Mohammed Sumon (@Mosumon8) header
Mohammed Sumon (@Mosumon8) profile picture

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
Joined April 2013
Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)Photo by Mohammed Sumon (@Mosumon8)
Some thoughts on how artificial intelligence and agents might transform sales and business development within enterprises... Musings #6: Sales, AI and the Human Experience mohammed-sumon.com/musings.h…
1
275
"Let me talk to my team" The team: - Me - Everyone on my X feed - Claude code - The 10 agents I got from someone else's repo
“Let me talk to my team” The team: - Me - Claude code - 4 coffees - Macbook Pro
2
145
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.
1
1
65
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.
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/…
45
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.
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?…
5
235
Okay but I hope we remember what happened last time when we had Terra and Luna orbiting the system...
Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work. openai.com/index/previewing-…
27
The government is effectively putting a lid on the growth trajectory of these companies. We are going to see a continuation of token share loss to open source models as those continue to get better and become more accessible.
🚨 BREAKING: U.S. government will decide who gets access to GPT-5.6 OpenAI will release GPT-5.6 only in a limited preview to a small group of partners. Sam Altman told staff the government would be "approving access customer by customer." Commerce Sec Lutnick personally called Altman warning: don't launch without approvals from other agencies. A de facto licensing regime. ITS HAPPENING
19
The combination of this + share of token used for US models moving to open-sourced Chinese models appears to be the canary in the coal mine. As compute costs continue to rise, foundational model companies will need to raise more capital to fund their business. However, only one has a business with enough free cash flow to steady the rough waters. What happens when OpenAI or Anthropic have a down round?
BREAKING: OpenAI is now "leaning toward" pushing its IPO until 2027, per NYT. Details include: 1. "Choppy" markets in recent weeks have led OpenAI to reconsider the timeline of the IPO 2. The company is worried it may not find much enthusiasm from retail investors 3. Advisors are recommending OpenAI either wait until 2027 to IPO at $1 trillion or lower the valuation for a quicker IPO Recent volatility in tech stocks has raised concerns around OpenAI's IPO.
43
Having run commercial diligence on software targets, I'd be careful reading too much into a vibecoded replica. A replica answers: how hard is this software to rebuild? This is useful but for a software business, the code is rarely the defining moat. Most SaaS products are a clean interface on top of a database wrapper. The UI and the workflows are easy and AI can reproduce them in an afternoon now. What it can't reproduce is everything that took years to build around the code: the integrations wired into a customer's daily workflow, the people trained on the system, the enterprise data structured to be ingestible and portable, and the thousands of edge cases an incumbent has already solved that a prototype has never seen. That last one is the most underrated advantage legacy software has. A mock-up handles a sandboxed environment which never breaks. It has no idea about the exception that breaks for one customer on the third Tuesday of the quarter. And no prototype touches what actually keeps a contract renewing: the switching costs, the financial lock-in, the trust of a vendor that's been in the stack for a decade. Anyone can vibecode the prototype. I have shamelessly done this. I have yet to see Claude vibecode relationship capital. One buyer in the piece reportedly walked from a deal partly on the strength of a replica. If a mock-up is enough to move the decision, I'd ask what the diligence was actually measuring. If anything, the panic reads backwards. With strong SaaS names down by a third this year, this may be one of the better windows in a decade to buy the ones with real, durable moats. The replica just isn't where you'll find the moat.
FT: Bain is testing takeover targets by using vibecoding to rebuild rough AI-made copies of their software. These replicas are not perfect clones, but they can expose whether a product’s interface, analytics, automation, or workflow logic is easier to reproduce than sellers claim. The replicas help potential buyers assess how difficult it would be to recreate a software product and whether the company has a defensible competitive advantage. Say a company is selling an analytics tool that takes sales data, cleans it, finds patterns, creates dashboards, and recommends what a sales manager should do next. Bain may ask AI coding tools to quickly build a basic version that uploads sales data, generates similar dashboards, adds simple recommendations, and feels somewhat like the target product. Bain says its teams have built hundreds of prototypes, and the tool has moved from specialist engineers to regular consultants doing AI diligence. For investors, the Bain method is built to reveal where a software company’s value actually sits, from its code and workflows to customer relationships, data, distribution, or other business layers. Using AI-made prototypes also lets buyers see how a product could develop in the years ahead as AI continues changing enterprise tech. This strategy comes at a time when generative AI is making software development cheaper and challenging the strength of SAAS business models. --- ft .com/content/e5bac4d1-b1f8-43a4-bd54-b182d5357af0?syn-25a6b1a6=1
79
Seems like the most durable companies within AI x Financial Services will be the ones selling governance, security, and compliance capabilities and tools. Multiple companies within this space raising big Series-X lately.
1
17
1. Most Business Development pipelines are stale by the time leadership asks about them. It's not a discipline problem.
1
38
5. The layer leadership actually uses: an agent in Microsoft Copilot Studio that answers any pipeline question in plain English — what's slipping, which mandates have gone quiet, what's stuck in diligence.
1
14
6. In Copilot Studio you point it at the pipeline as a knowledge source, wire in connectors for live deal data, write the instructions, and publish to Teams. Happy to compare notes on how I wired it up.
15
I just want to be successful thats all..
627
12,184
870
62,346
Looking to connect with those building with AI within private credit / private markets. Exploring conversations around career paths and how AI is being utilized.
42
How AI Agents Will Change Businesses 1/5: Many large corporations do not, and will not, grasp how artificial intelligence and agents (collectively, agents from here on) will transform their businesses. For these corporations, that means landing on the wrong side of it. Large organizations carry too much legacy knowledge, bias, and friction, and all of it gets in the way of what would otherwise be transformational change. I believe agents will deliver the most change across three areas: institutional context, speed-to-insight, and scope of action.
1
35
4/5: Scope of action is what expands the business. The first two legs extend current activity within your existing surface area. Scope-of-action agents widen that surface area. By understanding what has worked for you across different environments and under different exogenous variables, you can build agents that pursue new business: entering a new market, developing a new product, or moving up or down the customer value chain. These agents function as mini-CEOs. They act on your behalf, autonomously or semi-autonomously, to scan the commercial landscape and identify which avenues are best to enter on your own or to pursue through partnership.
1
14
5/5: Anyone can theorize about what agents might look like. I believe the real disruption will happen in the spaces between the workflows we can already see. There is tacit knowledge transfer and execution that lives inside steps no playbook ever captures. That is what we should be building AI businesses around. Companies that only replicate existing workflows, performed the same way they always have been, will lose to the ones that use technology to redefine those workflows entirely. My three-legged stool of agentic development offers a framework for deciding where to spend an AI budget and where to point your people: learning, implementing, and refining agentic workflows.
1
18