What is an AI software factory? And how to tell a real one.
An AI software factory is a delivery system that turns a specification the business has confirmed into a finished, running business system. AI agents write most of the code. People hold the gates on intent, architecture, quality and compliance. What makes it a factory is that it delivers the system to the business and carries the risk of building it.
The name is now used for four different things, and in the Gulf it is easily confused with a fifth. This page separates them and gives a buyer six tests to apply to any vendor that uses the word, including us.
An AI software factory is not an AI factory
In the Gulf the word “factory” already belongs to something else. NVIDIA describes AI factories as “a new class of infrastructure built to manufacture intelligence”, converting energy into tokens. When HUMAIN and NVIDIA announced AI factories for Saudi Arabia in May 2025, they meant data centres with a projected capacity of up to 500 megawatts. Deloitte’s “AI Factory as a Service” for the Middle East is a suite of generative-AI capabilities built on NVIDIA’s platform.
That is infrastructure, and the AI tooling that runs on it. It produces capacity and models. A business that needs a tenancy system, a lending platform or an e-invoicing workflow does not get one by buying capacity, any more than owning a power station gets you a car.
| Aspect | AI factory | AI software factory |
|---|---|---|
| What it is | Computing infrastructure: data centres built around accelerated chips | A delivery system: specification, build, verification and governance run as one chain |
| What it produces | Model training and inference, measured in tokens | A finished business system that an organisation runs |
| Who buys it | Governments, cloud providers and model builders | The business that needs the system: a bank, a developer, a distributor, a ministry |
| What it is measured in | Megawatts, tokens per second, cost per token | Whether the system went live and whether the number it was built for moved |
| What you hold at the end | Capacity, or access to it | The system, its source code and its specification, on the terms the contract sets |
The two are complementary: a software factory can run its models on an AI factory in the region. They are still different purchases, made by different people.
Four things are sold under the same name
The term is old. Hitachi adopted it for its Software Works in 1969, and it has meant a structured, repeatable way of producing software ever since. More recently AI agents made the production step cheap, and the phrase is now used for several quite different offers.
| What is offered | Who uses it | Who carries the delivery risk |
|---|---|---|
| Prompt-to-application builders | Business users and founders, mostly for first versions and prototypes | The enterprise |
| Coding agents and agent platforms | The enterprise’s own developers, working in its codebases | The enterprise |
| Specification and lifecycle control planes | Product and engineering teams upstream of the code | Mostly the enterprise |
| Factories that deliver to the business | The delivering firm, working with the enterprise’s owners | Shared, and in some models carried by the deliverer from go-live |
The first three are tools. They are often excellent, and they make an enterprise’s own people more productive. They do not change who carries the risk. Much of what is called a software factory, in other words, is sold to your developers. The question for a buyer is whether the factory delivers to the business. This classification comes from Position Paper No. 8.
Six things a buyer can check for
Each is a property of how the vendor works, not of what it claims. Each comes with a question you can put in a meeting.
The business confirms something it can use before anything is built.
A prototype the users can walk through removes more requirement risk than any signed document.
Ask: Can our people click through it before you build it?
The specification is an artifact that outlives the people who wrote it.
Requirements, use cases, stories and test cases that the next person can read. “It is in the code” is not an answer.
Ask: Who owns the specification when its author leaves?
Whoever wrote the code does not mark its own work.
The tests are written by something other than the agent that wrote the code, and kept where that agent cannot change them. People then run and verify the result.
Ask: Who tests the work, and did they write it?
The work does not begin from an empty repository.
Shared foundations and domain components that already exist and have already been tested, so the project’s effort goes on what is specific to you.
Ask: What already exists on day one, and can we open it?
The running system is governed, not only the process that built it.
Decisions that touch money, entitlement or compliance stay deterministic and auditable after go-live.
Ask: What does the system do when the language model is switched off?
The enterprise holds the source and the specification, and can run the system without the vendor.
With a right to host and change it that does not expire, and commercial terms that put the vendor at risk if the system does not go live.
