Egypt told AI startups to compete on local knowledge, not model access

THE SIGNAL IN ONE SENTENCE

If every startup can call the same model, the model is not the startup. That was the useful provocation from Ahmed Elzaher, chief executive of Egypt's Information Technology Industry Development Agency, at Techne Summit Alexandria. As ITIDA summarized his argument on October 3, smaller teams now have access to capabilities that once required much larger resources. That access also raises the competitive bar. The differentiators become a deeper understanding of the customer, specialized data, industry expertise and strong customer relationships. In plain English: the clever demo is getting cheaper. Knowing what should happen after the demo is still expensive. Egypt wants local startups to build that second part. ITIDA named healthcare, education, financial technology, agriculture and government services as areas where Egyptian context can become an advantage. It also said it is preparing an AI competition with a government ministry in which startups and other teams would work on real-world challenges. The ministry was not named. Neither were the challenges, launch date, budget, rules, datasets, procurement route or evaluation method. So the plain signal is not that Egypt has already built a perfect demand engine for AI startups. It is that the country has identified the right bottleneck. Access to a model is useful. Access to a problem owner, trustworthy local data, a testing environment and a buyer is what turns a model call into a company. This matters because the current AI startup market makes it easy to confuse product assembly with durable advantage. A team can connect a language model to a polished interface in a weekend. It can add document upload, retrieval, voice, a dashboard and an agent that politely announces each step. Investors and officials can watch a smooth demonstration. Then the system meets an Egyptian clinic with handwritten records, an Arabic dialect the benchmark did not test, a public office with several databases that disagree, a farmer with unreliable connectivity or a small business that cannot afford the workflow. The model was the easy part. Local knowledge begins with the actual job. In healthcare, it means understanding how a patient moves through a clinic, which records exist, who may see them, how Arabic and English terminology mix, what happens when the system is uncertain and which professional remains accountable. In education, it means the national curriculum, teacher workload, classroom size, device access, dialect, disability support and the difference between helping a student learn and merely producing the answer. In agriculture, it means crop cycles, irrigation constraints, regional practices, weather, soil, prices and whether advice can reach the farmer at the right moment in a usable form. In government services, it means law, procedure, identity, appeals, records, accessibility and the unpleasant fact that a wrong answer from a public system can block a benefit or send a citizen to the wrong office. None of those problems is solved by declaring that a model can reason. They require domain experts, users and operators who can say what success actually looks like. That is why ITIDA's four stated priorities are more interesting than another startup showcase. Elzaher named skills, resources, real-world application and scale. Each one is a gate. Skills means more than prompt technique. A serious team needs product judgment, software engineering, data work, security, evaluation, sales, design and enough domain knowledge to recognize a harmful shortcut. Resources includes cloud and computing access, but also clean data, secure environments, legal help, test users and time. A credit balance can buy experiments. It cannot persuade an institution to open the right dataset or approve a pilot. Real-world application means a problem owner is willing to define the need, expose the messy workflow and let the product be measured against what happens today. Scale means the company can repeat the result without rebuilding itself for every customer. Egypt already has pieces of that support system. ITIDA's Start IT programme offers a one-year incubation path for early-stage technology ventures. Current agency material describes a package worth EGP 480,000 with no equity requirement, up to $10,000 in Amazon Web Services credits, workspace, hiring support, technical mentoring, AI consultation, investor access and market exposure. CREATIVA Innovation Hubs provide facilities and programmes across multiple governorates. ITIDA also points to accelerator partnerships with Plug and Play and 500 Global, and to the Venture Ready programme with GIZ. Those are inputs. The harder question is whether they reliably produce customer evidence. Cloud credits can create a runway, but credits expire. Mentoring can improve a pitch, but a pitch is not product-market fit. An expo can introduce a buyer, but a photograph beside a booth is not a contract. The planned government competition could connect the missing pieces if it is designed as a procurement and evaluation pipeline rather than a one-day performance. Start with the challenge statement. A ministry should describe a specific public problem, the current process, affected people, available data, legal limits, service standard and failure cost. It should publish what the team may change and what remains outside scope. The statement should not prescribe AI before the teams investigate the problem. Sometimes the best solution is a better form, a reliable search system, a rules engine or two databases finally agreeing about a person's name. A challenge that rewards an AI label rather than a better outcome will select for theatre. Next comes data access. Local data can be a competitive advantage only if it is lawful, representative, documented and usable. A pile of records is not a dataset. Teams need a data dictionary, provenance, known gaps, consent or other legal basis, retention rules and a secure place to work. Public-sector