{"id":14088,"date":"2026-07-01T01:30:48","date_gmt":"2026-07-01T08:30:48","guid":{"rendered":"https:\/\/www.solix.com\/blog\/?p=14088"},"modified":"2026-07-01T02:36:38","modified_gmt":"2026-07-01T09:36:38","slug":"a-thousand-tables-deep","status":"publish","type":"post","link":"https:\/\/www.solix.com\/blog\/a-thousand-tables-deep\/","title":{"rendered":"A Thousand Tables Deep","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"<blockquote class=\"wp-block-quote blue\">\n<p>Where Enterprise AI Quietly Gives Up, and What Finally Gets It Past the Wall<\/p>\n<\/blockquote>\n<p>There is a moment in almost every enterprise AI program that nobody puts on a slide.<\/p>\n<p>The demo works. Someone asks a question in plain English, the AI writes a flawless query, the right number comes back, and the room nods. Budget gets approved. Then the same system gets pointed at the real thing, the actual production estate with its thousands of tables and decades of accumulated business logic, and it quietly comes apart. Not with an error. With confident, well formatted, wrong answers, and a steadily growing list of questions it simply cannot handle.<\/p>\n<p>The pattern is now well documented enough that it has a shape. AI trained and tested on public data scales beautifully. AI pointed at real enterprise data often does not. As one practitioner put it after this year&#8217;s data conferences, the gap is not about how good the models are. It is the data. Public data is organized, labeled, and easy to understand. Enterprise data is scattered across systems, defined differently by every team, and missing the context that explains what it actually means.<\/p>\n<p>I have spent close to thirty years selling into enterprise IT, and I want to walk through why this keeps happening, why it is getting worse rather than better, and what finally clears the wall.<\/p>\n<h2>The wall is made of your own data<\/h2>\n<p>When a model is evaluated on a clean academic benchmark, it looks brilliant. On the long standing industry standard, the best systems land around 85 to 86 percent accuracy. Then researchers rebuilt the benchmark out of real enterprise databases, the kind with hundreds of columns, brutal joins, and undocumented legacy tables, and the same caliber of frontier models fell to roughly 17 to 21 percent. Same models. A collapse from 91 percent to the high teens the moment the data looked like a real company.<\/p>\n<p>That is not a rounding error. It is a cliff. And it is almost entirely a function of scale and mess.<\/p>\n<p>Why does it happen? Three reasons every enterprise data leader will recognize immediately:<\/p>\n<ul class=\"cbpoints\">\n<li><b>The schemas are enormous.<\/b> Benchmark databases have three to ten tables. Real production estates have thousands. One widely cited enterprise test ran against 1,251 tables and more than 750 columns, scale that overwhelms even the largest context windows on the market. You cannot paste the whole schema into a prompt and hope. It does not fit, and the moment you truncate it to make it fit, you throw away the exact context the AI needed.<\/li>\n<li><b>The naming is hostile.<\/b> Real columns are not labeled customer_name. They are labeled consumer_1 and consumer_2, with the real values buried inside JSON, and a column called MARG_FORM sitting next to one called acctScope. The relevant table is almost never the obvious one. Finding it is a search problem the model cannot solve from the question alone.<\/li>\n<li><b>The meaning lives outside the database.<\/b> Active user excludes employees. Net revenue handles refunds a particular way. Those rules live in product docs, runbooks, and ten year old threads, never in the schema itself. Without them, the query comes back syntactically perfect and semantically wrong.<\/li>\n<\/ul>\n<p>Put those three together and you get the single most dangerous failure mode in enterprise AI: an answer that looks right, reads clean, and is quietly false. No error message. No red flag. Just a number that moves in the wrong direction, and an executive who signs off on it.<\/p>\n<h2>This is why the pilots are stalling<\/h2>\n<p>This is not a fringe concern. It is what the entire market is hitting at once.<\/p>\n<p>IBM described the pattern plainly this spring: a global bank builds an AI agent that works beautifully in isolation, then watches it fail to scale because it still depends on a small team curating data and validating every output by hand. The model performs well. The system does not. By IBM&#8217;s read of the analyst data, at least half of generative AI projects are abandoned after proof of concept, much of it traced to data that was never ready.<\/p>\n<p>The macro numbers tell the same story from every angle. BCG found 60 percent of enterprises generating no material value from AI, with only 5 percent creating substantial value at scale. McKinsey found 88 percent of organizations using AI in at least one function, but only 39 percent seeing any earnings impact at all. MIT&#8217;s widely cited work landed on roughly 5 percent of pilots achieving real revenue acceleration. Gartner now projects that 60 percent of AI projects unsupported by AI ready data will be abandoned through this year.