<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Gregory Gimenez on Medium]]></title>
        <description><![CDATA[Stories by Gregory Gimenez on Medium]]></description>
        <link>https://medium.com/@gregory_gimenez?source=rss-88607504616a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Gregory Gimenez on Medium</title>
            <link>https://medium.com/@gregory_gimenez?source=rss-88607504616a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 07:25:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gregory_gimenez/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[One Email, Many People]]></title>
            <link>https://medium.com/@gregory_gimenez/one-email-many-people-7b51c936e088?source=rss-88607504616a------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b51c936e088</guid>
            <category><![CDATA[salesforce]]></category>
            <category><![CDATA[data-architecture]]></category>
            <category><![CDATA[master-data-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Gregory Gimenez]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 17:59:45 GMT</pubDate>
            <atom:updated>2026-11-11T17:59:45.254Z</atom:updated>
            <content:encoded><![CDATA[<h4>Nobody lied. Nobody was wrong. “Master data” simply meant two different things.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_JP6fwPxQamv65rUtATY3w.png" /></figure><p>We had been handed a customer master. A reference, a single source of truth for who the customers were. That is what the word promised, and we took it at face value, the way anyone would.</p><p>Then we looked at the data. The same email address kept coming back, and each time the name behind it was different. First name, last name, both. One stable identifier, and a person who would not hold still.</p><p>It was not an error in the data. It was how these customers actually behaved. One contact booked for a partner, for the children, for a friend, for someone doing a favor. The email was the account. The people passing through it changed by the week.</p><blockquote>Nobody lied. Nobody was wrong. We read one word two different ways.</blockquote><p>To the client, “master data” meant exactly what they had, a record of their customers as they knew them, reached through whoever did the booking. To us, the same word carried a quiet promise, one stable identity per customer, the thing you build a model on. The word made us all nod. Underneath, we had set off in opposite directions, each sure we were aligned.</p><p>That is the thing about this kind of gap. It does not show up in a meeting. Everyone uses the same term, everyone agrees, everyone means something slightly different. It only shows up in the data, when the identifier you trusted turns out to point at more than one person.</p><p>So the real question is when you catch it. In pre-sales would be ideal, and it is usually out of reach. At that stage you have the words, not the data. You answer quickly, on what the client tells you, and a word like “master data” sounds as solid to you as it does to them. Missing it there is not a failure, it is the nature of the moment.</p><p>The real appointment is scoping. There you have a little time, and you can ask to see a real sample before anything is fixed. That is where a divergence like this should surface, while it still costs almost nothing to resolve.</p><p>And that is where AI earns its place. Point it at a real extract early, and it needs no hint. It sees that one email maps to many names, that the identifier is not the identity, and it says so plainly. Not to prove anyone right. To put the gap in black and white, and turn an unspoken difference into a shared fact. Once it is on the page, drawn from the client’s own data, it stops being one person’s reading against another’s. It becomes something everyone can see, and therefore something everyone has to weigh.</p><p>Then the real conversation can happen, the one worth having before a line of code is written. What is a customer here. A person, or a point of contact. There is no universal answer, and that is the point. It is a business decision, and it stays one. The data cannot make it, and neither can the AI. All the AI did was make sure the question got asked in time, on evidence no one could wave away. And because this is personal data, how you answer carries real weight, so you handle it with care.</p><p>Looking back, that is the lesson I keep. A master data model should not describe the customer a word lets you imagine. It should describe the one the data, and a shared definition, actually point to. The two are not always the same, and the gap between them stays quiet until it turns expensive.</p><p>The job, then, is to align the words with the data as early as you can. Ideally in pre-sales, realistically at scoping, but early, while a difference of meaning is still just a conversation and not a rebuild. AI does not settle what the words should mean. It makes the gap visible soon enough that everyone takes it into account. The deciding stays ours. The seeing, it can give us sooner.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b51c936e088" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI Works for the Client, Not Just for Us]]></title>
            <link>https://medium.com/@gregory_gimenez/when-ai-works-for-the-client-not-just-for-us-d6c48366003a?source=rss-88607504616a------2</link>
            <guid isPermaLink="false">https://medium.com/p/d6c48366003a</guid>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[salesforce]]></category>
            <category><![CDATA[consulting]]></category>
            <dc:creator><![CDATA[Gregory Gimenez]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 18:17:52 GMT</pubDate>
            <atom:updated>2026-11-10T18:17:52.893Z</atom:updated>
