Ashish Koshy talks about Artificial Intelligence as if it were a utility – not a spectacle but a grid. “At the base you have Khazna, then Core42 for GPU-as-a-service,” he said. “On top is Inception – the model-building and product company.” The stack matters because it explains how the Abu Dhabi group G42 tries to turn computation into outcomes without the usual detours. “We work hand in glove with domain companies – energy, healthcare, weather – so use cases are grounded in reality, not just lab demos.”
Inception began life in 2017 as a research institute for frontier models. That phase produced Jais, a large language model with an Emirati dialect spine, and cousins for Hindi and Kazakh, with Swahili and Azerbaijani under way. In 2024 the group pivoted. “We moved from ‘Inception 1’ – research first – to ‘Inception 2’,” Koshy said. “Same scientific depth but aimed at market-needed applications.” He became chief executive late that year, though he had been with G42 since its earliest days. The move reads less like a handover than a narrowing of the beam.
Local by design
The regional tilt was not a retrofit. Years before the present hype the UAE appointed a minister for AI and built MBZUAI, a university dedicated to the field. That foresight set a tone. “AI must be democratised,” Koshy said. “Nations should not be limited to models trained elsewhere with assumptions that do not fit local culture or policy.”
For him language is not a cosmetic layer but a vessel of norms. “Jais carries a strong Emirati flavour,” he said. “We are doing the same for other languages because accuracy and tone in the local language affect trust and adoption.”
The practical stakes are obvious enough. Arabic content engines that understand idiom. Hindi-first customer service at scale. Enterprise tools that respond to how people actually speak rather than how documentation imagines they should. The point, Koshy said, is utility over theatre – a model that reads you as you are.
Beyond the demo
Koshy is wary of AI that never leaves the slide deck. “Durable ROI comes when AI is embedded in core operations – not when it sits in a demo,” he said. His team ships an In-Business suite across procurement, productivity and process improvement, including no-code tools for building agents. But he resists the idea that agents are plug-and-play.
“A procurement agent must reflect your delegation of authority, your policies, your ERP integrations,” he said. “You need subject-matter experts in the loop. We co-design, set KPIs with the business, then measure against them.”
Some paybacks are quick. He points to a “board observer” agent that acts as a companion in board meetings – compressing long decks into briefings, surfacing risks and action items so limited time can be spent on strategy rather than page-turning. Others take quarters. “End-to-end procurement or investment agents span the value chain,” he said.
“They need robust integration and change management.” The discipline is case by case: run proofs that touch the real workflow, scale only when both the technical and cultural bars are cleared.
Partnerships keep the machine from trying to build every cog in-house. In energy, Inception leans on AIQ. For procurement, Fairmarkit. For model work, universities. “Do not reinvent what you can partner on,” Koshy said. “Invest where you can differentiate.”
Data and the border
If models are the surface, data is the bedrock. In a region where sovereignty and privacy rules vary, Koshy begins with governance. “We deploy to the posture the client chooses – air-gapped on-prem, sovereign cloud or public cloud with client-held keys,” he said.
“We ship models into the client’s tenant so data stays under their controls.” For proofs without access to sensitive records, his teams generate synthetic datasets. Cross-border rules are “design inputs, not obstacles,” he said. The non-negotiables are simple: transparency and auditability.
Responsible AI shows up in the plumbing rather than the press release. Group-level governance playbooks arrive with the product. Teams annotate trustworthy data, design auditable processes and implement risk controls. “Decisions must be explainable and repeatable – not hallucinated,” Koshy said. The bias is to publish when possible, even down to training datasets, because traceability invites scrutiny and hardens practice.
Two fronts, one platform
Inception’s market stance is tidy. “We run two fronts on one platform,” Koshy said. The first is Corporate AI: back-office and knowledge-work functions like procurement, finance, legal and HR, wired with agent workflows customised per enterprise. The board observer sits here, licensed alongside the base platform so organisations can widen use over time.
The second is domain-specific – stacks like Investment AI, fine-tuned on sector data to support sourcing, diligence and even voting workflows. The company is industry-agnostic yet enterprise-focused, which lets clients mix the horizontal with the vertical where impact is clearest.
Ambition is measured in returns rather than slogans. “In three years I want ‘AI in the region’ to be synonymous with G42 and Inception because customers can point to measurable ROI,” Koshy said. The plan involves a team that can handle national-scale problems, plus partnerships for when demand outpaces current capacity. He also wants early adopters to become messengers. “Leaders who took a leap of faith with us should have the metrics to show it,” he said.
The human stays
For all the talk of agents, Koshy is blunt about the boundary. “Skynet is great cinema, not a roadmap,” he said. “We are in narrow AI with fast progress in physical AI, but a human remains in the loop in our systems.” High-stakes domains demand proportionality: a missed calendar entry is tolerable, moving money is not.
The human sets thresholds, supervises exceptions, owns the trade-offs. “AI should move people from the mundane to the strategic,” he said. The posture extends to culture. Talent joins for pay, environment and meaningful problems. Inception sells the third. “Work on oil and gas optimisation one day, investment workflows the next, frontier models the day after,” Koshy said.
He looks for skill and attitude – people who like hard problems and accept the pace. The result is a shop where a 1 a.m. call happens because the challenge matters, not because anyone is forced.
A builder’s habit
Koshy’s answers circle back to the idea of impact. Before Inception he helped build G42’s healthcare arm from employee one through the peak of COVID, including the Emirati Genome Programme that many thought impossible at that scale. “We turned a lab into what I call the United Nations of sequencing,” he said, “by bringing together partners who rarely collaborate.” AI is the next frontier – an “open lab” of compute, data and permission to build – and he sounds happiest when the work moves from pitch to production. “I want us to leave deployed systems that matter,” he said.
The grid metaphor holds. A base of compute. A layer of models. A set of applications that live or die by whether they change the working day. Local language as a design choice, not a gloss. Governance as engineering, not theatre. The human – still, stubbornly – in the loop. “Durable ROI is when AI becomes part of the operation,” Koshy said. He made it sound less like a revolution than a craft – a set of fits, tests and integrations that, if you are patient, turn into outcomes.
