AI’s $1 trillion problem is missing data This year, Enterprises will spend $ 140bn on cloud storage alone. Objects stored in Amazon S3 have grown 250x over 13 yrs to 500 trillion+ today. Enterprises are storing vast amounts of data in the cloud. This decade, the promise is that... God LLMs will parse Enterprise data and through voodoo magic solve all problems. What has it done so far? Per McKinsey, only 37% of Enterprises can point to positive P&L impact from AI. Precisely 6% of Enterprises had a ≥ 5% EBITDA uplift from using AI. The $1 trillion Q is: Why are AI deployments failing in Enterprises?show more

The simple articulation is: Manuals, docs & decks are stored. How to use this material is missing. Systems record the final action but discard the reasoning, exceptions, corrections and outcomes surrounding it. The table below captures the types of data which Enterprises aren’t capturing today. Databricks and UiPath have a non-trivial solution to the problem: Record all work meetings, create an automated context gathering engine and use a librarian style AI agent to ask real-time Qs to team members. Databricks calls this the Genie Ontology (because the Genie fulfills your every wish). UiPath calls this Cartographer (because it creates a “map of work”) I call mine the “Bambu Master” because its job is literally to “bambu” my colleagues to gather intent, rationale & context data.

Today, Enterprise AI agents are like new employees who are given a work manual but aren’t given training, peer support or mentorship. Therefore, they operate with access to “historic data” but without institutional knowledge (both tacit & tribal) Today, “Bambu Master” is a human in the loop system where I interject to ask Qs (I’ve explained the philosophy behind this in the video). My 12-month goal is that Bambu Master becomes a self-learning system (like Databrick’s Genie) My new mental model is: “If you don’t capture how you manage the process, you can’t expect AI to do the process reliably for you.”

I could genuinely give so many reasons from my own personal experience while dealing with Enterprises. - These enterprise leaders have no idea about the breadth & depth of AI, or what can be achieved with it. - The people in power or influence either feel threatened or are completely noob and know nothing. Most of the Enterprise today is still in migration mode. - Outdated infra. Real example: an international bank that we met was keeping all the data in a DBMS software called "Fox" & I had no idea what that was; turns out it's a DBMS software from 1984. It would be a miracle for them to first migrate from there & then do something about AI. - Humans solve problems at the outcome & system level, not AI. These leaders are unable to decide what needs to be done. - We met one of the largest security providers in the country, doing $200 Mn annually. They have nothing; they are still figuring out their HRMS. - Enterprises in case of India are even worse; they have no intent of moving forward, no intent to actually grow; they are always in defense mode, protecting what they have. Indian Enterprises by 2030 will face a reckoning if they don't change themselves & how they carry & operate themselves. Because the firms that do will move 5x faster & slowly eat up others' market share.

because enterprises think in terms of "AI deployment" rather than here is a tough problem, let's solve this with the best tech possible. They have separate digital transformation teams who keep building reports on what next to automate without much thought

Excellent talk, the clearest framing I have seen of why transcripts alone are not enough. One q though, when a decision worked out by luck, how would your system keep a convincing post-hoc explanation from being learned as a real rule?

I'd save the reason when someone overrides an agent's recommendation. Otherwise the next run retrieves the same policy and repeats the same mistake.

context problem more than a data problem imo. the files exist, the reasons don't. been poking at this on the dev side, curious if the video says the same for enterprise

Complete data >= more data

The data is not missing. The problem is that enterprise data is stored in 40 different tools with 40 different formats. LLMs cannot fix a plumbing problem with more intelligence.

500 trillion objects and nobody can find the Q3 deck.

Problem is not AI but rather, portability of AI

Facilitator in MS Teams can have this functionality since it already is a part of daily standup calls and meets.

Enterprise AI is starting to look like a systems problem. Storage, retrieval, inference and data access all have to work together before the model can actually create value.

I think company have data but critical thinking and asking right question isn't executable. Think like you have lots of peanuts in hand but you don't wanna to end up with bad taste . So you unwrap the rosted peanut shell and taste each and exclude and include based on taste. 😇

So why does India still does not have a alt to Scale AI which can create real world and synthetic data for #AI? We have the scale of human and data sources , and we are democratic, so why still not have such a service which western countries can use? Is it data or cultural bias ?