Dubai-India Agentic AI Pact: What Businesses Should Build First

Dubai’s latest AI agreement is an ecosystem deal, not a deployment contract.

On 19 August, Dubai Chambers signed a memorandum of understanding with India’s National Association of Software and Service Companies, or NASSCOM. The official announcement covers partnerships, knowledge exchange, business events and support for Indian agentic AI and deeptech companies expanding through Dubai. NASSCOM represents more than 3,500 companies across an Indian technology industry it values at US$315 billion.

That creates a useful route between a large supplier base and a city trying to move businesses from AI experiments to systems that can take action. It does not, by itself, tell a buyer which agent is safe to connect to finance, customer data or production systems.

What the agreement changes

The MoU gives Dubai Chambers and NASSCOM a formal structure for bringing companies, investors and entrepreneurs together. Indian vendors can use Dubai as a base for regional expansion. Dubai businesses get a wider pool of specialists, products and implementation partners.

This is the newest piece of a programme announced earlier in the year. In May, the Government of Dubai launched a two-year private-sector transition programme. It called for training across Dubai Chamber of Commerce business councils, dedicated incubators and investment funds. In June, Dubai Chambers formed an Executive Committee for Agentic AI to develop the implementation framework.

The new India link can help fill that framework with suppliers and expertise. The National and NDTV Profit both reported the agreement as a route for business expansion and private-sector adoption.

The announcement does not name selected vendors, signed customer contracts, funding allocations, deployment dates or technical controls. Companies still have to do the hard part: pick a narrow workflow, set boundaries and prove that the agent behaves safely.

Start with authority, not the model

A chatbot produces text. An agent may read records, call tools, change a ticket, issue a refund or start a payment. The important design question is therefore not only which model performs best. It is what authority the system receives.

Give every agent its own workload identity. Do not let it borrow a staff member’s account or reuse a shared API key. The permissions should match one defined job, with separate approval for any wider scope.

This is moving into standards work. The US National Institute of Standards and Technology has an AI Agent Standards Initiative covering interoperability, security and identity. Its National Cybersecurity Center of Excellence is also examining standards-based agent identity and authorization, including how organisations identify agents and control their access and actions.

For a Dubai buyer, the practical requirement is simple: the audit log should show which agent acted, which policy allowed it, what data it used and which human approved the step when approval was required.

Move through four adoption gates

The safest first deployment is usually boring. Let the agent observe a workflow and produce a recommendation without changing the underlying system.

A customer-service agent might classify cases and draft responses. A finance agent might flag invoice mismatches. A logistics agent might suggest a reroute. The team can compare its output with real outcomes before granting any action rights.

The second stage adds human approval. The agent can prepare an action, attach its evidence and wait. This exposes weak reasoning while the cost of a mistake is still contained.

The third stage permits reversible actions inside a strict boundary. An agent could update an internal field, create a draft or reserve a slot that can be cancelled. Use allowlisted tools, transaction limits, complete logs and a tested rollback path.

High-impact autonomy comes last. Payments, legal commitments, account closures, medical decisions and changes to regulated records need stronger approval, continuous monitoring and an immediate kill switch. Some actions should remain human decisions even when an agent can prepare the work.

Agentic AI adoption gate. Sources: NIST AI Agent Standards Initiative, NIST agent identity and authorization project and OWASP Top 10 for Agentic Applications 2026. Original SultanByte infographic, 20 August 2026.

Treat tool access as an attack surface

Agents receive instructions from more places than the user prompt. They may read emails, documents, websites, database rows and messages produced by other agents. Any of those sources can contain hostile or misleading instructions.

That is why ordinary application controls still matter. Validate tool parameters outside the model. Separate untrusted content from system instructions. Restrict destinations, file types, recipients and transaction values in code. Never rely on a sentence in a prompt as the only control protecting a payment or deletion.

The OWASP Top 10 for Agentic Applications 2026 groups the major risks facing autonomous systems. Buyers do not need to turn its list into a paperwork exercise, but they should expect vendors to show how the product handles prompt injection, excessive privilege, unsafe tools, poisoned memory and failures across several connected agents.

A useful test is to feed the agent a legitimate-looking document that contains an instruction to ignore policy and send data elsewhere. The surrounding application should block the action even if the model tries to comply.

What Dubai buyers should ask for

A polished demo is easy because the vendor controls the data, tools and failure cases. Procurement should request evidence from conditions closer to production.

Ask the vendor to document the agent’s identity model, permission boundaries and supported approval points. Require an exportable audit trail and a way to replay a failed run. Confirm where prompts, memory, tool results and customer data are stored. Check whether administrators can revoke one agent without disabling the whole platform.

Demand measurable acceptance criteria for the chosen workflow. Accuracy alone is weak when the system can act. Track unauthorized action attempts, approval rejection rates, rollback frequency, time saved after review and the cost of human supervision. A system that completes more tasks but creates a larger review queue may not improve productivity.

Contracts should also cover incident notification, log retention, model or tool changes, subcontractors and exit. If a vendor changes the underlying model, the business needs to know whether earlier evaluations still apply. If the relationship ends, the customer needs its logs and data in a usable format.

What Indian vendors should bring to Dubai

The MoU can open doors, but vendors will still compete on operational fit. A strong regional offer needs more than an agent demo with an Arabic interface.

Show how the product handles Arabic and English content in the same workflow, including names, dates, numbers and right-to-left documents. Explain data hosting and administrative access precisely. Support the identity, logging and approval systems already used by the customer instead of asking the customer to replace every control layer.

Pick one sector and one workflow where the economics can be measured. A vendor that reduces manual reconciliation time for a bank or cuts exception handling in logistics has a clearer case than a platform promising autonomous transformation across the enterprise.

Local partners should also be able to operate the system after the pilot. Training cannot stop with prompt writing. Security teams, process owners and auditors need to understand permissions, failure handling and rollback.

The test is controlled production use

Dubai’s agreement with NASSCOM can bring more suppliers, training and investment into the same programme. That is useful infrastructure for adoption. It does not remove the buyer’s responsibility to control what an agent can see and do.

The strongest projects will start with a narrow workflow, give the agent a distinct identity, require approval before consequential actions and expand only after the logs show that the controls work. The partnership may increase the number of agentic AI options in Dubai. Production evidence will decide which ones last.

Cover and infographic: original SultanByte visuals based on the official and technical sources linked above.