't look private to a machine, that doesn't look private to a machine, but it might be a super sensitive phrase but it might be a super sensitive phrase but it might be a super sensitive phrase cuz it ties into a bunch of confidential cuz it ties into a bunch of confidential cuz it ties into a bunch of confidential work. Context is doing a lot of the work work. Context is doing a lot of the work work. Context is doing a lot of the work in understanding PII and confidential in understanding PII and confidential in understanding PII and confidential information. And some of that context information. And some of that context information. And some of that context exists only in our heads as we work. So, exists only in our heads as we work. So, exists only in our heads as we work. So, we have to give Airlock that context, we have to give Airlock that context, we have to give Airlock that context, and then Airlock can check the document, and then Airlock can check the document, and then Airlock can check the document, bring the patterns it recognizes, plus bring the patterns it recognizes, plus bring the patterns it recognizes, plus the terms I told it to protect, all the terms I told it to protect, all the terms I told it to protect, all under one screen. If there's any under one screen. If there's any under one screen. If there's any uncertainty, the starting choice is to uncertainty, the starting choice is to uncertainty, the starting choice is to just hide that item. That's our default, just hide that item. That's our default, just hide that item. That's our default, right? I can keep it, but I have to right? I can keep it, but I have to right? I can keep it, but I have to decide to keep it on purpose. And this decide to keep it on purpose. And this decide to keep it on purpose. And this is the question that matters a lot more is the question that matters a lot more is the question that matters a lot more than just building a nice detector. What than just building a nice detector. What than just building a nice detector. What do we need to do to get this job done, do we need to do to get this job done, do we need to do to get this job done, right? Like the whole point is to use AI right? Like the whole point is to use AI right? Like the whole point is to use AI to get useful work done. The model needs to get useful work done. The model needs to get useful work done. The model needs to know that all three warehouses move to know that all three warehouses move to know that all three warehouses move in the same week. So, we want to keep in the same week. So, we want to keep in the same week. So, we want to keep that detail. It needs to understand that that detail. It needs to understand that that detail. It needs to understand that the plan assumes an ERP integration will the plan assumes an ERP integration will the plan assumes an ERP integration will be ready before the first billing cycle.

be ready before the first billing cycle. be ready before the first billing cycle. We got to keep that. It needs to know We got to keep that. It needs to know We got to keep that. It needs to know that training is supposed to happen that training is supposed to happen that training is supposed to happen during normal shifts. Yeah, so we got to during normal shifts. Yeah, so we got to during normal shifts. Yeah, so we got to keep that, too. But does it need to know keep that, too. But does it need to know keep that, too. But does it need to know that this made-up person, Alice that this made-up person, Alice that this made-up person, Alice Meridian, has a home address? No, it Meridian, has a home address? No, it Meridian, has a home address? No, it doesn't need to know that stuff. Does it doesn't need to know that stuff. Does it doesn't need to know that stuff. Does it need her email? Obviously not. Does it need her email? Obviously not. Does it need her email? Obviously not. Does it need any private medical note that we need any private medical note that we need any private medical note that we have here from the operations director? have here from the operations director? have here from the operations director? Definitely not. So, all of these things Definitely not. So, all of these things Definitely not. So, all of these things that are personal, like API key, there's that are personal, like API key, there's that are personal, like API key, there's no conceivable reason for that no conceivable reason for that no conceivable reason for that credential to be inside the task. It credential to be inside the task. It credential to be inside the task. It just happens to be bundled into the just happens to be bundled into the just happens to be bundled into the files, which is how so much confidential files, which is how so much confidential files, which is how so much confidential information gets leaked into the cloud information gets leaked into the cloud information gets leaked into the cloud and into frontier models. It's just kind and into frontier models. It's just kind and into frontier models. It's just kind of bundled in. We don't really need it. of bundled in. We don't really need it. of bundled in. We don't really need it. Now, the price is a little more Now, the price is a little more Now, the price is a little more interesting. If I were asking the model interesting. If I were asking the model interesting. If I were asking the model to evaluate the pricing itself, the to evaluate the pricing itself, the to evaluate the pricing itself, the number might be part of the work. And number might be part of the work. And number might be part of the work. And for this question, what assumptions for this question, what assumptions for this question, what assumptions could break the operating plan, it's could break the operating plan, it's could break the operating plan, it's not. So, in this case, I can leave the not. So, in this case, I can leave the not. So, in this case, I can leave the price behind as well. So, I have to price behind as well. So, I have to price behind as well. So, I have to start with a job, not the file. The same start with a job, not the file. The same start with a job, not the file. The same price can be essential for one question price can be essential for one question price can be essential for one question and irrelevant for another. And once I and irrelevant for another. And once I and irrelevant for another. And once I make those choices, Airlock builds a new make those choices, Airlock builds a new make those choices, Airlock builds a new Word document for me. It does not draw Word document for me. It does not draw Word document for me. It does not draw black rectangles over the old one, black rectangles over the old one, black rectangles over the old one, right? That matters because Word files right? That matters because Word files right? That matters because Word files are strange little containers, right?

