A 90-Day AI Roadmap for SMEs Based on TurkStat 2026 Data

An AI roadmap for SMEs: what TurkStat’s 2026 data says

For small and medium-sized businesses in Türkiye, an AI roadmap is no longer a “someday” item. The Artificial Intelligence Usage in Enterprises and Individuals bulletin published by TurkStat (TÜİK) on 2 October 2026 shows that the share of enterprises with at least 10 employees using at least one AI technology rose from 7.5% to 14.0% in a single year. According to the same bulletin, the share of individuals who say they use generative AI went from 19.2% to 37.6%. In other words, many of your employees are already experimenting with these tools; the real question is whether the company can turn that into a controlled, measurable plan that complies with KVKK, Türkiye’s personal data protection law.

In this article we read the TurkStat figures from an SME perspective and then propose a three-phase plan you can run in 90 days: days 0–30 for discovery (use-case inventory, data inventory, KVKK check), days 31–60 for pilots (document/OCR, a customer support assistant, sales and stock forecasting) and days 61–90 for ERP/CRM integration, measurement and a scale decision. The goal is not a sweeping “digital transformation” programme but a small, reversible way of working with a written decision at the end of every phase.

In short

• TurkStat 2026: AI use reached 14.0% among enterprises with 10+ employees and 12.8% among those with 10–49.
• For enterprises considering AI, the biggest barrier is lack of expertise (72.3%), followed by legal uncertainty and data protection concerns.
• The 90-day plan has three gates: pilot approval, integration approval and a scale decision.
• When the data is ready, document/OCR, a support assistant and sales/stock forecasting give SMEs the fastest feedback.
• Lasting value appears once the model is connected to ERP/CRM through a middleware layer and measured.

What do TurkStat’s 2026 figures mean for SMEs?

The TurkStat bulletin measures two separate groups: individuals aged 16–74 and enterprises with at least 10 employees. On the enterprise side the picture is clear. AI adoption was 2.7% in 2021, reached 7.5% in 2025 and 14.0% in 2026. Growth is fast, but adoption is still low: only about one in seven enterprises uses at least one AI technology.

The size gap: large firms lead, SMEs are speeding up

The breakdown by headcount carries the main message for SMEs. Adoption stands at 37.1% among enterprises with 250 or more employees, 17.0% among those with 50–249 and 12.8% among those with 10–49. In 2021 the same figures were 9.6%, 3.6% and 2.3%. Small businesses grew fastest in relative terms, yet the gap with large companies has not closed. In our experience that gap usually comes less from budget and more from the absence of a plan that decides who tries what, and how it is measured.

Indicator (TurkStat, 2026)	Value	What it means for SMEs
Enterprises using AI (10+ employees)	14.0% (2025: 7.5%)	Competitors have started; there is still an early-learning advantage.
Enterprises with 10–49 employees	12.8%	Works at small scale too; choosing the right pilot is critical.
Use for marketing or sales	51.0%	The most common starting point: content and proposal drafting.
AI users processing personal data	16.6%	A KVKK assessment should come before the pilot.
“Lack of expertise” barrier among those considering AI	72.3%	A case for outside support and a narrow first scope.

What is AI used for, and how is it sourced?

Among enterprises using AI, the most common purpose is marketing or sales (51.0%), followed by R&D or innovation (46.3%) and production or service processes (43.4%). On sourcing, 67.8% use open-source software and 49.1% use closed-source software. 46.5% use systems developed by external providers and 32.1% use systems developed by their own staff. This suggests most SMEs combine ready-made models with outside expertise rather than training models from scratch, and your roadmap should rest on that same realistic foundation.

Barriers: expertise, legal questions and data protection

8.3% of enterprises that do not use AI say they are considering it. In this group the leading barriers are lack of expertise (72.3%), legal uncertainty about liability (66.4%) and data protection and privacy concerns (65.4%). All three are governance problems more than technical ones. That is why a good roadmap does not start with “which model should we pick?” but with “which data will we process, with whose approval and with what record?”

• • •

Why build the roadmap as 90 days and three gates?

Ninety days is long enough for an SME to see results on real data without losing focus. A shorter plan often ends as a demo; a longer one delays the first tangible output and the team loses interest. Putting a gate at the end of every 30 days splits the investment into small slices: at each gate you make a written decision to continue, narrow the scope or stop. It is the discipline of planning a software project, adapted to AI.

