Three factories, one product thesis
On 21 September 2026 at HUAWEI CONNECT in Shanghai, Huawei Digital Finance BU CEO Jason Cao launched the Huawei Fintelligent AI Solution under the slogan “Own Your AI, Own Your Intelligence.” Coverage from PR Newswire describes a stack built around three “intelligent factories”: Agent Factory, Token Factory, and Data-Knowledge Factory, plus the open-source multi-agent platform openJiuwen, the TokeNexus token lifecycle layer, and a Financial Agentic Data Solution. Huawei says it already serves more than 7,100 financial customers in 80+ countries, including 54 of the world’s top 100 banks.
For MENA product and engineering leaders, this is not a chip war headline. It is a packaging of agentic banking as an operating model: agents as managed assets, tokens as a cost centre with governance, and bank knowledge as a factory input—not a one-off RAG demo.
Why “own your AI” is a UX and org-design problem
Cao’s key message: competition is shifting from model capability to enterprise-wide AI system capability. Owning AI means independent choice, secure operation, scale replication, and continuous evolution—not merely calling a frontier API. That maps cleanly onto what Gulf banks already hear from SAMA, CBUAE, and board risk committees: where does the model run, who can revoke an agent, what is the unit economics of inference, and can the bank explain a credit or KYC decision in Arabic and English?
iFynx work with bank digital units shows the failure mode clearly. Teams buy copilots, then discover they have no inventory of agents, no token budget owner, and no knowledge pipeline that survives staff turnover. Fintelligent’s “factories” language is useful because it forces those ownership questions into the roadmap.
Product moves MENA teams should make this quarter
1. Treat agents as versioned products, not chat widgets. Agent Factory’s pitch—build, run, manage, accumulate, reuse—implies an internal catalog with owners, SLAs, and retirement dates. Ship a bilingual agent registry in your bank app ops console before you ship the tenth chatbot.
2. Make token economics visible to product managers. TokeNexus covers planning, production, scheduling, and optimization across the token lifecycle. Even if you never buy Huawei gear, copy the product surface: per-journey cost dashboards, soft budgets per agent, and alerts when a collections agent burns 3× its forecast. Token opacity kills agent programs quietly.
3. Upgrade data supply to knowledge supply. The Data-Knowledge Factory story is the hard part: documents, rules, and legacy code become assets agents can trust. Pair ontology-centric modernization (Huawei’s AI Coding Solution angle) with human review queues. In Islamic banking and Sharia-compliant products, “knowledge” includes fatwa references and product rulings—design for that, not generic PDFs.
4. Prefer open multi-agent kernels with audit hooks. openJiuwen is positioned as financial-grade, long-running, and compute-aware. Whatever platform you choose, insist on session replay, tool-call attestation, and kill switches that compliance can exercise without paging an SRE at 2 a.m.
5. Use lighthouse cases as RFP criteria, not wallpaper. Huawei unveiled a 2026 Global Financial Lighthouse set of nine practices. Translate those into measurable acceptance tests: time-to-first-agent in production, percentage of journeys with human-in-the-loop gates, Arabic content coverage for agent replies.
Implementation checklist (iFynx craft)
- Agent asset register with owner, risk tier, revoke path
- Token budget + anomaly alerts per agent and per journey
- Knowledge pipeline with provenance stamps and bilingual QA
- Sandbox that mirrors production tool permissions
- Synthetic monitoring for agent drift (prompt, tool, policy)
- Board-ready one-pager: “what we own vs what we rent”
iFynx takeaway
Fintelligent reframes agentic banking as industrial process, not model shopping. MENA banks that own agent inventories, token economics, and domain knowledge will outrun teams that only rent chat endpoints. Design the factories first—then pick the models that plug into them.
Field notes from MENA delivery teams
When we run agent discovery workshops with Gulf banks, three gaps appear within the first hour. First, nobody owns the agent catalog—IT thinks product owns it, product thinks the vendor owns it, risk thinks it does not exist yet. Second, token spend is invisible until finance asks why the cloud bill doubled. Third, knowledge is trapped in SharePoint folders with no provenance, so agents invent policy. Fintelligent’s factory framing is useful precisely because it assigns those gaps to named workstreams.
A practical 90-day plan: weeks 1–2 inventory every bot, copilot, and RPA script that touches customer data; weeks 3–6 stand up a minimal agent registry with risk tiers; weeks 7–10 connect token metering to the registry; weeks 11–13 pilot one knowledge pipeline for a single journey (for example, card dispute FAQs in Arabic and English) with human QA. Do not wait for a perfect platform RFP. The factories are a management system you can start on current vendors.
Also watch vendor lock-in language. “Own your AI” should include portable prompts, portable evaluation sets, and portable audit logs. If a platform cannot export agent definitions and conversation traces in a documented format, you do not own much. Insist on exit drills the same way banks insist on core-banking exit plans.
Originally published on iFynx.
Top comments (0)