From Service to Product: The Hardest Transition in the AI Era

The AI era has made it easier than ever to start a services business and harder than ever to escape it. This sentence explains a lot about the current startup landscape.

Every enterprise globally wants AI insights now, and a technically strong team can deliver these insights as a project. The work is real, and with the best intentions, the client is satisfied with new insights, and revenue for the startup comes in early. This is exactly the trap. These projects scale with headcount, but not with software. Each engagement is bespoke and needs continuous customization for the client. Two years in, the startup might have meaningful revenue and impressive logos, but nothing that compounds long-term. The startup converts to a consultancy business with a technical demo. Much of the value destruction in this AI cycle will come from consultancies with good demos being priced as product startups.

Byanat Founders - Ahmed Alghadani and Ahmed Albadi

The relevant question is not how to avoid services. It's how to escape them. I want to lay out what I think the escape route could look like, and show it in practice through our portfolio company Byanat, an Omani company we backed at Golden Gate Ventures that has made the jump in one of the hardest markets imaginable: critical infrastructure operations in the Gulf.

Why AI halts

Start with why so many enterprise AI efforts fail, because the failure mode and the escape route are the same thing seen from opposite sides.

McKinsey found that eight in ten companies cite data limitations as the roadblock to scaling agentic AI, and fewer than one in ten have scaled it to tangible value. MIT research puts the share of generative AI pilots with no measurable P&L impact at roughly 95%. The bottleneck is not the model layer. It's the data layer.

Nowhere is this clearer than in operations. Generative AI swept through customer service, marketing, and back-office work, but inside critical infrastructure: data centers, telecom networks, utilities and it has barely moved. The reason is simple: an operator cannot act on an answer it cannot verify. A general-purpose model asked about the state of a facility guesses from general knowledge. Even a model fine-tuned on the operator's data is still predicting the likeliest answer rather than retrieving the actual one, and it cannot tell you where the answer came from. In a control room, a confident, unsourced, occasionally wrong answer is worse than no answer at all.

There is a third path: govern the operational data into well-defined data products — each with an owner, a contract, and a lineage — and have AI query them at runtime. Ask the same question and the system reads the current value from the governed source, answers, and shows the data, the timestamp, and the reading behind it. Verifiable, current, sourced. That's the difference between an AI that guesses and an AI someone can act on.

Hold on to that idea, because it carries the whole thesis: the model is not the competitive edge. The foundation that makes AI trustworthy is. Models are commoditizing rapidly and the application layer is a crowded space. The durable position is underneath in the data layer, on the customer's side of the firewall, embedded deeply that it is nearly impossible to grasp.

The product inside the service

Now connect that to the service dilemma. If the data foundation is where AI value actually accrues, then the companies best positioned to build it are not the ones writing whitepapers about it. They're the ones already elbow-deep in customer data doing service work (if they can recognize what they're sitting on).

Here's the pattern to look for, whether you're an investor evaluating a company or a founder evaluating your own: in every engagement, what is the thing you rebuild? Not the deliverable the customer pays for — the substrate underneath it. If the answer is the same across customers, that substrate is the product. If there is no answer, there is no product, and no amount of AI branding will change that.

The service years, read this way, are not a detour. They are funded product research inside the environments the company will later sell into. Knowledge that doesn't exist in any market report, because it lives behind the customer's firewall. The transition fails not because companies lack this asset, but because they never stop to name it, and commercial pressure keeps them shipping bespoke work until the window closes.

Byanat: A Case Study

Byanat is the cleanest example of this pattern I've seen up close.

The company was founded in Muscat in April 2022 by two Omani engineers. Ahmed Alghadani came from mechatronics and embedded systems; Dr. Ahmed Albadi from four years of machine-health analytics research at the University of Sheffield in collaboration with Rolls-Royce. Their founding observation was that Gulf telecom operators were generating enormous volumes of machine data with no intelligent way to act on it. They chose to build in Muscat rather than Dubai or London — a decision that put them inside the operator relationships that would define everything that followed.

Their early work looked like sophisticated services: deep engagements with operators, benchmarking live network data, producing insights the operators' own tooling couldn't. A conventional read would call it a consultancy. But every engagement forced the same underlying work — extracting operational data from fragmented, multi-vendor systems, cleaning it, structuring it, governing it to the point where it could be queried and trusted. The insights were the visible output. The data foundation underneath was identical every time, at every operator, across every vendor stack.

