DXwand Built the AI Nobody Wanted to Build - and Made Money Doing It

There is a moment in every enterprise AI demo where the polished part ends and the real part begins. Someone in the room types a question the way they would actually type it. In Cairo, that question comes out in Egyptian Arabic, spelled phonetically, half in Latin characters, with a word borrowed from English in the middle. The model, trained on a mountain of Modern Standard Arabic, blinks and returns something confident and wrong.

Ahmed Mahmoud watched that moment happen enough times at Microsoft, where he spent years leading enterprise AI work across the Gulf, that it stopped being an anecdote and started being a business. In 2018 he and Mahmoud Gomaa left to build DXwand, on a premise that sounded narrow at the time and looks obvious now: Arabic is not one language for machine-learning purposes. Egyptian, Gulf, Levantine and Maghrebi dialects diverge enough that a system trained on the formal written register fails the second it meets a real customer. Global vendors treated this as a localisation footnote. DXwand treated it as the product.

We've focused on making AI actually work for businesses. It's not about just tech for tech's sake; it's about solving real problems.

Ahmed Mahmoud, co-founder and CEO

The pivot that made the company

DXwand did not start where it ended up. Its first product helped small merchants sell on Facebook and WhatsApp - a customer-facing chatbot for shops that had no call centre and no CRM. It worked. It also had a ceiling. Small businesses in emerging markets are price-sensitive by definition, and a chatbot is a feature, not a system of record.

In 2021 the founders moved upmarket, to banks, telecoms and government bodies. The technology did not change dramatically. The problem did. Enterprises did not need a friendlier front end. They needed access to their own institutional memory: two decades of policy documents, contracts nobody had opened since 2011, call transcripts sitting unlabelled in a data lake. The industry term is knowledge mining. The practical version is that a compliance officer can ask a question in plain language and get an answer sourced from the organisation's own paperwork.

That shift is what investors bought into. Tamer Azer of Shorooq Partners, which co-led the company's Series A, described the appeal as technology "that mines institutional knowledge" to make companies faster and governments more efficient. Karim Hussein of Algebra Ventures, the round's other lead, pointed to "an impressive roster of corporate and government clients" as evidence the tools were solving real problems at scale.

What ORXTRA actually is

The current platform is called ORXTRA, and it is best understood as three things stacked on each other.

At the bottom sits the knowledge layer: a retrieval-augmented generation engine that ingests documents, structured data and roughly eighty other source types, and - crucially - does not require manual data labelling. That last detail is unglamorous and load-bearing. Labelling is where most enterprise AI pilots quietly die. Someone has to tag thousands of examples so the system knows what a refund request looks like versus a complaint. It takes months, costs a fortune, and by the time it is finished the executive sponsor has moved teams.

In the middle sits Agent Builder, a visual no-code editor with a scripting escape hatch for engineers who want one. It orchestrates more than twenty language models - open-source, commercial and DXwand's own fine-tuned versions - selecting between them for cost and accuracy rather than betting the product on any single vendor.

On top sit the agents themselves. Some are pre-built: customer service, which DXwand says can absorb up to 90 percent of routine workload; procurement, which automates RFP creation, bidder management and proposal evaluation with time savings the company puts at up to 85 percent; and HR, covering onboarding, self-service and resume filtering. Others are custom - document intelligence for policy and compliance review, and a generative analytics module that produces dashboards and, notably, charts explaining how it reached its answer.

Figure 1 - Where the agents live

WhatsAppMicrosoft TeamsFacebook Messenger Web chatSMSCall centre / voiceEnterprise apps CRMERPHR platformsProcurement tools

The cost argument, which is the real argument

Ask most AI vendors what makes them different and you will hear about accuracy. Ask DXwand and you end up talking about GPUs.

The reason is arithmetic. A bank can afford a brilliant AI agent for a thousand customers. It cannot necessarily afford one for ten million, and a national ministry serving an entire population certainly cannot. The economics break before the technology does. So DXwand built Prism, a small language model tuned for Arabic and English that the company says uses up to 85 percent less GPU than comparable alternatives, and can be fine-tuned for harder reasoning tasks. Speaking to Arab News, Mahmoud described the platform as delivering over 90 percent cost optimisation while raising accuracy by more than 30 percent, using smaller models paired with retrieval tooling.