<\/p>\n<p>Read those together and a single sentence emerges. The models are not the bottleneck. The data, at enterprise scale, is. And almost every tool on the market was built to be impressive on a small, clean schema, not to survive a real one.<\/p>\n<p>That is the mood right now. Quiet frustration. A great deal of money spent, a great many demos that dazzled, and a growing pile of pilots that could not make the jump from the slide to the system of record.<\/p>\n<h2>What it actually takes to clear the wall<\/h2>\n<p>Here is the part that matters. The wall is not permanent. But getting past it requires a fundamentally different approach than pointing a general purpose model at a giant database and praying the context window is big enough.<\/p>\n<p>It requires two things, working together.<\/p>\n<p>The first is the ability to genuinely understand the data, all of it, at full enterprise scale. Not a sample. Not the thirty tables that happen to fit in a prompt. The entire estate, across live operational systems and decades of archived history, with every relationship between system, table, and field mapped out and, critically, tied to what those things actually mean to the business. When the AI knows that consumer_1 is the billing entity and consumer_2 is the ship to address, and it knows how your organization actually defines net revenue, the guessing stops.<\/p>\n<p>The second is making that power usable by an ordinary human being. A question in plain English. An answer in seconds. No SQL, no ticket to the data team, no three week wait for someone to write the query.<\/p>\n<p>This is exactly what we built at Solix, and it is why I wanted to write this piece.<\/p>\n<h2>DataSense and DataAsk<\/h2>\n<p>Two products, one job: turn your largest, messiest, most valuable data into something your whole company can simply ask questions of, and trust the answers.<\/p>\n<p>DataSense is the engine. It profiles your entire data landscape automatically, across live applications and archived data, and builds a living, business aware understanding of how everything connects and what it means in your terms. This is the part that handles the scale the rest of the market trips over. Where a general model is handed a phone book and told to find one person, DataSense gives the AI a guided, business aware map of the whole estate. It is the reason accuracy holds at a thousand tables instead of collapsing at thirty.<\/p>\n<p>DataAsk is the front door. It is the plain English layer that sits on top. Anyone in the business, finance, operations, sales, compliance, can ask a question the way they would ask a colleague, and get a precise answer back in seconds, drawn straight from the real systems of record.<\/p>\n<p>What the combination delivers, in the terms that matter to a buyer:<\/p>\n<ul class=\"cbpoints\">\n<li><b>Scale that does not break.<\/b> Hundreds or thousands of tables across live and archived systems, handled as a matter of course, not as the edge case that takes the whole thing down.<\/li>\n<li><b>Accuracy where the category collapses.<\/b> Near perfect results in exactly the conditions, large schemas, hostile naming, real business logic, where general purpose AI falls to single digits.<\/li>\n<li><b>It asks instead of guessing.<\/b> When a question is ambiguous, the system asks a clarifying question rather than confidently inventing an answer. That one behavior is the difference between AI you can put in front of an executive and AI you have to double check.<\/li>\n<li><b>Hours, not months.<\/b> Setup is measured in hours, not the multi month data engineering project most of the market quietly requires before anything works.<\/li>\n<li><b>Dramatically lower cost per question.<\/b> Roughly an 87 percent reduction in per query cost versus brute force approaches, because the system is precise about what it actually needs to touch.<\/li>\n<li><b>Governance built in.<\/b> Federated control across both live and archived data, so the answers are not just fast and accurate, they are compliant.<\/li>\n<\/ul>\n<p>I am deliberately not going to detail how DataSense does what it does. The mechanism is the moat, and it is ours. What I will tell you is what it produces: the one thing the rest of the market has not been able to deliver, which is near perfect accuracy against the full, real, sprawling data estate of a large enterprise, asked and answered in plain English.<\/p>\n<h2>What I would tell a CIO<\/h2>\n<p>If you are sitting on a stalled pilot right now, I would ask one question before you spend another dollar tuning the model. Did it fail on a small, clean test set, or did it fail the moment it touched the real thing?