            <content:encoded><![CDATA[<h4>We judge AI by how fast it makes us. The better question is what the client gets to keep.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IX4zi_33_Jawh9LjypMHRw.png" /></figure><p>When we talk about AI, we talk about ourselves. How fast we work, what we save, how many hours we claw back. We rarely stop to ask what the client actually gets out of it. One old bid keeps reminding me to.</p><p>It was a small engagement. Part of it was straightforward, bring products, prices, quotes, and orders into the CRM, the usual backbone of a sales setup. But one piece sat apart. The heart of how this company priced and quoted lived in a bespoke application of its own, built over years, outside the CRM and outside almost anyone’s full understanding.</p><p>On paper, folding that piece into the same system was the right move. One place for the commercial process, fewer bridges to maintain, a cleaner map of the business. Everyone in the room could see the sense in it.</p><p>And everyone could also see why it would not happen. The logic in that application was dense and barely documented. Understanding it well enough to rebuild it safely would have taken weeks of careful, expensive work before a single useful line was written. Measured against the budget, it did not add up. So the piece stayed where it was. A silo, kept not by preference but by arithmetic.</p><p>That arithmetic is the part worth looking at again. The reason to leave the silo in place was never technical. It was the cost of the one task standing in front of everything else, understanding a large, undocumented body of logic and turning it into something you can act on.</p><p>That task is exactly where AI has become good. Not at deciding, at reading. Give it the tangle and it can map the rules, surface the patterns, flag the contradictions, and draft a first version of the specification that used to take a specialist weeks. The analysis that priced the whole idea out of reach gets a lot cheaper.</p><p>And when that cost drops, the arithmetic flips. What was too expensive to unify becomes something you can at least seriously consider. The piece that had to stay in its corner might finally come home.</p><p>Notice who gains when that happens. Not us, or not mainly us. The client. One system instead of two, one source of truth for how they price and sell, less shadow tooling to keep alive, less quiet debt piling up in a corner no one wants to touch. The value does not land on our side of the table. It lands on theirs.</p><p>I want to be honest about the edges of this, because they matter. AI made the analysis cheaper, it did not make the judgment optional. It does not know which pieces are worth unifying and which are better left alone. It does not own the business logic, and it never will. The rules it reads out of an old system still have to be checked, line by line, with the people who have carried them for years, because a confident summary of the wrong rule is worse than no summary at all. And a cheaper analysis is not a license to pull everything into one place. Some silos exist for good reasons. The point is to choose on purpose, not to be forced by the cost of looking.</p><p>Looking back, that is what I would do differently with the tools I have now. Not promise a faster build. Offer the client a cheaper way to understand their own system, so the decision about what to unify could be made on its merits instead of its price.</p><p>That is the version of AI I find most interesting, and the one we talk about least. Not the assistant that makes us quicker. The one that lowers the cost of understanding, until the good decisions a client had quietly shelved, for no reason other than the expense of looking, come back onto the table.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d6c48366003a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Bid We Could Not Have Answered by Hand]]></title>
            <link>https://medium.com/@gregory_gimenez/the-bid-we-could-not-have-answered-by-hand-b8f38f6c82e9?source=rss-88607504616a------2</link>
            <guid isPermaLink="false">https://medium.com/p/b8f38f6c82e9</guid>
            <category><![CDATA[presales]]></category>
            <category><![CDATA[consulting]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[proposal-management]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Gregory Gimenez]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 12:31:59 GMT</pubDate>
            <atom:updated>2026-11-10T12:31:59.550Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI did not write our proposal. It gave us back the hours to think about it.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7CiZ6d-wb03Dou7Pd1zSCA.png" /></figure><p>There was a bid I still think about. Not because of how it ended, but because of what it would have taken to answer it the old way.</p><p>The specification was long. The questions ran into the dozens. The window to respond was short, short enough that reading the document properly, from the first page to the last, would have eaten most of the time we had. Every proposal is tight. This one was past tight. Done by hand, we would not have made it. Not well, and not on time.</p><p>I used to think of a proposal as a writing job. It is a reading job first. You cannot answer a specification you have not fully understood, and understanding a long one against a short clock is where the hours quietly disappear.</p><p>That is the part where AI changed how I work. Not as a clever writer that speaks for me. As an assistant that takes the heavy, repetitive load off the front of the process and hands me back time I did not have.</p><p>So that is where we put it. It went through the whole specification and pulled every requirement into a compliance matrix, so nothing buried deep in the document slipped past unseen. It drafted first versions of the answers we had written many times before. It gave the pricing a structure to fill, section by section, instead of a blank page under pressure.</p><p>What it did not do was win the bid. It had no view on the price that would make us competitive without giving the margin away. It did not know the client, the history, the one angle that would make our answer land. Those stayed with us. They always do, and they are the parts that actually decide.</p><blockquote>On that bid, the AI was not a comfort. It was the condition for being in the race at all.</blockquote><p>Which brings me to the part I did not expect to write.</p><p>When a client sets a deadline that no team could meet by hand, they have already made a decision about method. They may not say it, but they have accepted that the answer will not be produced the old way. Modern tooling, AI included, is now part of what everyone quietly assumes. We keep up a polite fiction when we pretend otherwise.</p><p>If the deadline is an implicit yes to AI on our side, then honesty should not stay implicit either. We should say it. And we should stay responsible for every line, because that is what turns an unspoken assumption into something both sides can actually trust.