are strange little containers, right? are strange little containers, right? Comments, track changes, author names, Comments, track changes, author names, Comments, track changes, author names, old edits, and external relationships old edits, and external relationships old edits, and external relationships can accidentally remain even when the can accidentally remain even when the can accidentally remain even when the page you're looking at looks clean. So, page you're looking at looks clean. So, page you're looking at looks clean. So, the safer approach is simply to rebuild the safer approach is simply to rebuild the safer approach is simply to rebuild the approved material into a separate the approved material into a separate the approved material into a separate file and leave the original alone. Now, file and leave the original alone. Now, file and leave the original alone. Now, I have two docs. The original is still I have two docs. The original is still I have two docs. The original is still untouched on my computer. The smaller untouched on my computer. The smaller untouched on my computer. The smaller copy contains the warehouse plan and the copy contains the warehouse plan and the copy contains the warehouse plan and the relationships the model needs without relationships the model needs without relationships the model needs without the customer identity, without the home the customer identity, without the home the customer identity, without the home address, without all the stuff I talked address, without all the stuff I talked address, without all the stuff I talked about, right? The personal and about, right? The personal and about, right? The personal and confidential information that came along confidential information that came along confidential information that came along for the ride, to be honest. So, I open for the ride, to be honest. So, I open for the ride, to be honest. So, I open the clean copy and I read it. This is the clean copy and I read it. This is the clean copy and I read it. This is the only version I have to review now. the only version I have to review now. the only version I have to review now. And then I give that copy to any And then I give that copy to any And then I give that copy to any frontier model I want. And I ask, "Hey, frontier model I want. And I ask, "Hey, frontier model I want. And I ask, "Hey, can you help me think through this? Can can you help me think through this? Can can you help me think through this? Can you help me identify the assumptions you help me identify the assumptions you help me identify the assumptions that are most likely to break this plan that are most likely to break this plan that are most likely to break this plan before launch? Explain them to me before launch? Explain them to me before launch? Explain them to me clearly, and can you recommend a clearly, and can you recommend a clearly, and can you recommend a mitigation for me?" Right, this is the mitigation for me?" Right, this is the mitigation for me?" Right, this is the kind of work we like to do with our kind of work we like to do with our kind of work we like to do with our frontier models cuz it requires frontier models cuz it requires frontier models cuz it requires thoughtfulness. And look at what comes thoughtfulness. And look at what comes thoughtfulness. And look at what comes back. The model catches that the back. The model catches that the back. The model catches that the all-at-once warehouse migration and the all-at-once warehouse migration and the all-at-once warehouse migration and the ERP readiness assumption and the ERP readiness assumption and the ERP readiness assumption and the assumption that training will reduce assumption that training will reduce assumption that training will reduce throughput. They're all pretty big throughput. They're all pretty big throughput. They're all pretty big assumptions. They're load-bearing. It assumptions. They're load-bearing. It assumptions. They're load-bearing. It recommends staging the rollout, proving recommends staging the rollout, proving recommends staging the rollout, proving the integration before the first billing the integration before the first billing the integration before the first billing cycle, and building training time into cycle, and building training time into cycle, and building training time into the operating plan. Hey, these sound the operating plan. Hey, these sound the operating plan. Hey, these sound pretty sensible. This is why I built pretty sensible. This is why I built pretty sensible. This is why I built this. I don't want privacy to become a this. I don't want privacy to become a this. I don't want privacy to become a second project that cancels out the time

second project that cancels out the time second project that cancels out the time AI was supposed to save. I want help AI was supposed to save. I want help AI was supposed to save. I want help separating the information the task separating the information the task separating the information the task needs from the information the file needs from the information the file needs from the information the file happens to contain, and then I want to happens to contain, and then I want to happens to contain, and then I want to get back to work. get back to work. get back to work. I want to get back to work. Now, why I want to get back to work. Now, why I want to get back to work. Now, why does this feel newly urgent? Sensitive does this feel newly urgent? Sensitive does this feel newly urgent? Sensitive documents and cloud software have both documents and cloud software have both documents and cloud software have both been around for a really long time. One, been around for a really long time. One, been around for a really long time. One, I think AI changed