Most conversations about AI start with a tool name: which model, which platform, which subscription. In a small or mid-sized business, though, the outcome is decided by something else — which process you start with. A project that starts with the wrong process turns into a dashboard nobody opens, no matter how good the tool was.
This is not a product pitch. It sets out where Turkey actually stands using official figures, explains what cost really depends on, and proposes a concrete 90-day starting plan. Wherever we give a number we give its source; where we can't, we say why.
The figures first: where does Turkey actually stand?
According to TurkStat's Artificial Intelligence Statistics (2025) — published under its own heading for the first time — 7.5% of enterprises with ten or more employees use some form of AI technology. The same figure was 2.7% in 2021, roughly a threefold increase in four years.
The difference by size is striking: 6.6% among businesses with 10-49 employees, 9.6% among those with 50-249, and 24.1% among those with 250 or more. On the surface, AI still looks like a large-company game.
One scope note deserves emphasis: this TurkStat survey covers only enterprises with ten or more employees. Micro businesses — those with 0-9 employees, which make up the overwhelming majority of firms in Turkey — are not in the data at all. So it would be wrong to say "7.5% of businesses in Turkey"; with micro businesses included, the real rate is very likely lower.
The figure that completes the picture: TurkStat's SME Statistics (2024) counts 3,928,000 SMEs in Turkey, making up 99.6% of all enterprises, 68.5% of employment and 44.1% of turnover. In other words, the main body of the economy sits exactly where AI adoption is lowest.
Read that as timing information rather than as a shortcoming. In a market where adoption is low, a business that sets things up properly moves before its competitors do.
The wrong question: "which AI should we buy?"
The right question is: which of our tasks produces a repetitive, measurable load?
AI does not perform well on loosely defined work. It performs well on work that repeats dozens of times a day, whose rules can be written down and whose result can be counted: sorting incoming messages, booking appointments, requesting documents, answering the same fifteen questions, extracting data from a file.
So the first decision is not a software decision but a process decision. You choose the tool afterwards.
Three tests for choosing the process
- Frequency: does it repeat at least a few times a day, preferably dozens? Automating something done three times a month is a change nobody feels.
- Measurability: can you state today's situation as a number? "We reply to messages in four hours on average", "20% of appointments are no-shows", "we spend 25 minutes collecting documents per policy". If you can't measure a process, you can't prove it improved either.
- Cost of error: what do you lose when it goes wrong? A wrong appointment time can be corrected; a wrong price quote cannot. Start with a process where errors are cheap, and build trust from there.
The places where all three usually hold: answering inbound customer messages, appointment and reminder flows, responses to frequently asked questions, and repetitive document collection. These are also the areas with the lowest cost of error — because you can keep a human approval step at every stage.
What does the cost actually depend on?
Three lines determine the budget of an AI project, and most proposals discuss only the first.
The first is setup: mapping the process, connecting systems, the operator panel, testing. This is one-off, and most of the real work sits here.
The second is usage: running the model and the messaging infrastructure. Here your cost depends less on message volume than on message mix. On WhatsApp specifically the logic is this: within the twenty-four-hour window that opens when a customer writes to you, your free-text replies and utility templates are free — while marketing templates are charged even inside that window. So the bill of a business that answers incoming questions looks nothing like the bill of a business that sends bulk campaigns every day.
For users arriving from an ad or a page button the window extends to seventy-two hours, and messages are free throughout — which is why turning ad-driven interest into a conversation is comparatively cheap.
We deliberately avoid quoting per-message figures here. Meta updates these rates by country and at regular intervals; a fixed number would be wrong within months. Check the current rates on Meta's official WhatsApp Business Platform pricing page before deciding. Figures repeated by third-party blogs are often stale and contradict one another — we saw exactly that while verifying this article.
The third line never appears on an invoice: internal time. If it isn't clear who owns the process, who decides the exceptions and who watches the panel, the project dies even when it works technically.
What happens in 90 days?
The plan below is for a single process. Starting three at once is the most reliable way to leave all three unfinished.
Days 0-30 — measure and narrow
- Write down today's situation as a number: volume, average duration, error or loss rate. That number will later be the only honest answer to "did it work?".
- Narrow the scope: one channel, one language, one topic. Widening is easy; pulling a sprawling scope back together is not.
- Define the handover points up front: in which situations does the AI stop and pass the work to a person? Complaints, returns, price negotiations and moments of crisis need that threshold in writing.
- Make the data and privacy decisions here: what is stored where, for how long, and where the privacy notice is shown.
Days 30-60 — run it in a narrow scope
- Go live with real traffic but limited scope. Synthetic tests cannot imitate a real customer's typos, half-finished sentences and impatience.
- Make every reply auditable: what was said should be visible in the operator panel and reversible when needed.
- Watch only two things weekly: the share of cases the AI closes on its own, and the reasons behind the ones handed to a person. That second list is the agenda for your next improvement.
- Do not add features at this stage. Adding them breaks the measurement.
Days 60-90 — widen and hand over
- Compare the first measurement with today's number. If there is no difference, do not widen the scope — change the process or the thresholds instead.
- Add a second channel or a second topic — not both.
- Complete the handover of ownership: who owns the panel, who decides exceptions, and what the weekly review routine is, all in writing.
- Only then move on to a second process.
When does it pay back?
The honest answer: it depends on the process you chose, and you should be sceptical of anyone who gives you a single number.
A framework is still possible. For response time, missed enquiries and repetitive document work, the effect becomes measurable within the first months — because today's cost of those is already countable. For headings such as "brand perception" or "customer satisfaction", you measure the effect indirectly and over months.
The practical rule: don't start with a process whose payback you cannot measure in the first three months. An improvement you cannot measure cannot be defended in a budget discussion.
Five common mistakes
- Choosing the tool before the process. Decide what gets automated first; pick the tool second.
- Opening too many fronts at once. Most teams that start three processes together finish none of them.
- Not writing down the handover threshold. An AI saying "I don't know, let me pass you to a colleague" always beats an AI inventing an answer.
- Skipping the baseline measurement. If the before is unknown, the after cannot be proven.
- Mistaking a pilot for a permanent system. A narrow pilot is not permanent until it has a widening plan and a named owner.
The compliance calendar: what's happening with e-invoicing?
Digitalisation doesn't always start with efficiency; sometimes regulation starts it. Under the current version of the Turkish Revenue Administration's General Communiqué on Tax Procedure Law No. 509, taxpayers with gross sales revenue of 3 million TRY or more in the 2022 or any subsequent accounting period must move to e-Invoice. For those selling goods or services online, and for intermediary service providers, the threshold is 500,000 TRY.
The timing works like this: the start of the seventh month of the year following the accounting period in which the threshold was exceeded. For a business whose 2025 turnover exceeds those thresholds, that date is 1 July 2026.
There is also a recent deferral: General Communiqué No. 589 moved the "e-Archive Invoice regardless of amount" obligation to 1 January 2027 for taxpayers assessed on a simplified basis and those keeping books on a business-account basis. For small taxpayers that means an extra year of preparation.
We include this calendar for a reason: moving to e-invoicing already opens a software project for most businesses. Reviewing your processes while that project is open costs less than doing two separate jobs at two separate times.
In short
- Start with a process, not a tool; pick one that repeats, can be measured, and is cheap to get wrong.
- Write today's situation as a number before you start.
- One process, narrow scope, a written human-handover threshold.
- Count internal time alongside setup and usage when you budget.
- Measure at ninety days; if there's no difference, fix it rather than widen it.