The definitive voice of Sri Lanka’s technology industry since 2014.
Type at least two characters to search articles.
This week, Sri Lanka will see a lot of AI demos.
Some will be genuinely impressive. Some will be a chatbot connected to a PDF. Others will have the word ‘agent’ doing more work in the pitch deck than it does in the product.
That is not a criticism of Sri Lanka AI Week. In fact, the timing is useful. The country is finally treating AI as something bigger than a conference topic. The 2026 Budget allocates money for AI, cloud infrastructure, data centers , local-language datasets and startup support, while the government has also spoken about a National AI Institute and a broader digital-economy push.
Sri Lanka AI Week 2026 runs from 28 September to 4 October, organized by the Ministry of Digital Economy, GovTech Sri Lanka and SLT-MOBITEL. The official announcement describes an ambition to move Sri Lanka from being mainly a technology consumer to becoming a producer of AI-driven products and services.
Good, but that’s also where the easy part ends.
In my day-to-day work as a CTO, the most useful AI conversations are rarely about whether a model can generate a good answer. We already know it can. The uncomfortable questions start when somebody says, ‘Can we put this in production?’
A demo has ten users. A product has the student who uploads a 180-page PDF at 10:47 p.m., the customer who asks the same question six times because the answer is slow, the employee who pastes confidential data into a prompt, and the finance team that wants to know why the API bill doubled in a week.
A demo gets applause when it works once. Production software has to work again tomorrow.
That difference matters because access to AI models is no longer a competitive advantage by itself. OpenAI, Anthropic, Google, Meta, Mistral and others are making capable models available to almost everyone. A Sri Lankan startup and a company in San Francisco can often call the same API within minutes of each other.
The real advantage is everything around the model: the workflow, the data, the local context, the product design, the security controls, the cost discipline, and the ability to keep the system useful when the provider is having a bad day.
This is where AI gets very Sri Lankan, very quickly.
A team can build a prototype on free credits and conclude that the economics look fine. Then usage grows. Prompts get longer. Users upload documents. The product retries failed calls. Someone adds a second model for ‘better quality.’ Suddenly the feature that looked nearly free has a variable foreign-currency cost attached to every interaction.
For a company earning mostly in LKR, that is not a small implementation detail. It is part of the business model.
Before adding an AI feature, I want to know the maximum amount we are willing to spend on one successful task. Not per token. Not per API call. Per useful outcome.
If a student asks a short question, does it need the most expensive model available? If a system is classifying a support ticket, does it need the same reasoning budget as a long research task? If a document has already been processed, should we pay to process the same context again?
Those are product questions disguised as infrastructure questions.
Sri Lankan teams are usually forced to become cost-conscious earlier than companies in richer markets. That can actually be an advantage. Cost pressure makes you question waste before waste becomes architecture.
There is another mistake I see often: the team chooses the model first and designs the experience around whatever latency it gives them. It should be the other way around.
If somebody is chatting with a product, a long silent pause feels broken even if the final answer is excellent. Streaming can help, but it does not magically make every workflow interactive. Large-document analysis, report generation, and multi-step research often belong in a background job with progress updates, not inside one long HTTP request while the user stares at a spinner.
This sounds obvious once written down. It is surprisingly easy to ignore when everyone is excited about getting the first response from a new model.
The better question is: how long is this user willing to wait for this particular task? Then pick the model and architecture that fit the answer.
We learned this lesson with payment gateways, SMS providers, email services, and cloud platforms. AI should not get a special exemption because the technology is new.
Providers rate-limit. Regions have incidents. Models change. Pricing changes. Responses occasionally come back in the wrong format. A model can refuse a request that worked yesterday. A tool call can fail halfway through a workflow. If your application has no answer for these situations, the AI feature is not production-ready. It’s just a successful happy path.
The solution is not to connect five model providers on day one. That’s another form of overengineering. A small team can start with one provider and still keep model access behind an internal interface, set hard timeouts, limit retries, log failures, and make it possible to add a second route later without rewriting the product.
The important thing is to decide what failure should look like before users decide it for you.
The most serious problems begin when the model is allowed to do things, not just say things.
Reading a document is one level of risk. Sending an email, updating a medical record, issuing a refund, changing a student’s data, or calling an internal administrative tool is another.
A system prompt saying ‘only do this when authorized’ is not access control.
The model can suggest an action. The application still has to check who the user is, whether that user is allowed to perform the action, whether the arguments are valid, and whether a human needs to approve it.
The OWASP GenAI Security Project highlights risks such as prompt injection, sensitive-information disclosure, and excessive agency. These are not exotic research problems. They are practical software-security problems once an AI system is connected to real data and real tools.