the amount of I think AI changed the amount of I think AI changed the amount of information we want to move. A couple of information we want to move. A couple of information we want to move. A couple of years ago, a normal AI interaction was a years ago, a normal AI interaction was a years ago, a normal AI interaction was a general question in an empty chatbot. general question in an empty chatbot. general question in an empty chatbot. Write the email explain the concept, Write the email explain the concept, Write the email explain the concept, give me 10 ideas. That was super normal give me 10 ideas. That was super normal give me 10 ideas. That was super normal in 2024. A lot of the useful work begins in 2024. A lot of the useful work begins in 2024. A lot of the useful work begins with your material. The the real with your material. The the real with your material. The the real detailed proposal you are actually detailed proposal you are actually detailed proposal you are actually sending, the contract you're actually sending, the contract you're actually sending, the contract you're actually negotiating, the meeting notes that negotiating, the meeting notes that negotiating, the meeting notes that nobody really organized, right? Or or nobody really organized, right? Or or nobody really organized, right? Or or the code base your team is actually the code base your team is actually the code base your team is actually trying to understand. trying to understand. trying to understand. This is a much bigger chunk of work. This is a much bigger chunk of work. This is a much bigger chunk of work. We've we've hundred x or a thousand x We've we've hundred x or a thousand x We've we've hundred x or a thousand x the kind of data that we can reasonably the kind of data that we can reasonably the kind of data that we can reasonably give to a model. And for work like this, give to a model. And for work like this, give to a model. And for work like this, the relevant context is exactly what the relevant context is exactly what the relevant context is exactly what makes the model useful.

makes the model useful. makes the model useful. And that's why these files are so And that's why these files are so And that's why these files are so tempting to upload. Look, I get it. You tempting to upload. Look, I get it. You tempting to upload. Look, I get it. You want the model to do useful work, you want the model to do useful work, you want the model to do useful work, you have to give it a lot of context to do have to give it a lot of context to do have to give it a lot of context to do useful work, so you kind of need to put useful work, so you kind of need to put useful work, so you kind of need to put the docs in, but because you have 10 x the docs in, but because you have 10 x the docs in, but because you have 10 x or 100 x more docs, it's really hard to or 100 x more docs, it's really hard to or 100 x more docs, it's really hard to clean them all out. But along the way clean them all out. But along the way clean them all out. But along the way from 2024 to 2026, what happened to us? from 2024 to 2026, what happened to us? from 2024 to 2026, what happened to us? We we made intelligence almost We we made intelligence almost We we made intelligence almost frictionless, and then we handed all of frictionless, and then we handed all of frictionless, and then we handed all of us the job of deciding what can leave, us the job of deciding what can leave, us the job of deciding what can leave, where it can go, what has to remain for where it can go, what has to remain for where it can go, what has to remain for the answer to be useful. It's like we're the answer to be useful. It's like we're the answer to be useful. It's like we're all individually privacy filters on top all individually privacy filters on top all individually privacy filters on top of our other jobs. of our other jobs. of our other jobs. And nobody, including me, grew up with And nobody, including me, grew up with And nobody, including me, grew up with instincts for that. instincts for that. instincts for that. Word can hide old comments in places Word can hide old comments in places Word can hide old comments in places many of us don't look, right? We have many of us don't look, right? We have many of us don't look, right? We have retention, we have memory, we have retention, we have memory, we have retention, we have memory, we have enterprise logging, those are all enterprise logging, those are all enterprise logging, those are all different things. Yet we ask people to different things. Yet we ask people to different things. Yet we ask people to work it out when the upload button can work it out when the upload button can work it out when the upload button can save an hour. And then we put pressure save an hour. And then we put pressure save an hour. And then we put pressure on them to deliver faster. So no wonder on them to deliver faster. So no wonder on them to deliver faster. So no wonder the upload button is really tempting, the upload button is really tempting, the upload button is really tempting, right? I asked people in my community right? I asked people in my community right? I asked people in my community this week what they actually do around this week what they actually do around this week what they actually do around privacy, not what the policy says they privacy, not what the policy says they privacy, not what the policy says they should do. One person gave me an answer should do. One person gave me an answer should do. One person gave me an answer I cannot stop thinking about. I am I cannot stop thinking about. I am I cannot stop thinking about. I am relying on trust more than I'm relying on trust more than I'm relying on trust more than I'm comfortable admitting.

