Credit: BRKZ A transporter pulls up to a construction site outside Riyadh, unloads a delivery of bulk cement, and sends a photo of the delivery note through WhatsApp. For most of BRKZ’s history, that photo was the start of a chain of human work. Someone had to read the document, find the matching order, verify the quantity, update the system and close the delivery. Today, an agent we call Nusa does most of that.
It reads the image, extracts the details, matches them against our systems, verifies the quantities and closes the delivery. The team handles uncertain matches and exceptions. Nusa now closes the majority of deliveries in our reconciliation workflow, and its share grows every month. Bulk cement isn’t a glamorous AI use case.
That’s exactly why I like it. I recently came back from a trip to the US where AI dominated most of my conversations with founders, investors and operators. The conversation there is moving past copilots and headcount reduction toward a more interesting question. Can AI fundamentally change the operating model of a company?
I think it can. And I believe some of the biggest opportunities are in the industries where technology has historically struggled most: physical goods, fragmented supply chains, credit and logistics. We built the foundations first When we started BRKZ, our first problem was more basic: the industry needed systems. Building-material procurement was deeply manual and fragmented.
Requests for quotation (RFQs), pricing, supplier relationships, orders, deliveries, credit and collections lived across phone calls, WhatsApp messages, spreadsheets, PDFs and people’s heads. So we spent our first three years doing something less fashionable but essential. We brought our core workflows onto shared systems: how demand enters, how suppliers are evaluated, how quotations are created, how orders are fulfilled, how deliveries are reconciled, how customers pay and how every transaction is recorded. That gave us something valuable: every workflow started leaving a digital trail.
Every RFQ, quotation, supplier interaction, delivery and payment became data. Over time those trails accumulated into tens of millions of structured data points across products, suppliers, transactions and payment behavior. We don’t have to ask every contractor and transporter to change how they communicate. An RFQ can still arrive on WhatsApp and a delivery note as a photo.
Agents turn those inputs into structured records the business can act on. You can’t add meaningful intelligence to a business you haven’t first made observable. Systemize → Capture → Understand → Automate → Agentize I’ve come to think the AI journey for traditional industries follows a progression: systemize, capture, understand, automate, agentize. By agentize, I mean giving agents responsibility to act within clear limits.
The first two steps are essential. Without data, AI has no proprietary context. And without context you’re putting a general-purpose model on top of the same information available to everyone else. From prediction to execution Once we had enough transaction history, we started asking different questions.
Could the system learn how we price? Could it work out which suppliers were relevant for a given RFQ? Pricing building materials is surprisingly hard. The same product can carry different prices depending on quantity, location, delivery requirements, timing and market conditions.
Traditionally that knowledge sits in the heads of experienced procurement people. So we built a pricing engine on our historical RFQs and transactions. Mizan, our AI procurement agent, is now live on our first product categories. It generates price recommendations almost instantly for procurement officers to review and adjust before submission.
The data we’d spent years collecting stopped describing what had happened and started helping us decide what should happen next. Prediction is useful. Execution is more interesting. Nusa taught us that AI becomes far more powerful when it can act inside a system rather than simply answer questions about it.
Which parts of the transaction can an agent own? Instead of asking “where can we add AI?”, we started asking “which parts of the transaction can an agent own?” A transaction breaks naturally into domains (sales, procurement, credit, operations and finance), and each has inputs, decisions, actions and exceptions. In procurement, an agent can interpret an RFQ, identify products, predict prices, select suppliers and quote competitively. In credit, it can support underwriting, limits, exposure monitoring and collections, with human oversight where the decisions matter most.
In operations, agents coordinate orders, validate deliveries and reconcile the physical world with our systems. The goal is an agentic layer across the transaction, with clear responsibility for the work each agent carries out. The hidden cost is coordination Companies don’t only have labor costs. They have coordination costs.
A single customer request might touch sales, procurement, operations, logistics, finance and credit before it becomes a completed transaction. Each handoff creates latency. Someone sends a message. Someone asks for approval.
Someone copies information from one place into another. As organizations grow, coordination itself becomes work. AI can remove enormous amounts of this invisible work, and I suspect that will ultimately matter more than automating individual tasks. More output per person My strongest takeaway from the US was that the interesting question isn’t “how many people can AI replace?” It’s “how much more can every person accomplish?” I want AI to turn a great salesperson into someone with capabilities that would previously have required an entire support team.
Imagine every salesperson with a digital twin. It knows which customers are likely to reorder, which have quietly reduced their purchasing, which quotations didn’t convert and which products a customer should be buying but isn’t. Who should I call? Why now?
What should I sell, and at what price? Which relationship is deteriorating before I’ve noticed? The salesperson still owns the relationship. An agent can handle preparation, analysis and routine coordination, and eventually take on follow-ups and reorders within agreed limits.
The same logic applies to finance, procurement and operations. As agents take on routine work, people focus on where their judgment and relationships create the most value. The physical world has to be made visible In physical industries, much of what matters still happens outside the data a company captures. A truck arriving at a site is data.
A pallet being unloaded is data. A material failing a quality test, inventory sitting in a warehouse, a supplier repeatedly arriving late: all data. But if nobody captures those events, they don’t exist digitally. AI can only reason about what it can see.
This leads to a counterintuitive idea: operational complexity can become a moat. But if one operations employee can oversee dramatically more transactions because agents handle routine coordination, and one procurement person can manage dramatically more spend, then the economics change. The harder a workflow is to execute manually, the more valuable it becomes once you can teach machines to execute meaningful parts of it reliably. Don’t rebuild everything just because you can Not every problem needs the smartest model in the world.
The enterprise AI stack won’t have one model. It will have many, and the architecture will decide which one, local, specialized or frontier, fits each job. And AI makes building software dramatically easier. But if someone has already solved a generic problem extremely well, buying still makes more sense than building.
Our engineering capacity goes where BRKZ has something proprietary: our workflows, transaction data, pricing knowledge, supplier network and payment histor
Leave a Reply