My simplest rule is to treat model output the same way we treat input from any untrusted external system: useful, but not automatically trusted.
Sri Lanka also has a problem that does not show up neatly in global model benchmarks: we do not operate in one language.
A product may need English for the interface, Tamil for one user, Sinhala for another, and a mixture of English and local-language terms inside the same conversation. Anyone who has handled actual Sri Lankan customer support knows that people do not politely stay inside one language boundary.
A model saying it ‘supports Tamil’ or ‘supports Sinhala’ is not the same as the product being good in those languages. You need to test the language people actually use: colloquial phrasing, code-switching, local names, education terms, government terminology, and the wonderfully inconsistent spelling that appears in real chat messages.
The 2026 Budget specifically mentions funding for local-language datasets as part of the country’s AI push. That’s important because local context is one of the few areas where Sri Lanka can build something that’s not immediately commoditized by access to the same global model APIs.
The budget also sets aside funding for AI and cloud facilities, selected AI projects, and digital infrastructure. The published 2026 Budget Speech puts those investments inside a much wider digital-economy program.
I’m not worried that Sri Lanka is ‘too excited’ about AI. Excitement is useful. It gets students experimenting, companies investing, and people from different industries into the same room.
What I would like to see after AI Week is a change in the questions we ask.
Instead of only asking, ‘Which model are you using?’ ask: What does this save? What happens when it fails? What data does it send outside the organization? How much does one successful task cost? Can a user understand when AI is making a suggestion rather than a decision? Does it still work well in Tamil and Sinhala? Who gets called when it breaks at 2 a.m.?
Those questions are less exciting on a stage. They are exactly what turns a demo into infrastructure.
We are not going to win an AI infrastructure arms race against the United States, China, or the largest technology companies. We do not need to.
Sri Lanka’s opportunity is much more practical: use increasingly commoditized AI capability to solve problems we understand unusually well.
Education workflows that are still held together with WhatsApp messages and spreadsheets. Healthcare administration with too much manual work. Small businesses that cannot afford large software teams. Local-language services. Agriculture. Public-service workflows. Export software built by small teams that have learned to operate efficiently because they never had the luxury of wasting money.
The model does not have to be Sri Lankan for the product to be Sri Lankan. The product becomes ours when it understands our users, our constraints, and our economics better than a generic global interface does.
AI Week will end on 4 October. The booths will come down. The LinkedIn posts will slow down. Everyone will have a folder full of photos from panel discussions.
Someone has to turn the prototype into a service people will pay for. Someone has to monitor it, secure it, explain it to users, control the cost, handle the failures, and decide when AI is the wrong solution entirely.
That work is not glamorous. It is also where most of the value is created.
If Sri Lanka wants to become an AI producer rather than just an AI consumer, we should judge ourselves less by how many demos we can show this week and more by how many useful systems are still running a year from now.
AI can summarise a financial report in seconds. Getting hold of information that is difficult to find, commercially valuable and reliable is…
John Keells IT has signed a collaboration agreement with Andersen Consulting, giving the Sri Lankan technology business a further route to international…
An Indian customer finds a Sri Lankan hotel, chooses a room and reaches the payment page. Giving them a familiar way to…
AI announcements are easy to find. Evidence that customers are paying for the technology takes a little more digging. IFS’s latest financial…
MillenniumIT ESP has partnered with Shift Technologies Canada to bring fleet and transportation software to customers in Sri Lanka, the UAE and…
Photo: Adityamadhav83 / Wikimedia Commons (CC BY 4.0)A view of Kambalakonda Wildlife Sanctuary near Visakhapatnam, India, where proposed data-centre developments have raised…
Adding an AI tool to a team is easy. Working out whether it saves time, improves the work, or adds another subscription…
Dialog says its 5G network now spans all 25 districts in Sri Lanka. According to the company, over 1,000 5G towers are…
You tap “pay” in an app. Somewhere behind the screen, several systems need to talk to each other before your money moves.…
Tisankan Jeyakumar is the Chief Technology Officer at Yarl Ventures in Jaffna, Sri Lanka. He works hands-on across software architecture, cloud infrastructure, DevSecOps, AI and product engineering, with a focus on building secure, cost-conscious digital products for real-world operating environments.
A useful perspective, a sharp question or something we missed — bring it to the discussion.
Your email address will not be published. Required fields are marked *
Save my name, email, and website in this browser for the next time I comment.
GIPHY App Key not set. Please check settings
Please explain the reason why this comment should be removed
Tech News, Reviews, Start-ups, Fintech
News, analysis and culture from Sri Lanka’s technology ecosystem.
© 2026 ReadMe. All rights reserved.
Loading comments…