How DEWA Turned a Generative AI Pilot Into a Working Service

Rammas was already answering questions before Dubai’s utility added generative AI. The rollout shows what changes when an assistant has to interpret requests, connect to existing workflows and leave a clear path to human help.

In November 2023, Mohamed Obaid Alsharid, then chief digital officer at Dubai Electricity and Water Authority, described two versions of Rammas in a published interview. One was the virtual assistant DEWA had launched in 2017 to help customers with enquiries and transactions. The other was a generative AI beta that customers could try on the website and in the app. The distinction matters. DEWA was introducing a new way to understand and answer questions within a service customers already used. It also had an existing set of tasks and digital channels against which to judge the change. Telecom Review’s published interview with Mohamed Obaid Alsharid

The later public record shows Rammas moving further into customer care. A case study in Digital Dubai’s State of AI report describes the assistant extracting details from enquiries, identifying intent and starting workflows in internal systems. In July 2026, DEWA also announced that it had deployed agentic AI across its digital platforms. The public accounts describe what the service was intended to do and which channels were involved. They give much less detail about the tests, release approvals and engineering ownership behind each step. That distinction sets the limits of what this case can tell another team planning a launch. Digital Dubai’s State of AI report, case study 14, DEWA’s July 2026 rollout announcement

A pilot inside an existing service

In May 2023, DEWA announced that Rammas had been piloting ChatGPT on its website since April. By August, the utility reported that the enhanced assistant had answered 32,084 queries on the website and in the smart app from the end of April through the first half of 2023. It also reported a 90% customer happiness figure for the pilot. Those figures show that the beta encountered real customer traffic and that DEWA collected feedback. They cover different things. One counts questions answered. The other reflects a reported customer rating, with no published response denominator or scoring method in that announcement. DEWA’s May pilot announcement, Dubai Media Office’s report on the pilot

Alsharid’s later interview described the generative beta as available on the website and app and mentioned a different happiness rating. Without the sampling periods and calculation methods for both figures, they cannot establish a change in satisfaction. A separate 2024 DEWA update listed several channels through which Rammas was available in Arabic and English. That broader channel list describes the assistant. DEWA specifically identified the website and app when describing the ChatGPT introduction. The public record does not establish that each channel offered the same generative features at the same time. DEWA’s February 2024 update

The chronology also includes a social channel. In March 2023, before the public pilot dates described above, DEWA announced that customers could reach the existing Rammas assistant through Instagram’s Message feature. The social account thus served both as a place to publish utility updates and as an entry point for enquiries. An engineer planning a rollout across channels has to decide what each entry point can handle. A public-facing profile can offer general guidance and point to a secure route for account-specific work. A logged-in app can preserve more context across the customer’s journey. The fact that both routes begin with a message gives them a common appearance, while the information available to the service may differ. DEWA’s Instagram channel announcement

DEWA’s infrastructure reel illustrates the public side of the same account it identified as a route to Rammas.

Starting with an existing assistant gave DEWA a baseline in the form of a service with known tasks. A team in a similar position can use that history to select questions whose answers can be checked against an approved source, identify tasks that users already complete through another route and find the points where customers routinely need staff. These comparisons are strongest when a team can follow the same kind of request through both paths. Measuring a new conversational interface against an old phone queue without adjusting for the kinds of requests each receives can make a change look better or worse than it is.

The move from a reply to a workflow

The Digital Dubai case study, attributed to DEWA, describes pressure on the customer care centre from requests arriving across different channels. It describes Rammas for You as an initiative that can interpret requests, extract information, route tickets and start internal workflows. It also describes escalation for cases requiring staff and reports a 25% reduction in ticket resolution time. The case study does not publish the underlying dataset, the period used for the comparison or the distribution of tickets by type. The figure is useful as a reported operational measure and has limited value as a benchmark for another deployment.

There is a practical engineering distinction between giving an answer and starting a workflow. A customer who asks what documents a service needs can be directed to a published list. A customer who asks whether a specific request has already been lodged may need to be identified, and the answer needs to reflect a current service record. A request that changes the record involves another step. The system must know what the user intended, whether it is permitted to act and whether the action was confirmed by the service that actually owns the record.

For a team building such a system, that suggests testing the chain as a sequence of decisions. First, did the assistant interpret the question correctly? Then, did it select the permitted action and pass the correct details? Finally, did the underlying system complete it and return a result the customer can understand? If the assistant only receives a receipt that a request was accepted, the message should say that it was accepted. Any later confirmation has to come from the system handling the request. This is a design implication of the type of workflow described by Digital Dubai. It is separate from DEWA’s unpublished implementation details.

Consider an illustrative failure. A customer sends the same request twice after a slow response, then switches from the app to the website. If the service is designed only around individual chat sessions, the assistant may treat the second attempt as a fresh request. If the underlying system can recognise a duplicate, the second attempt can return the state of the first one. The important test checks both the customer message and the number of requests actually created. The answer can read perfectly while the back office still has duplicate work.

Evaluate the point where the customer needs help

An engineer preparing an assistant for wider use should distinguish a plausible reply from a correct response and a completed task. To assess the first two, review questions against the documents and policy versions that were active when they were asked. To assess the third, compare the conversation with the eventual service outcome. A customer who accepts an answer about the process may still return later because the process was never completed.

Bilingual testing should reflect the way people ask for help. A literal translation can miss informal Arabic phrasing, mixed Arabic and English or a technical term customers recognise only in one language. For an illustrative utility case, test a question about an old bill, a request with an incomplete reference and a message that uses the word for a connection to mean two different things. The model may need to ask a short clarifying question in the language the customer used. A review set should capture whether that question resolves the ambiguity and whether the answer stays consistent once the intent is clear.

A case that needs staff is part of the service design. The person receiving it needs enough context to continue without making the customer repeat everything. They also need to see whether the assistant already tried an action and what happened. An engineering team can measure handoff quality by reviewing repeated questions, reassignment between queues and the time staff spend establishing what happened. A fall in chat volume means little if customers reopen the same case by phone.

Give the service an owner after release

DEWA’s July 2026 announcement describes Rammas drawing on information on the utility’s website. That makes content maintenance an operational question as well as a writing task. If instructions change, someone needs to verify that the website reflects the change and that the assistant can find the correct information. If two pages give conflicting instructions, the team needs a way to identify the preferred source or send the question to staff. When the assistant can start a request, changes in the back office can also require fresh workflow tests.

After a change, review a fixed set of requests that covers both languages and the highest-impact actions the service can initiate. Include the paths customers use when a step times out or returns an unfamiliar status. Give the reviewer the expected result in the underlying system as well as the expected conversation. If an update improves how answers read but makes confirmed actions harder to distinguish from pending ones, the review should catch that before exposure expands.

A controlled rollout gives an engineering team a way to notice those changes without losing sight of ordinary operations. It can release one request type to a limited audience, record where cases go and compare the results with the established channel. If there is a rise in duplicate requests or an important source becomes unavailable, the team can narrow the affected functionality while keeping other routes available. Before release, someone needs the authority to make that call and access to the records that show when to make it. This describes a practical operating model for a similar service. DEWA’s public accounts do not disclose its internal rollback thresholds or on-call procedures.

Alsharid’s published account identifies what customers could try in 2023. The later project records show what the organisation said it had put into operation. Between those milestones, a service team has to define what the assistant can do, tie those actions to the systems that confirm them and give staff enough information to pick up cases that need judgement. For an engineer taking a pilot into wider use, the measure of that work is whether the customer can see what happened and still has a workable route when something fails.