Ask: What exactly do we hold at the end, under which licence, and what do you lose if it never runs?
A chain, not a tool
A factory that passes the six tests runs the same seven steps, in the same order, every time. People own the intent and the outcome. The factory owns the execution, and the delivery risk that comes with it.
- 01SpecifyIntent distilled by people who know the business.
- 02ConfirmUsers click through a prototype first.
- 03BuildA governed build, on a proven core where one fits.
- 04VerifyLayered QA, independent of the builder.
- 05DeployAdoption owned jointly with your team.
- 06OwnSource and specification are delivered to you.
- 07GovernDecisions stay deterministic and auditable.
A development company, coding tools, and a factory
These are general patterns, not a description of any one firm. The last column is what the six tests look for.
| Aspect | Development company | AI coding tools | AI software factory |
|---|---|---|---|
| What you buy | People and time | A licence for your developers | A finished, running system |
| Where the build starts | Usually a blank repository | Your existing codebase | Components and cores already built and tested |
| Who tests the work | Often the team that wrote it | Your team | Tests written and held independently of the builder, then human QA |
| What a change costs | A change request, priced | Your developers’ time | Depends on the contract; ask whether changes are billed |
| Who is exposed if it is late or wrong | The buyer | The buyer | The factory, to the extent payment depends on go-live |
The longer version, risk by risk, with when each route is the right one: software factory, development company or AI coding tools.
Xamun against the same six tests
Each row says what we do and where on this site it is shown.
| Test | What Xamun does | Where to check |
|---|---|---|
| 1 · Confirm before build | The people who will use the system click through a prototype before the build starts. | Software Factory, step 02 |
| 2 · Durable specification | Use cases, user stories and test cases are held as artifacts that queue for your acceptance, and the specification is delivered with the source. | Software Factory, XamunForge orchestration and step 06 |
| 3 · Independent verification | Test-driven, with the roles separated. The tests are written by a different AI agent, on a different model from the one that writes the code, and are locked away from the code-writing agent so it cannot change them. Automated quality gates run, and human QA runs and verifies the result before your own acceptance testing. | Software Factory, step 04 |
| 4 · Starts from built | Twenty-eight systems in the OS Series, most with a live demo you can open and the rest shown by walkthrough; a new build starts from the nearest core where one fits. | OS Series |
| 5 · Governed when running | A deterministic engine holds rules, calculations and the audit trail, runs first and has the final say; the language model never does. | Two Minds |
| 6 · Yours to run | Built from a blank specification, the system is yours outright and you own the code. Built on an OS Series core, you get the full source with a perpetual right to run, brand, host and modify it, without exclusivity on the core. Usage is paid from go-live; on a licence the final payment is due at done. | How we price, Turnkey and the published licence form |
What this page does not claim: that Xamun is the only software factory, or that every project belongs in one. Under the usage model nothing is billed before the operation runs through the system; under a licence there are milestones, with the final payment due at done. For three years after done, if the agreed number slips, we fix it without asking why. The terms are on Outcome-as-a-Service and Turnkey.
When you should build your own engineering
When software is the product. A technology company, or a bank whose digital channel is its competitive position, should own its engineering.
For small internal tools with little regulatory exposure and a short life.
For exploration. Prototypes built to learn what is possible are cheap and valuable. The risk arises only when they are promoted to production without being respecified.
The line runs through systems of record, regulated workflows, and anything that must still behave correctly in three years. Those are where a factory earns its place.
What changes for a buyer in the Gulf
Two things, whether the buyer is in the UAE, Saudi Arabia, Qatar, Kuwait, Bahrain or Oman. The first is the vocabulary above: ask a vendor which kind of factory it means before the meeting, not after.
The second is that tests five and six carry more weight. A system that decides anything about money, tenancy, credit or personal data should be able to show a regulator which component had the final say, and where the data sat when it did. The rules that bind a private firm’s AI system in the UAE are set out on AI governance in the UAE, and the government programme for agentic AI on Agentic AI in Dubai.