That was the substrate, and Byanat named it. They productized the foundation, not the insights, and rebuilt the company around it: a platform that makes operational infrastructure legible to AI, running entirely on the customer's own hardware, inside their own walls. The applications on top — monitoring, predictive maintenance, capacity planning — demonstrate what the platform can do. The platform is what they sell. Services showcase; product scales.

The discipline that keeps them on the product side of the line is worth quoting because any founder can adopt it: the product ships at the next customer without code changes. Customer-specific work happens through configuration and deployment, not engineering. A feature that only works for one customer doesn't belong in the platform. It sounds obvious. Holding it under commercial pressure is what separates a product company from a consultancy with a roadmap.

The market noticed. After incubation at Omantel Innovation Labs and a 2023 seed round from 500 Global, Sanabil Investments, Omantel, and Al Jabr MENA, we led Byanat's Pre-Series A in February 2025, alongside Qatar Development Bank, Omantel Innovation Labs, Salcia Oryx Fund, Waad Investment, and PlusVC. The company now operates from Muscat, Riyadh, and Doha, with a team drawn from the likes of Rolls-Royce and Microsoft.

The Arena: Data centers and the sovereign AI significance

What makes the thesis urgent rather than academic is where it's now playing out. The same foundation that made telecom networks legible to AI applies to data centers, and data centers are the defining infrastructure buildout of this decade. AI-driven data center capital expenditure is expected to exceed $1 trillion in 2026 and reach around $1.7 trillion by 2030. Deloitte puts global commitments to sovereign AI compute above $100 billion in 2026 alone. Gartner sizes the sovereign cloud market at roughly $80 billion this year, growing 35.6%, with the Middle East and Africa growing at 89%, the fastest in the world.

The Gulf sits at the center of that curve, building compute capacity at hyperscale pace with regional team sizes and a hard constraint most global vendors can't meet: the data cannot leave. Operators of critical infrastructure need AI that runs on their own hardware, under their own governance. Sovereignty here isn't a compliance checkbox; it's the precondition for the purchase. Hyperscalers are responding with in-region cloud offerings, but a regional cloud with local directors is not the same as the operator owning the data on its own iron — and the buyers know the difference. Companies that are sovereign by architecture, not by branding, are selling into a structurally underserved market. Byanat, giving data center operators detailed operational insight in power, cooling, capacity, service-level exposure across multi-vendor estates, is built for exactly that buyer.

For founders doing service work inside these industries, the message is that the constraint is the opportunity. For investors, it's that the least crowded, most defensible layer of the AI stack is being built by teams that look, at first glance, like the thing you're trained to avoid.

Our Takeaways

For founders: the service dilemma is real, but so is a way out, and it runs through your own delivery work. Name the substrate you rebuild every engagement. Invest in it as an asset. Adopt a rule that forces generalization, and hold it when a customer waves money at an exception. Don't aim for zero services. Top-tier AI companies all win enterprise accounts with forward-deployed engineering, a point a16z made well in "Trading Margin for Moat." The test is never the presence of services. It's whether they feed the compounding asset or substitute for it.

For investors, whether you're writing checks into companies or into funds: the AI-era diligence question is not "is this a services business?" but "does the service work compound into something owned?" Revenue quality, not revenue source. The companies that pass that test in unglamorous industries: infrastructure, utilities, logistics, are where I'd argue the durable value in this cycle sits.

For everyone building ecosystems in emerging markets: Byanat is also proof of what happens when pieces line up. A company founded in Muscat, incubated inside a national telecom operator, backed first by regional institutions and then by international capital, now carrying Omani engineering into the region's largest markets. The Gulf is full of technically strong teams doing sophisticated service work inside critical industries, sitting on exactly the kind of substrate Byanat productized. There will be more of these stories. The pattern is now visible.

I'll be writing more about the region's AI developments in the coming months. If you're building or investing in this layer of the stack, I'd be glad to compare notes.

---

Sources: McKinsey, "Building the foundations for agentic AI at scale" (May 2026); MIT NANDA GenAI pilot research; Deloitte, "State of AI in the Enterprise 2026"; Gartner sovereign cloud forecasts (2026); Dell'Oro Group data center capex forecasts; a16z, "Trading Margin for Moat" (January 2026); Byanat funding announcements (2023, 2025).