<\/p>\n<p>If it was the real thing, the model was never your problem. Scale was. And no amount of prompt engineering, no larger context window, no next model release fixes a system that was never designed to understand your data at the size it actually exists.<\/p>\n<p>The companies pulling ahead are not the ones with the cleverest model. They are the ones who solved the data understanding problem first, then let everyone ask questions of it. That is the whole game this year.<\/p>\n<p>Your data is almost certainly rich enough to be AI ready in value. The only real question is whether anything you have put in front of it can actually handle how big, and how real, it is.<\/p>\n<h2>References<\/h2>\n<ul class=\"cbpoints\">\n<li>ProArch, Enterprise AI Failing? Fix Data First, March 2026. <a href=\"https:\/\/www.proarch.com\/blog\/enterprise-ai-data-challenges-fabric-2026\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/www.proarch.com\/blog\/enterprise-ai-data-challenges-fabric-2026<\/a><\/li>\n<li>BlazeSQL, Natural Language to SQL: The Complete 2026 Guide (Spider 2.0 collapse to 10 to 20 percent), February 2026. <a href=\"https:\/\/www.blazesql.com\/blog\/natural-language-to-sql\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/www.blazesql.com\/blog\/natural-language-to-sql<\/a><\/li>\n<li>Spider 2.0 evaluation, frontier model accuracy of 17 to 21 percent on large enterprise schemas (arXiv). <a href=\"https:\/\/arxiv.org\/pdf\/2601.11687\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/arxiv.org\/pdf\/2601.11687<\/a><\/li>\n<li>Datost, How Accurate Is Text-to-SQL, Really? Spider, BIRD, and the Enterprise Cliff, 2026. <a href=\"https:\/\/datost.com\/blog\/text-to-sql-accuracy-benchmarks\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/datost.com\/blog\/text-to-sql-accuracy-benchmarks<\/a><\/li>\n<li>Promethium, Enterprise Text-to-SQL Accuracy Benchmarks (1,251 tables, 750 plus columns), December 2025. <a href=\"https:\/\/promethium.ai\/guides\/enterprise-text-to-sql-accuracy-benchmarks-2\/\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/promethium.ai\/guides\/enterprise-text-to-sql-accuracy-benchmarks-2\/<\/a><\/li>\n<li>IBM, Why most enterprise AI projects stall before they scale, April 2026. <a href=\"https:\/\/www.ibm.com\/think\/insights\/why-most-enterprise-ai-projects-stall-before-scale\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/www.ibm.com\/think\/insights\/why-most-enterprise-ai-projects-stall-before-scale<\/a><\/li>\n<li>BCG, The Widening AI Value Gap, September 2025. <a href=\"https:\/\/talyx.ai\/insights\/enterprise-ai-implementation-failure\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/talyx.ai\/insights\/enterprise-ai-implementation-failure<\/a><\/li>\n<li>McKinsey Global AI Survey, November 2025. <a href=\"https:\/\/talyx.ai\/insights\/enterprise-ai-implementation-failure\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/talyx.ai\/insights\/enterprise-ai-implementation-failure<\/a><\/li>\n<li>MIT NANDA Initiative, The GenAI Divide, 2025. <a href=\"https:\/\/talyx.ai\/insights\/enterprise-ai-implementation-failure\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/talyx.ai\/insights\/enterprise-ai-implementation-failure<\/a><\/li>\n<li>Iris.ai, Why 95 Percent of Enterprise AI Projects Fail, And How to Fix It, March 2026. <a href=\"https:\/\/iris.ai\/blog\/why-enterprise-ai-projects-fail-roi\" rel=\"noopener noreferrer nofollow\" target=\"_blank\">https:\/\/iris.ai\/blog\/why-enterprise-ai-projects-fail-roi<\/a><\/li>\n<\/ul>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"<p>Where Enterprise AI Quietly Gives Up, and What Finally Gets It Past the Wall There is a moment in almost every enterprise AI program that nobody puts on a slide. The demo works. Someone asks a question in plain English, the AI writes a flawless query, the right number comes back, and the room nods. [&hellip;]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":123478,"featured_media":14094,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[139],"tags":[],"coauthors":[334],"class_list":["post-14088","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-enterprise-ai"],"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/posts\/14088","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/users\/123478"}],"replies":[{"embeddable":true,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/comments?post=14088"}],"version-history":[{"count":4,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/posts\/14088\/revisions"}],"predecessor-version":[{"id":14092,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/posts\/14088\/revisions\/14092"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/media\/14094"}],"wp:attachment":[{"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/media?parent=14088"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/categories?post=14088"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/tags?post=14088"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/www.solix.com\/blog\/wp-json\/wp\/v2\/coauthors?post=14088"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}