Xamun has an office at the Dubai AI Campus in DIFC, a registered office in London and its build team in Manila. Jurisdiction packs exist for the UAE, the Philippines and the United Kingdom, and coverage varies by product. A system for another Gulf state starts with writing that state’s rules as a pack, and we have not published one yet.
Questions buyers ask about software factories
What is an AI software factory?
An AI software factory is a delivery system that turns a specification the business has confirmed into a finished, running business system, with AI agents writing most of the code and people holding the gates on intent, architecture, quality and compliance. What makes it a factory, and not a tool or a team for hire, is that it delivers the system to the business and carries the risk of building it.
Is an AI software factory the same as an AI factory?
No. An AI factory is computing infrastructure. NVIDIA describes AI factories as infrastructure built to manufacture intelligence, converting energy into tokens. The AI factories announced for Saudi Arabia in 2025 are data centres. An AI software factory produces something different: a finished business system, such as a lending platform or an escrow ledger, that an organisation runs and owns. One can run on the other, but they are bought by different people for different reasons.
How is an AI software factory different from a software development company?
A software development company usually sells effort: people and time, billed by the day or the sprint, with change requests priced separately, so the buyer carries the risk that the system is late or wrong. A software factory sells the finished system. The build runs through one repeatable chain, starts from components that already exist, and is paid for in a way that depends on the system going live.
How is it different from AI coding tools?
Coding agents make your developers faster. They are sold to your developers, and your organisation still owns the requirement, the testing, the adoption and the risk. A software factory covers the whole chain from specification to a governed running system and delivers the result to the business.
How can a buyer tell whether a vendor is really a software factory?
Apply six tests. The business confirms something it can use before anything is built. The specification is a durable artifact that outlives the people who wrote it. Verification is independent of whoever, or whatever, wrote the code. The work starts from something already built. The running system is governed, not only the process that built it. And the enterprise holds the source and the specification, and can run the system without the vendor.
Who owns the code a software factory produces?
It depends on the factory and the contract, which is why this is one of the six tests. With Xamun there are two cases. A system the Software Factory builds for you from a blank specification is yours outright: you own the code. A system built on an OS Series core comes under a source-available licence: the full source of the delivered version, a perpetual and irrevocable right to run, brand, host and modify it for your own business, and ownership of the modifications you make. The licence is non-exclusive, so Xamun keeps the core and builds on it for others. You cannot resell it, offer it to others as a service, or share the source beyond your own employees and contractors. Ask any vendor the same question and ask to see the clause.
When is a software factory the wrong choice?
When software is the product itself, as for a technology company or a bank whose digital channel is its competitive position, building your own engineering organisation is the right decision. The same is true of small internal tools with little regulatory exposure, and of exploration, where prototypes built to learn are cheap and valuable. A factory earns its place on systems of record, regulated workflows and anything that must still behave correctly in three years.
Is there an AI software factory in the UAE or the wider Gulf?
Xamun has an office at the Dubai AI Campus in DIFC, a registered office in London and its build team in Manila. The six tests on this page apply to us as they do to anyone else, and our answer to each one names the page where it can be checked. Jurisdiction packs exist for the UAE, the Philippines and the United Kingdom, and coverage varies by product.
- NVIDIA, “AI Factories: The New Infrastructure of Intelligence”, 27 May 2026.
- NVIDIA, “HUMAIN and NVIDIA Announce Strategic Partnership to Build AI Factories of the Future in Saudi Arabia”, 13 May 2025.
- Deloitte Middle East, “AI Factory as a Service” press release, 12 October 2024.
- Wikipedia, “Software factory”, for the 1969 Hitachi Software Works.
- Xamun, Position Paper No. 8, “You hired a coder. You needed an owner.”, sections 5 and 10, for the four kinds of factory, the six tests (the sixth is restated here in terms of the two ownership cases) and where in-house engineering is right.
Bring the six questions. We will answer them on a real system.
Pick the core closest to your operation and ask to see it. Then ask what you would hold at the end, what you would pay and when, and what happens if it never goes live.