Tecof • September 9, 2026
AI in Turkey: Progress, Policy and 2030 Vision
In Brief
Turkey has moved past the pilot stage on artificial intelligence. Autonomous platforms fly operational missions, banks catch card fraud in milliseconds, factories replace bearings before they fail, and fertiliser now lands only where a field actually needs it. The national strategy anchors on lifting the technology's contribution to GDP toward 5 percent. For smaller companies the shift starts quieter: models that draft product copy, agents that answer overnight questions, dashboards that predict stock and returns. The obstacles are equally clear, from brain drain and chip dependency to energy bills and legal uncertainty.
A card is used in Ankara at 03.12 and again in Frankfurt at 03.14. The second transaction never reaches a human approver, because an anomaly model made the call. That same week a grower on the Konya plain reads a multispectral drone image and fertilises 40 of his 300 donums. The two look unrelated. They run on the same plumbing.
For years the national conversation stayed in the future tense. By September 2026 it is concrete: uncrewed aircraft fly real missions, e-Nabiz health data turns into epidemiological forecasts, cameras handle quality control in Bursa and Kocaeli. The open question is how fast this spreads, and who benefits.
Three inputs set that pace: data, trained specialists, and the bill for compute. Defence and banking reach all three, while a twenty person brand usually has only the first.
Policy, national strategy and regulation
How far a technology spreads tends to track the framework the state draws around it. In Turkey that centres on the Digital Transformation Office of the Presidency, which connects siloed public databases and publishes anonymised datasets for researchers and founders. Public data is the best raw material for training local models.
The number the strategy hangs on
Action plans revolve around one figure: raising the contribution of artificial intelligence to gross domestic product to around 5 percent. Universities have opened AI engineering departments, training for data scientists has accelerated, and TUBITAK and KOSGEB channel grants into early stage ventures.
Regulatory sandboxes
Tightly supervised sectors need legal room to test an algorithm before production. A sandbox provides that: limited users, a limited window, real data under supervision. Trying a credit scoring model there protects the institution and the consumer at once.
What KVKK and IYS change in practice
Turkey's personal data law, KVKK, constrains what you may feed a model. Before pushing customer conversations into training data, check that your privacy notice covers that purpose. Marketing is the same: no SMS goes to a number without consent in IYS, even when a model wrote the copy. Servers abroad add transfer rules.
| Body | What it does | Why a business cares |
|---|---|---|
| Digital Transformation Office | Policy coordination, open data | Free datasets, standards for public projects |
| TUBITAK and KOSGEB | R and D grants, startup support | Part of development cost covered |
| KVKK authority | Personal data supervision | Notices, retention limits, transfer rules |
| IYS | Commercial message consent registry | Consent required before SMS and email |
| Technoparks | Tax exemptions, incubation | Lower payroll cost for R and D staff |
Defence, the most visible showcase
Defence is where the country turns research into fielded product fastest. A sector once defined by mechanical advantage now runs on software and sensor fusion, and a platform's value sits in its code rather than its payload weight.
Intelligence at the edge
Bayraktar TB2, Akinci, Anka and Aksungur are not just airframes. Video from electro optical cameras and radar returns is processed in the air rather than shipped to a ground station. That is edge AI, and tracking continues through a lost datalink because the model runs on board. Separating civilian from military objects finishes in well under a second.
Swarm behaviour
Dozens of small aircraft can operate together without a central controller thanks to swarm algorithms. Each vehicle sets its route relative to its neighbours, targets are divided among the group, and losing one aircraft triggers an immediate redistribution of tasks. The point of that architecture is that no single strike stops the system. The civilian analogue is a drone fleet running coordinated search patterns over a disaster zone.
Loyal wingman and autonomy at sea
Uncrewed combat aircraft such as Kizilelma and Anka-3 are designed to fly formation with piloted jets. The pilot sets intent, the aircraft computes the manoeuvre. At sea, vessels like ULAQ and Marlin navigate autonomously in rough conditions and classify contacts on their own, while ground vehicles handle reconnaissance and mine clearance with nobody aboard. Computer vision that matures under defence budgets later shows up in industry and hospitals, carried by the same engineers and libraries.
Banking, payments and fintech
Turkish banking sits well above the global average on digital infrastructure, and transaction volume per second passed what rule based legacy systems could police years ago.