</p><p>Transparency has its own limits, and they matter. Being open about using AI does not mean saying a machine wrote the answer. It means owning that it helped, while a human stays accountable for the content, the accuracy, and the judgment. A tight deadline is not permission to do anything. If a client sets rules, or rules it out, we follow that, or we negotiate the time. And openness never comes at the cost of confidentiality. You can be transparent that you used AI without ever feeding a client’s specification into a public tool. Honest about the method, careful with the data.</p><p>Looking back, the lesson was not about speed for its own sake. The AI did not write our proposal. It gave us back the hours to think about it, and on that timeline, it gave us the chance to answer at all. The time it returned went straight to the work that wins, the strategy, the price, the one idea that sets you apart.</p><p>That is the trade I will take every time. Let the assistant carry the load it carries well. Keep the judgment, and the responsibility, where they belong. And tell the client, plainly, how the work was done. The deadline already took the help for granted. The least we can do is be open about it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b8f38f6c82e9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Skew in Salesforce: Would the AI Have Caught It?]]></title>
            <link>https://medium.com/@gregory_gimenez/data-skew-in-salesforce-would-the-ai-have-caught-it-645315fc12bd?source=rss-88607504616a------2</link>
            <guid isPermaLink="false">https://medium.com/p/645315fc12bd</guid>
            <category><![CDATA[data-skew]]></category>
            <category><![CDATA[data-architecture]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[salesforce]]></category>
            <category><![CDATA[solution-architecture]]></category>
            <dc:creator><![CDATA[Gregory Gimenez]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 15:48:46 GMT</pubDate>
            <atom:updated>2026-11-09T15:51:24.538Z</atom:updated>
            <content:encoded><![CDATA[<h4>The data skew was never hidden. It was in the numbers we already had.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s6fCylFje3z4bosyoWEglw.png" /></figure><p>In 2021, finding a data skew in Salesforce was something you did by hand. It was slow, a little tedious, and it usually came after the platform had already started to hurt. Sharing recalculations piling up, locks, transactions timing out. You went looking for the cause once the symptoms were loud enough to force you to.</p><p>We had run into one. The reason was not carelessness, it was constraints. The data was segmented for legal and operational reasons, and a lot of permissions were set manually and dynamically. So by default, most records ended up owned by a small set of generic users.</p><p>On paper that is reasonable. In Salesforce it breaks a rule that does not forgive much. When a single owner holds more than ten thousand records of an object, every change to the role hierarchy or the sharing rules triggers long, heavy recalculations across the sharing tables. That is ownership data skew, and at scale it brings the org to its knees.</p><p>This year I felt like I had changed jobs. When I step back, it is more nuanced than that. I am still an architect, doing the work I have always done. What changed is the tooling and the method. The questions are the same. The way I ask them, and when, is not.</p><p>So I ran a small experiment on myself. I went back to that project and asked one thing. If I had handed the design documentation to an AI the day the specifications were signed off, with the volumetry we already had, would it have seen the skew coming?</p><p>The setup matters, so let me be precise. Specs just finished. No production yet, no incident to point at. The only things I give the AI are the design document and the volumetry, including the record counts per owner on the largest objects. That last part is the key. We knew those numbers from the start. I am not feeding it the answer after the fact, I am feeding it what was already on the table before anyone wrote a line of code. Then I ask a single question. Reading this, where do you see a scalability or sharing risk?</p><p>The answer is yes, and it is not a lucky guess. It reads the numbers, it does not divine them. Core objects at tens of millions of records. Sharing tables sized, in the design itself, up to close to a hundred million rows. A role hierarchy seven levels deep, with cross-visibility handled through public groups. And a small set of generic owners concentrating a large share of the records, far past the ten thousand per owner line.</p><p>Put those four facts next to each other and there is nothing left to debate. This is ownership data skew, with a sharing recalculation problem sitting right behind it. Anything that touches sharing later, a merge, a reassignment, a reload, will recalculate in bulk, hit the ten thousand row limit, and burn through the Apex CPU budget.</p><p>I want to be honest about the one thing the AI would not have known. It would not have known why the records were owned by generic users. That choice came from legal and operational constraints that are nowhere in the volumetry. It could not infer the reason, and it should not pretend to. But that does not soften the risk, it only changes the justification. The job was never to guess the intent. The job was to refuse to wave through a sharing model at that volume without raising a loud flag.</p><p>Here is what stays with me. The skew was not hidden. It was not waiting quietly inside the data for someone to stumble on it. It was deducible from numbers we already had, in documents we had already written and approved.</p><blockquote>What was missing was not information. It was a review that put the sharing model and the volumetry side by side and read them together, before the build, not months into production.</blockquote><p>That is what I mean by designing with AI in the loop. Not an oracle that makes the call for you. A reviewer that asks the right question at the right moment, and does not let you move on until you have answered it. The architect still decides. The AI just makes sure the critical question gets asked while it still costs nothing to answer.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=645315fc12bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>