The plan needs an owner, usually a business unit manager paired with a technical lead. At board level, only gate decisions and measurement results are reported. That way AI stops being a side project of the IT department and gets tied to a business goal.

Days 0–30: what happens during discovery?

The first month may pass without running a single model. That is not wasted time; it is the preparation that determines how productive the next two months will be. Discovery consists of three inventories.

Use-case inventory

Ask each department to list repetitive, time-consuming tasks: keying incoming invoices into the system, answering the same customer questions, calculating the weekly stock order, drafting proposal texts and so on. For each item, record frequency, time spent, cost of errors and the owner. Then score the use cases on two axes: business value and feasibility. The one or two in the top-right corner are your pilot candidates.

Data inventory

Work out which data the chosen use case needs: document scans, email archives, ERP sales history, CRM records, support tickets. For each source, note where it is stored, who can access it, how far back it goes and how good it is (missing fields, duplicates). If the data does not exist or is scattered, the pilot’s first job will be collecting it, and knowing that up front keeps the schedule realistic. A structure such as an enterprise data analytics and management reporting platform can bring scattered sources onto one measurement baseline.

KVKK check

According to TurkStat, 16.6% of enterprises using AI process personal data in those systems. If the pilot touches personal data, answer these questions before you start: Which personal data will be processed? Are the legal basis and privacy notice in place? Will data be transferred abroad? Does the model provider store requests or use them for training? Is masking or anonymisation possible? Work through these with your legal adviser; this article is not legal advice. For businesses that need to keep data in the country, evaluating domestic model services where data is processed on infrastructure in Türkiye, such as the EVREN API, has also been an option since 2026.

Gate 1 output: a one-page pilot definition covering the use case, data source, KVKK note, success metric (for example processing time per document), owner and budget ceiling.

Days 31–60: which pilots make sense for an SME?

The pilot phase does not answer “does AI work?” but “does it make a measurable difference in this workflow?” Pilots run on real data with a limited scope and a human approval step. Three candidates tend to work well for SMEs.

Document and OCR data extraction

In many SMEs, supplier invoices, delivery notes, contracts and application forms are still read and keyed in by hand. Combining OCR with language models lets you extract fields such as date, amount, tax number and line items and drop them into a review screen where an employee only checks and approves. Measurement is simple: time per document and correction rate. Make sure low-confidence fields are flagged automatically and nothing reaches accounting without approval.

Customer support assistant

For repetitive requests such as FAQs, order status and return conditions, an assistant grounded in company documents can help the support team draft replies faster. In the first phase the assistant should not answer customers directly; it should suggest replies to agents. If the knowledge base is outdated, the assistant will be wrong too, so half of this pilot is really about documentation. Measure first-response time, how often agents use the suggestion unchanged and customer satisfaction.

Sales and stock forecasting

For businesses with a few years of sales history in the ERP, demand forecasting by product or product group is a way to cut the cost of overstock and stock-outs. Classic time-series and machine learning methods often do more here than a large language model; the language model can turn results into a summary a manager can read. Measure forecast error, stock turnover and lost sales. Adding campaigns, seasonality and supplier delays to the data directly affects quality.

Use cases by department

The table below is a starting list for the discovery phase. Risk and effort are general guidance; re-score them against your own data.

Department	Use case	Data needed	Risk	Effort
Accounting / finance	Field extraction from invoices and delivery notes (OCR)	Document scans, account records	Medium: wrong amounts, human approval required	Low–medium
Customer service	Support assistant that drafts replies	FAQ, returns policy, past tickets	Medium: personal data and misinformation	Medium
Sales / marketing	Proposal and product copy drafts	Product catalogue, price list	Low: internally approved before publishing	Low
Warehouse / procurement	Sales and stock forecasting	ERP sales history, stock movements	Low–medium: cost of a wrong order	Medium–high
Human resources	Internal policy and procedure search assistant	Regulations, handbook	Medium: keep employee data out of scope	Low–medium
Management	Weekly report summaries	Reporting platform outputs	Low: protect access to source reports	Low

Gate 2 output: a pilot measurement report covering the baseline, the post-pilot value, error examples, user feedback and an estimated scope for integration.

Days 61–90: how do you integrate with ERP/CRM and decide whether to scale?