Anomaly detection
Machine learning models build a behavioural profile per customer: usual hours, usual devices, usual ticket sizes. Movement outside it gets flagged and frozen, and the models are retrained as fraud tactics shift. On the commerce side the same logic drives return abuse scoring and fake order detection. The hard part is the false positive rate, since a model tuned too tightly blocks genuine shoppers and floods the call centre.
Alternative credit scoring
Traditional scoring reads banking history, which a recent graduate or a six month old business does not have. Local fintechs build a trust score from alternative signals such as utility payment patterns and e-commerce spending. In practice that pulls people outside the banking system into credit and instalment options, and on the corporate side supplier payment history feeds the same term risk model.
Assistants and algorithmic trading
Most first line bank support now runs through assistants built on Turkish natural language processing, which analyse spending and suggest actions rather than only parsing commands. On Borsa Istanbul, institutional desks run algorithmic systems that read news flow and technical signals faster than any human. Payments show similar maturity: local providers decide with a risk model whether a card needs a 3D Secure step, and removing an unnecessary verification screen visibly reduces cart abandonment. We covered that mechanic in more depth in our guide to raising ecommerce conversion rates.
Factory floor, field and hospital
The most measurable gains sit in manufacturing, agriculture and health. The technology rarely invents a new job there, it lowers the cost of an existing one.
Predictive maintenance and machine vision
Nothing costs a plant more than an unplanned line stop. Sensors on motors and presses stream vibration and acoustic data, and a model hears a bearing failing two weeks out, moving the repair into a scheduled shutdown. On automotive lines quality control has shifted to cameras, where sub millimetre scratches and bad welds get flagged instantly.
Precision agriculture
Satellite and drone imagery is processed by the light wavelengths a plant reflects, showing which part of a field is under water stress or short on nitrogen down to centimetres. That feeds variable rate application, where the tractor treats only the flagged patches. Input costs fall and the soil absorbs fewer chemicals, while yield forecasting helps the grower decide when to sell. Irrigation follows the same arithmetic, with moisture sensors opening water by need rather than by clock.
Radiology, e-Nabiz and drug discovery
Medical imaging is exposed to fatigue driven human error. Software built by Turkish engineers alongside clinicians marks lung nodules and haemorrhage risk at pixel level, offering a second read rather than replacing the radiologist. De-identified e-Nabiz data shows weeks ahead how a seasonal outbreak spreads, which saves money on drug procurement and bed planning. In biotech labs molecular reactions are simulated in silico, compressing years of trial and error into months. Approval as a medical device is a separate hurdle that absorbs much of the calendar.
Turkish language models and data sovereignty
On the generative side the test facing any country is cultural as much as technical. Widely used commercial models were trained mostly on English corpora, so Turkish morphology, idiom and legal terminology ride in the back seat.
Why Turkish is its own problem
A word carrying five suffixes gets shredded into far more pieces than it should by tokenisers designed for English. That raises both cost and error rate, so processing the same paragraph is simply more expensive in Turkish. Teams around TUBITAK BILGEM, several universities and the T3 Foundation keep building Turkish focused models for that reason. A vocabulary built for Turkish covers the same sentence in fewer tokens, so it runs cheaper and more consistently.
Where it actually pays off
Turkish first models need fewer corrections on work ranging from case law search to summarising patient histories. The commercial version is plainer: tagging reviews by sentiment, sorting return requests by reason, mapping product titles to marketplace categories. A few points of accuracy there becomes real hours saved for a team processing hundreds of records a day.
Where the data is processed
Sending a hospital record or a bank transcript to a server abroad is a legal decision, not a technical preference, which is why open source models on local infrastructure spread quickly. Smaller merchants face the same question, since an audit asks first which country processes your customer addresses and order history. Read the processing terms as carefully as the price list. Our walkthrough on building an AI powered store covers those checks step by step.
What this means for small businesses
Most of the above belongs to organisations with large budgets, yet most companies in Turkey employ fewer than twenty people. For them the work is not training a model but wiring an existing one into a workflow. The two approaches differ in cost by two orders of magnitude.