data can contain health information, identity records, financial details or histories of people who never volunteered to help a startup win a competition. Privacy cannot be added after the finalist has already copied everything into a commercial model. The competition should provide a controlled environment, minimize personal data, use synthetic or de-identified material where appropriate and require a data-protection review before any live pilot. Arabic performance needs deliberate testing. Modern Arabic, Egyptian Arabic, mixed Arabic and English, transliteration, technical vocabulary and regional speech patterns are not interchangeable. A system may look competent in a prepared demonstration while failing on the language people actually use. Evaluation sets should be created with local experts and then kept partly hidden from competitors. Otherwise teams will tune to the test instead of the public need. Metrics must follow the service. A health workflow may measure missed cases, false alarms, clinician time, subgroup performance and the rate at which a professional overrides the system. An education tool may measure learning, teacher workload, accessibility and whether students without premium devices receive the same opportunity. A government assistant may measure completed transactions, wrong referrals, appeal rates and human escalation. Latency and benchmark scores matter. They are not the citizen outcome. The buyer also needs a baseline. How long does the existing process take? What does it cost? Where do people drop out? How often do staff correct an error? Without that starting point, almost any pilot can claim improvement. The competition should fund the work required to collect the baseline rather than demanding that startups invent a success story around missing records. Procurement is the next trap. A team may win the challenge and still have no path into a contract. Government pilots often occupy an awkward zone where the agency wants innovation but normal purchasing rules begin only after the demonstration. The rules should say what winners receive, who owns the intellectual property, whether the ministry can buy a successful system, how conflicts are handled, what security standards apply and how unsuccessful teams recover or delete the data. Payment milestones should reward evidence: a reproducible prototype, a safe pilot, verified user benefit and a maintainable deployment. Not simply the best five-minute pitch. The government's role is not to guarantee that a startup succeeds. It is to make the test honest. That means publishing evaluation criteria before teams build, separating mentors from judges, declaring conflicts, including affected users, allowing independent technical review and releasing a result report after the competition. The report should include failures. If a tool performed worse for women, rural users, disabled people or a particular language group, that is not embarrassing debris to sweep behind the stage. It is the information the next team needs. ITIDA's October announcement also named ten startups receiving space in its Techne Summit pavilion: KoCyber, STEMulator, Fabritec, Aman 7, Dackatra, Just2Pay, Polaris GRC, Dragons, Syncode X and EVRAID. The agency says the group spans cybersecurity, education technology, software services, property technology, health technology, financial technology, AI, automation, deep technology and electric mobility. A pavilion can create useful introductions. It does not prove that each company has a defensible local dataset, revenue, an effective product or a global path. The announcement provides no comparable performance information for the ten firms. That limitation should remain visible. Government startup policy often counts the easiest things: applications, cohorts, events, training hours, booths and investment meetings. Those figures show activity. They do not show whether customers kept using the product, public outcomes improved or the company survived after support ended. Egypt's National AI Strategy for 2025 to 2030 sets ambitious goals. It calls for responsible governance, accessible data, scalable infrastructure, support for local startups and stronger talent. It targets more than 250 successful AI companies and 30,000 AI professionals by 2030. Those are targets, not achievements. The strategy itself recognizes that the AI industry needs governance, technology, data, infrastructure, an ecosystem and talent. The Techne Summit message translates that national list into a founder's problem: what do you know about a customer that the shared model does not? There is one more complication. Model access is not actually equal. Startups face different prices, rate limits, data-location choices, language performance, enterprise terms and access to advanced features. A company with more capital can buy more experiments and absorb more failed runs. A provider can change the model, price or policy beneath a product. Local advantage cannot become an excuse to ignore platform dependence. Egyptian startups need portable data, modular architecture, model comparisons, fallback options and a plan for costs after cloud credits end. Public buyers should require exit terms and test whether a service can move between models without losing its records or evaluation history. Open models and national infrastructure may help in some cases. Commercial systems may be better in others. The choice should follow evidence, language, security, cost and maintainability rather than flag-waving or vendor worship. The durable moat is not merely local data. It is a trusted loop. A real user describes the problem. A local team builds with lawful data. Domain experts test the result. The customer reports failure. The team repairs the system. The buyer measures the outcome. The knowledge from that cycle becomes difficult to copy because it lives in relationships, operations and evidence. That is the version of local knowledge worth competing on. Egypt's proposed ministry challenge could build the loop. First it has to publish the challenge.