A pilot can succeed on a separate screen, but if staff copy and paste between two systems every day, the gain soon disappears. The third month is about connecting the model to the existing workflow.

Connecting through middleware

Instead of wiring applications straight to a model provider, putting a layer in between is cheaper in the long run, even for SMEs. This layer connects securely to ERP and CRM, can mask personal data, decides which request goes to which model, logs cost and latency, and switches to a fallback when a provider fails. We cover the architecture in detail in taking AI agents to production and the AI Ops layer, LLM router and KVKK. The layer itself is usually a custom software development job and should be designed to fit your ERP and CRM.

Measurement and the decision

Once the integration is live, measure for at least two to three weeks. Calculate the success metric defined in the pilot the same way in the integrated flow; otherwise the comparison is meaningless. On cost, look at model usage fees, infrastructure and maintenance effort together. The third gate has three options: scale (extend to another department or site), fix (solve the data or workflow issue and measure again) or stop (record what you learned and move to the next use case). Stopping is not failure; learning in ninety days and on a small budget that “this doesn’t work for us” is valuable too.

Discovery: use cases and weekly tracking

Pilot: tracking the success metric

Scaling: rolling out to daily operations

AI roadmap checklist

Before each gate, make sure the following items are ticked:

• Every department has listed repetitive tasks and scored them for value and feasibility.
• The pilot’s data source, access rights and data quality have been confirmed in writing.
• A KVKK assessment has been done with a legal adviser; transfers abroad and retention are clear.
• The success metric and its baseline were measured before the pilot.
• Human approval and error reporting are defined in the pilot workflow.
• Model access goes through a logging, masking middleware layer rather than straight from applications.
• Cost (usage, infrastructure, maintenance) is reported monthly.
• At every gate a written continue, fix or stop decision has been made and shared.

How can Aksiyon Soft help?

Aksiyon Soft is headquartered in Samsun and works remotely with businesses across Türkiye, with planned on-site visits for discovery or go-live when needed. In AI work we follow the same rhythm we use for enterprise software solutions:

• Discovery: use-case and data inventory, a KVKK question list and a measurable success metric.
• MVP: one pilot going live on real data, with human approval and a limited scope.
• Sprint demos: working software and measurement results shared every two weeks.
• Hypercare and SLA: close monitoring after integration, then maintenance under a written service level.

Our aim is not to lock you into a model or provider, but to build a structure where switching models is easy thanks to the middleware layer, and where measurement and records stay with you.

Frequently asked questions

Does a small business really need an AI roadmap?

Yes, but it should be short. TurkStat data shows adoption reached 12.8% among enterprises with 10–49 employees. Unplanned experimenting raises the risk of staff moving company data to outside services through personal accounts. A ninety-day, three-gate plan keeps that risk manageable.

Which area should we choose for the first pilot?

Pick a task where the data is ready, the work repeats often and the result is easy to measure. Document/OCR extraction, reply drafting for the support team and sales/stock forecasting are common choices for SMEs. Starting with a use case that involves little personal data also lightens the KVKK workload.

Do we need to train our own model?

In most cases, no. TurkStat data also shows that most enterprises use open-source software or systems from external providers. For an SME the real work is connecting ready-made models to the right data, workflow and measurement.

What is the most critical point under KVKK?

Knowing which personal data goes to which provider and which country. Masking and logging in the middleware layer, preferring services that process data in Türkiye and updating privacy notices are the basic steps. Work with your legal adviser for the final assessment.

What happens if it fails at the end of ninety days?

The plan is built for that possibility. Investment at each gate is limited; a stop decision is recorded with the lessons learned and you move on to the next use case. Outputs such as the data inventory and the middleware layer are reused in the next attempt.

Can this process be run remotely?

Yes. Discovery sessions, sprint demos and measurement reports can all be handled remotely. Where on-site observation is needed, a visit can be planned around the project schedule.

Sources

• TurkStat — Artificial Intelligence Usage in Enterprises and Individuals, 2026 (2 Oct 2026)
• Anadolu Agency — Share of people in Türkiye using generative AI doubles to 37.6% (in Turkish, 2 Oct 2026)
• Anadolu Agency — Defence industry’s national AI platform EVREN opens for use (in Turkish, 27 Sep 2026)

Let’s talk about your project

If you would like to build your ninety-day AI roadmap around your own data and processes, get in touch for a discovery call. In the first conversation we can review your pilot candidates and data situation together.