comfortable admitting. comfortable admitting. That's a really honest answer. And this That's a really honest answer. And this That's a really honest answer. And this was not someone being casual, right? was not someone being casual, right? was not someone being casual, right? This is someone who has a team plan with This is someone who has a team plan with This is someone who has a team plan with training off, scoped connectors, and training off, scoped connectors, and training off, scoped connectors, and agents limited to development data. agents limited to development data. agents limited to development data. They're being really responsible, right? They're being really responsible, right? They're being really responsible, right? They're doing the things a sophisticated They're doing the things a sophisticated They're doing the things a sophisticated user is supposed to do. And still, the user is supposed to do. And still, the user is supposed to do. And still, the just don't paste anything sensitive rule just don't paste anything sensitive rule just don't paste anything sensitive rule has stopped being a realistic operating has stopped being a realistic operating has stopped being a realistic operating model. And I think that that's just model. And I think that that's just model. And I think that that's just true. Like at this point, asking someone true. Like at this point, asking someone true. Like at this point, asking someone to not do that is sort of malpractice to not do that is sort of malpractice to not do that is sort of malpractice because the models need so much data and because the models need so much data and because the models need so much data and our sensitive information is so our sensitive information is so our sensitive information is so intermixed. Someone who works in health intermixed. Someone who works in health intermixed. Someone who works in health care gave me the opposite answer in my care gave me the opposite answer in my care gave me the opposite answer in my community. The cost of one public community. The cost of one public community. The cost of one public incident involving patient information incident involving patient information incident involving patient information feels so high that his organization just feels so high that his organization just feels so high that his organization just mostly keeps all the valuable data away mostly keeps all the valuable data away mostly keeps all the valuable data away from AI. They use AI only around the from AI. They use AI only around the from AI. They use AI only around the edges where the information looks much edges where the information looks much edges where the information looks much more like every other industry. While more like every other industry. While more like every other industry. While the data that could support the most the data that could support the most the data that could support the most powerful work remains absolutely out of powerful work remains absolutely out of powerful work remains absolutely out of reach because it's just too risky. Think reach because it's just too risky. Think reach because it's just too risky. Think about those two answers. One careful about those two answers. One careful about those two answers. One careful person goes forward and relies partly on person goes forward and relies partly on person goes forward and relies partly on trust. Another careful person abstains.

