Twitxy Saved

Mohammed Sumon

Verified account This account is verified. Learn more

@Mosumon8

A decade inside legacy businesses, executing the workflows that drag them down. Now showing how to use AI to stay relevant and transform the way we do work.
New York mohammed-sumon.com Joined April 2013
902 Following 194 Followers 819 Posts

Latest posts 5

Mohammed Sumon Verified account @Mosumon8
Every enterprise AI roadmap has a model strategy. Almost none have a document strategy.

That gap is where most AI transformations quietly die.
I've spent years inside diligence processes and loan trading desks. The pattern is identical everywhere: the business runs on documents - contracts, amendments, invoices, certificates, the deck from 2019 that still governs a live obligation. And those documents live in five systems, three naming conventions, and one shared drive folder called "Final_v3."

Then leadership buys an AI platform and asks it to reason over that.

The model isn't the problem. The model is fine. Ask it a question it can see the answer to and it performs like a senior analyst. The problem is that the answer lives in a scanned PDF, page 141, in a folder nobody indexed.
It's not an intelligence gap. It's a custody gap.

Which is why the unglamorous middle of the stack matters more than the model layer right now. Document extraction and normalization - the Reducto and Hebbia category - plus the boring disciplines nobody puts in a keynote: a single system of record, consistent metadata, amendments actually linked to the agreements they amend.

Here's the test I'd give any transformation plan: can you retrieve every document that governs one customer relationship, in one query, with confidence nothing is missing? If no, the AI initiative is lip service.

The re-keying just moved from analysts to prompts.
The companies pulling ahead aren't the ones with the best model access. Everyone has the same models (even I do). They're the ones whose documents are machine-readable enough that the model has something true to read.

Clean data custody is the transformation. The model is the reward.
Mohammed Sumon Verified account @Mosumon8
Meta is rewiring its business away from being reliant on advertisers. Will be huge if they're able to execute and deliver to market. Next step is to leverage their trove of human data to create hyperspecialized inference which can be sold at what will likely be a hefty premium.
Bloomberg Verified account @business Jul 1
Meta is developing plans for a cloud infrastructure business that will sell access to AI computing power and models, setting up a new vector of competition with industry leaders like Amazon Web Services, Microsoft Azure and Google Cloud bloomberg.com/news/articles/…
View quoted post on X
Mohammed Sumon Verified account @Mosumon8
Bain says the next wave of enterprise software will win by operating closest to governed truth. Almost right. The wrong word here is "truth."

Truth inside a company is contestable. Ask two business-unit leaders for the real number and you'll often get three, each sufficiently defensible and each built on a different definition. Governance does not apply when it comes to truth because it opens the possibility for disagreement. Imagine if different agents started to rely on different truths and eventually ran into each other while making an update to a dashboard or report. Who would win?

I believe what you can govern is something more precise: the fundamental unit of information. An atom. Not truth, but a specific, named fact that is replicable, transparent, and traceable. It points to one insight. It returns the identical result when you run it again. It carries a record of how it was generated. Unlike "truth," it is not an argument. A fundamental unit of information is an accurate receipt of history. Agents need receipts.

That distinction reframes who actually wins the next phase of enterprise AI.

As foundation models commoditize, the winner won't be whoever has the best agent. It'll be whoever lets you run agents under strict governance: control over what an agent sees, how it interprets the data, how it acts on it, and how it leaves a traceable map of those actions for accountability. The agent is becoming the cheap part. The control around it is the valuable part.

Start with what an agent stands on. To plug into real workflows, it needs a context layer over all the tools and data: a machine-readable model of what the business actually means. Think about it. When you hire someone new, you (ideally) give them an onboarding runway with the materials they need, their context, so they can eventually do the job well. This is part of what makes humans sample-efficient. Give us enough good material a few times and we get really good at understanding what applies and what doesn't.

For a while the industry tried to brute-force this by enlarging the context window and stuffing in more tokens. But performance degrades well before the limit. Past a couple hundred thousand tokens, the model starts losing the plot. We have now realized that more context isn't the fix (it's also economically the wrong path to follow). Better-structured context is. That's what Databricks' Genie Ontology strives for: context codified into distinct, retrievable units the business defines once, so every agent draws from the same knowledge base instead of rebuilding it.

It matters for the work itself. The best decisions don't come from the most data. They come from the most relevant information. So a good agent's job isn't to ingest everything. It's to collate what exists, structured and unstructured, decide what's relevant, synthesize it into insight, act, and iterate as feedback arrives. Get the context layer right and you raise decision quality and cut token cost at the same time, because the agent reasons over curated units instead of wading through noise.

Then there's the part no one can afford to get wrong. Doing the task isn't enough. Doing it inside strict security boundaries is the rest of the job. The nightmare is an agent running shadow operations or drifting off the rails until the damage is irreversible. Unity AI Gateway addresses this as a runtime control layer. It bounds what an agent can touch, spend, and trigger, with audit on by default. It's as if you have a whole enterprise risk team moving with the agent, seeing and controlling everything it does.

And the security question runs in two directions at once. You're securing the agent operating inside the enterprise, and securing the enterprise from the agent. It sounds adversarial, but the two work in a symbiotic way. If either side is weak, the whole thing is exploitable, because the agent is both an actor inside your perimeter and a possible way through it. Companies need to govern them together or not at all. The timing matters too: controlling at runtime, rather than midway or after the fact, is what holds down both damage and cost. It is impossible to pursue a "no-error" state. Perfection does not exist, and edge cases will always arise. You have bad actors who defraud companies, manipulate markets, and commit corporate espionage. You cannot control for everything, but you can create situations that reduce the surface area of catastrophe. Starting at runtime is the best way to begin that pursuit.

We have seen the alternative. The SaaS revolution left most companies with application sprawl: a different tool for every task, none of them talking, employees inheriting workflows that were more about updating trackers than thinking about what's next. Many are now repeating it with agents. Annual AI budgets get burned through in a quarter. Teams run siloed, inefficient agent operations that won't scale, which quietly caps the whole benefit AI was supposed to deliver.

This leads to two kinds of companies. The first replicates the SaaS era: agents bolted onto sprawl, each a local fix, the mess just faster now. The second collapses its databases and process layers into something coherent and connected, so an agent can plug in, reference the most current information, and decide well because it's standing on good data. The gap between the two compounds, the way plumbing always does.

Which brings it back to the word that started this. The winners won't be whoever owns the most data or the best model. They'll be whoever puts agents closest to information that is governed, precise, and reproducible. Atoms. Not the company's truth, which is always up for debate, but its fundamental units of information, which aren't. You argue about truth. You act on atoms. And a business built on atoms compounds, because every agent is standing on the same ground.
Databricks @databricks Jun 20
Missed the #DataAISummit 2026 keynotes? We’ve got you covered.

Full keynote replays and select breakout sessions are now available on-demand to help you catch up on the biggest moments from the week.

Start watching: databricks.com/dataaisummit?u…
View quoted post on X