Where to start
- Drafting product and category copy, then an editor review
- An agent giving the first answer to overnight and weekend questions
- Mapping titles and attributes for Trendyol and Hepsiburada listings
- Flagging high return risk orders before the courier collects
- Generating ad copy variants and testing them
A worked example
The table models a brand with 400 products and 1,200 orders a month. Swap in your own figures, the logic holds.
| Task | Old way | With a model | Monthly delta |
|---|---|---|---|
| 400 product descriptions | Agency, 12,000 TL | 900 TL usage, 4 editor hours | Around 9,000 TL saved |
| 60 overnight questions | Answered next morning | Instant first reply | Fewer abandoned carts |
| Ad copy testing | 2 variants a month | 12 variants a month | Faster learning cycle |
| Return risk flagging | None | Confirmation call on risky orders | Lower shipping and return cost |
Marketplace versus your own store
On a marketplace the algorithm belongs to the platform, whose model ranks your title and images. On your own domain you own search, recommendation and pricing logic, and the balance comes down to margin and repeat purchase rate. What makes product pages readable to language models sits in our AEO and GEO guide. For orders and stock, Tecof AI agents ship with ready workflows, while the small business package keeps accounting and courier integrations in one panel.
Obstacles and the road to 2030
However ambitious the vision, four obstacles stay put, and ignoring them makes any target impossible to assess.
Brain drain
Machine learning architects and data engineers are among the most contested profiles worldwide. As remote work normalised, companies abroad began hiring engineers who never leave Istanbul, and technopark tax relief alone does not close that gap. Deeper venture funding and a simpler path to founding a company keep specialists here.
Chips and energy
Training models demands serious compute, and that hardware is imported, so the bill arrives in foreign currency however strong the software skills. Data centre electricity is a second line item, which makes renewable supply a competitive question rather than only a cost question. Cloud providers opening a region inside Turkey cuts latency and price at once.
Legal responsibility
When an autonomous vehicle crashes or a diagnostic aid points the wrong way, liability is still contested. Does it sit with the team that wrote the software, the manufacturer, or the operator? Rules guaranteeing auditability of algorithmic decisions, plus courts with the expertise to hear such cases, will close that gap.
What to do tomorrow morning
Rather than watching the national picture, start inside your own operation. Tomorrow morning, export the last 90 days of customer questions from your admin panel, put the 20 most repeated ones in a sheet, and count how many concern stock, shipping or returns. Two thirds usually collapse into five topics. Wiring an agent that answers those five is a day of work, and the first measurable gain lives there. For the step after that, use the checklist in our piece on next generation ecommerce and automation.
Frequently Asked Questions
Which sectors in Turkey use AI most?
Defence, banking and manufacturing are the most mature: autonomous platforms, fraud detection and predictive maintenance respectively. Agriculture and healthcare are converging fast, while retail mostly buys ready made services.
What does the national AI strategy target?
The headline anchor is raising the technology's contribution to gross domestic product toward 5 percent. Training specialists, opening public data and funding startups are the pillars, coordinated by the Digital Transformation Office of the Presidency.
How does KVKK limit AI use?
KVKK requires you to declare why you process personal data and stay inside that purpose. Using customer conversations for training is a new purpose and must appear in your privacy notice. Servers abroad bring transfer rules as well.
What budget should a small brand set aside?
Model usage starts as a monthly line item of a few thousand lira. The real cost is human: the editor or operations lead reviewing the output. Starting with one workflow and measuring it beats a large upfront commitment.
Why does Turkish need its own language model?
Turkish agglutination works badly with tokenisers built for English, splitting text into more pieces and raising cost and error rate. Legal and medical terminology needs local context, and data sovereignty is a separate reason.
How is AI affecting employment?
Repetitive data entry and first line support are shrinking, while data preparation, model supervision and process design roles grow in their place. The safest position now is mid level expertise combined with fluency in the tools.
Is independence possible without domestic chips?
Software skill alone is not enough, since training depends on imported hardware. The near term answer is shared compute and smaller purpose built models. Domestic chip design stays a long horizon item for defence and critical infrastructure.
Who is liable when an AI system gets it wrong?
Current legislation gives no clean answer, so practice is shaped case by case. Liability in most disputes lands on the institution operating the system. Keeping a human on the final approval and logging the reasoning has become standard practice.
Does AI generated content hurt search rankings?
The risk lies in output quality, not in how the text was produced. Unreviewed, near duplicate pages get penalised, while content an editor has checked and enriched with your own data performs normally. On product pages the test is whether the copy answers the shopper's question.