trust. Another careful person abstains. trust. Another careful person abstains. This is why I don't think another annual This is why I don't think another annual This is why I don't think another annual training course solves the problem. The training course solves the problem. The training course solves the problem. The benefit arrives immediately and the benefit arrives immediately and the benefit arrives immediately and the possible cost is delayed. It's possible cost is delayed. It's possible cost is delayed. It's uncertain. It's often invisible. Verizon uncertain. It's often invisible. Verizon uncertain. It's often invisible. Verizon saw this in its enterprise telemetry saw this in its enterprise telemetry saw this in its enterprise telemetry this year. The share of employees using this year. The share of employees using this year. The share of employees using an AI platform at least once every 15 an AI platform at least once every 15 an AI platform at least once every 15 days on a corporate device rose from 15% days on a corporate device rose from 15% days on a corporate device rose from 15% to 45%. And among those users 2/3 of to 45%. And among those users 2/3 of to 45%. And among those users 2/3 of them were accessing AI through them were accessing AI through them were accessing AI through non-company accounts. It's the shadow IT non-company accounts. It's the shadow IT non-company accounts. It's the shadow IT problem. In the data policy events problem. In the data policy events problem. In the data policy events Verizon observed involving outside Verizon observed involving outside Verizon observed involving outside systems, source code was the most common systems, source code was the most common systems, source code was the most common material submitted because again, material submitted because again, material submitted because again, there's that pressure to work and there's that pressure to work and there's that pressure to work and deliver more value. Now, you can read deliver more value. Now, you can read deliver more value. Now, you can read that as a story about irresponsible that as a story about irresponsible that as a story about irresponsible employees. I think that misses the more employees. I think that misses the more employees. I think that misses the more useful question here. useful question here. useful question here. Why is the unapproved route so much Why is the unapproved route so much Why is the unapproved route so much easier to find and use than the approved easier to find and use than the approved easier to find and use than the approved route? NIST has a very plain phrase for route? NIST has a very plain phrase for route? NIST has a very plain phrase for what happens next, security fatigue.

what happens next, security fatigue. what happens next, security fatigue. When the same security decisions pile When the same security decisions pile When the same security decisions pile up, the easiest option starts to win. up, the easiest option starts to win. up, the easiest option starts to win. We already know how to build around We already know how to build around We already know how to build around that, right? Your phone asks about the that, right? Your phone asks about the that, right? Your phone asks about the camera when an app wants the camera, not camera when an app wants the camera, not camera when an app wants the camera, not randomly, right? Your browser handles randomly, right? Your browser handles randomly, right? Your browser handles the certificate check without asking you the certificate check without asking you the certificate check without asking you to become some kind of cryptographer to become some kind of cryptographer to become some kind of cryptographer first. We know that we can't expect first. We know that we can't expect first. We know that we can't expect users to carry that load in so many users to carry that load in so many users to carry that load in so many other parts of software, but AI has a other parts of software, but AI has a other parts of software, but AI has a lot of this backwards. We put lot of this backwards. We put lot of this backwards. We put astonishing intelligence behind an empty astonishing intelligence behind an empty astonishing intelligence behind an empty box and then put the privacy system in box and then put the privacy system in box and then put the privacy system in policy pages, in admin consoles, in policy pages, in admin consoles, in policy pages, in admin consoles, in document menus, in vendor contracts, and document menus, in vendor contracts, and document menus, in vendor contracts, and the user is expected to hold all of this the user is expected to hold all of this the user is expected to hold all of this in memory and somehow make the right in memory and somehow make the right in memory and somehow make the right decision. Do you always make the right decision. Do you always make the right decision. Do you always make the right decision? It's right? Like do you? Do I? decision? It's right? Like do you? Do I? decision? It's right? Like do you? Do I? How can we know? It lives in our How can we know? It lives in our How can we know? It lives in our memories. The obvious answer is memories. The obvious answer is memories. The obvious answer is redaction. Just remove the sensitive redaction. Just remove the sensitive redaction. Just remove the sensitive information. But the more I worked on information. But the more I worked on information. But the more I worked on building Airlock, the more I realized building Airlock, the more I realized building Airlock, the more I realized that redaction is easy if you don't care that redaction is easy if you don't care that redaction is easy if you don't care whether the model can still help you.