Those figures are the company's own and should be read as vendor claims. But the strategic logic behind them is sound and worth stealing: in emerging markets, the binding constraint on AI deployment is almost never ambition. It is the invoice.

Figure 2 - Funding, round by round

2021–22 Seed

$1.3M

2022 Pre-A

$1.0M

2023 Series A

$4.0M

Who is actually buying

DXwand's customer list reads like a directory of institutions that cannot afford to be embarrassed. Financial services: ADIB, EFG, the fintech Valu. Healthcare: Bupa. Government: Egypt's Ministry of Communications and Information Technology. Education: Kuwait's Ministry of Education, where DXwand built a multilingual student support assistant named Hamad. Add legal services provider NIABA and retail group Al Adal, and the pattern is clear - regulated buyers with large volumes and low tolerance for a model that invents a policy.

That is a harder market to enter and a much better one to occupy. Regulated buyers move slowly, demand on-premise or offline deployment, and then stay for years. Nobody swaps out their central bank's AI vendor because a competitor shipped a nicer interface. DXwand leaned into this deliberately, building offline and air-gapped deployment as a first-class option rather than a reluctant concession.

The commercial results follow the pattern. Contracts run as annual subscriptions sized to the problem - measured by conversation volume or user count - reported publicly in the range of roughly $50,000 to $400,000 a year. The company told TechCrunch it reached about $5 million in annual recurring revenue in 2023, roughly double the prior year, and has said it grew by a multiple of two every year since 2020 while maintaining profitable unit economics.

If I go to a client now and say generative AI, it has a meaning in their mind that I don't need to explain. A year ago, it may take us a month to explain why this is important.

Ahmed Mahmoud, on the cost of being early

Where it sits in a crowded market

DXwand is not trying to beat OpenAI, and says so. It supports more than twenty models and competes on everything that happens after the model: the connectors, the Arabic handling, the compliance posture, the price per conversation, and the speed of getting live. The company claims deployments measured in hours or days rather than months, and cites 8X faster time to market against traditional approaches.

Figure 3 - The alternatives a MENA enterprise is weighing

Option	Strength	Where it strains
Global platforms (Kore.ai, Yellow.ai, Ada)	Mature tooling, deep integrations, large partner networks	Dialectal Arabic, local data-residency rules, price at national scale
Hyperscaler tools (Copilot Studio, Dialogflow)	Already in the procurement contract, familiar to IT	Generic out of the box; heavy internal build required
Build in-house	Full control, no vendor lock-in	Scarce ML talent, long timelines, ongoing model ops burden
DXwand / ORXTRA	Arabic dialects, no manual labelling, offline option, cost engineering	Small team, regional brand recognition, limited footprint outside MENA

Geography is the other half of the strategy. The Series A was earmarked for expansion into Saudi Arabia and Africa. DXwand secured a MISA licence and moved to open a Riyadh office, working with the Gulf expansion platform AstroLabs, and has said it aims to deliver two nationwide-impact initiatives a year. Given the scale of Saudi digital-transformation spending, that is where the contracts are.

The quiet advantage

The most interesting number in DXwand's story is not the $4 million or the $5 million. It is the word "profitable."

In a sector where the default operating model is to raise enormous sums and outrun the burn, a company with positive unit economics gets to make different decisions. It can turn down a badly-fitting enterprise deal. It can let a nine-month government procurement run its course without a board meeting about runway. It can negotiate its next round from a position that is not desperation. Capital efficiency is usually described as a virtue. Here it functions more like a weapon.

Mahmoud frames the company's purpose in terms that are notably free of the usual disruption vocabulary. AI, he has said, exists "to empower individuals by automating mundane and repetitive tasks, allowing them to focus on more creative and valuable work" - and the goal is finding ways for humans and AI to collaborate rather than replace. The numbers underneath that framing are specific: 90 percent of a customer service queue, 85 percent of the hours spent assembling an RFP, 70 percent of a trading platform's advisor workload. None of that is the job. All of it is the part of the job people complain about.

Whether DXwand becomes the enterprise AI layer for the Arabic-speaking world or gets outmanoeuvred by a hyperscaler that finally takes dialects seriously is an open question. What is already settled is the lesson. The founders found a problem the industry had labelled too small to bother with, checked whether it was actually small or merely inconvenient, and built the specific, unfashionable thing. Seven years on, that thing answers questions for banks and ministries in a language the market said was an edge case.