whether the model can still help you. whether the model can still help you. It's real hard if you care whether the It's real hard if you care whether the It's real hard if you care whether the model can still help you. model can still help you. model can still help you. Because if you just say delete every Because if you just say delete every Because if you just say delete every name, every number, every date, every name, every number, every date, every name, every number, every date, every role, every relationship, every price, role, every relationship, every price, role, every relationship, every price, and every location, you can delete all and every location, you can delete all and every location, you can delete all of that and you can produce a of that and you can produce a of that and you can produce a wonderfully empty document. It will also wonderfully empty document. It will also wonderfully empty document. It will also be useless because the hard part is be useless because the hard part is be useless because the hard part is meaning making. Which facts make this meaning making. Which facts make this meaning making. Which facts make this task solvable and which facts only came task solvable and which facts only came task solvable and which facts only came attached for the ride and we can get rid attached for the ride and we can get rid attached for the ride and we can get rid of them. That is why when I demonstrated of them. That is why when I demonstrated of them. That is why when I demonstrated Airlock earlier in this video, I called Airlock earlier in this video, I called Airlock earlier in this video, I called out that you have to make some of those out that you have to make some of those out that you have to make some of those decisions. Take a contract. If you want decisions. Take a contract. If you want decisions. Take a contract. If you want AI to find renewal dates or unusual AI to find renewal dates or unusual AI to find renewal dates or unusual indemnity language or termination indemnity language or termination indemnity language or termination obligations, the model may need the obligations, the model may need the obligations, the model may need the clauses and the dates. It probably clauses and the dates. It probably clauses and the dates. It probably doesn't need the home addresses of the doesn't need the home addresses of the doesn't need the home addresses of the signatories, right? Negotiated prices, signatories, right? Negotiated prices, signatories, right? Negotiated prices, well, the model's going to need the well, the model's going to need the well, the model's going to need the numbers, right? There's no way around numbers, right? There's no way around numbers, right? There's no way around that. If you want to rewrite the cover that. If you want to rewrite the cover that. If you want to rewrite the cover email, maybe you don't need those email, maybe you don't need those email, maybe you don't need those numbers. It's about your intent. And numbers. It's about your intent. And numbers. It's about your intent. And I've said that a lot on this channel. I've said that a lot on this channel. I've said that a lot on this channel. It's about intent and understanding what It's about intent and understanding what It's about intent and understanding what we want to do with the models. Take a we want to do with the models. Take a we want to do with the models. Take a medical record. Sometimes the full medical record. Sometimes the full medical record. Sometimes the full history is the reason the analysis has history is the reason the analysis has history is the reason the analysis has value. Removing the sensitive value. Removing the sensitive value. Removing the sensitive information would remove the task. And information would remove the task. And information would remove the task. And in that case, Airlock, honestly, it's in that case, Airlock, honestly, it's in that case, Airlock, honestly, it's the wrong route. The work belongs in a the wrong route. The work belongs in a the wrong route. The work belongs in a governed environment built for the full governed environment built for the full governed environment built for the full record or it should not be involved with record or it should not be involved with record or it should not be involved with AI at all. This is why I keep coming AI at all. This is why I keep coming AI at all. This is why I keep coming back to the job. What are you asking the back to the job. What are you asking the back to the job. What are you asking the model to do? What is the minimum context

model to do? What is the minimum context model to do? What is the minimum context that makes the answer useful? And once that makes the answer useful? And once that makes the answer useful? And once you know that, you can start making you know that, you can start making you know that, you can start making intelligent decisions about the file. intelligent decisions about the file. intelligent decisions about the file. The next time you open a document and The next time you open a document and The next time you open a document and you think AI could help, but I can't you think AI could help, but I can't you think AI could help, but I can't upload this file, begin with the job. upload this file, begin with the job. upload this file, begin with the job. What does the model actually need to see What does the model actually need to see What does the model actually need to see and where is that smaller copy allowed and where is that smaller copy allowed and where is that smaller copy allowed to go? You should not have to become a to go? You should not have to become a to go? You should not have to become a privacy engineer before lunch to get privacy engineer before lunch to get privacy engineer before lunch to get your work done. AI made intelligence your work done. AI made intelligence your work done. AI made intelligence frictionless. Intelligence needs frictionless. Intelligence needs frictionless. Intelligence needs information and safety has to live in information and safety has to live in information and safety has to live in the same path as convenience to make all the same path as convenience to make all the same path as convenience to make all of this work. So, that's why I made it. of this work. So, that's why I made it. of this work. So, that's why I made it. If you want to check it out, the link is If you want to check it out, the link is If you want to check it out, the link is below. Tell me what you think. Tell me below. Tell me what you think. Tell me below. Tell me what you think. Tell me how you're dealing with privacy. I'd how you're dealing with privacy. I'd how you're dealing with privacy. I'd love to see some privacy horror stories. love to see some privacy horror stories. love to see some privacy horror stories. Give me some privacy horror stories in Give me some privacy horror stories in Give me some privacy horror stories in the comments. the comments. the comments. And let's build something great And let's build something great And let's build something great together.