AI Implementation Guide for SMEs: From Process Inventory to a 90-Day Roadmap

Who is this guide for?

This article is written for businesses with roughly 3 to 100 employees that run their daily work on WhatsApp, email, Excel and an accounting or industry-specific application. You might be an auto repair shop, a car dealership, a rental company, a clinic, an accounting firm, a boutique manufacturer, an online store or a service company. What you have in common: plenty of work, few people, and the same tasks repeating every day.

We will not spend long defining "what AI is." Instead, we will answer three questions:

1. Which tasks in my business can be handed over to AI, and which should not be?
2. At what level should I do it: with an off-the-shelf tool, a connected automation, or custom software?
3. How do I manage the risks (wrong information, data breaches, cost control), and how do I measure success?

Positioning AI correctly: the "smart intern" model

The most useful mental model for businesses is this: today's AI tools are like an intern who reads and writes extremely fast and never gets tired, but does not know your business or your rules. Give it a clear task, the information it needs and the rules it must follow, and it will turn work that takes minutes into a draft in seconds. Send its output to a customer without checking, and it may occasionally state something wrong with complete confidence.

Three practical rules follow from this model:

• AI drafts, a human decides. Every text that goes to a customer and involves pricing, a quote, health or legal content passes through a person's approval.
• You provide the information. AI cannot know your price list, stock levels or contract terms on its own. For correct answers, that information has to be supplied to it.
• The clearer the task, the better the result. Instead of "write something," you need to say "write a reply draft for this customer, using this information, in this tone, in no more than 5 sentences."

Where AI is strong, moderate and weak in an SME

The key point in this table: AI's greatest strength is language and document work. Calculating figures and making definitive decisions are still the job of conventional software and people. The best results come from systems that use both together.

Step 1: Build a one-week process inventory

Most businesses start with AI by asking "which tool should we buy?" The right question is "which of our tasks costs us the most time?" To answer it, keep a small inventory for one week.

Open a simple spreadsheet (Excel or Google Sheets is enough) and ask everyone on the team to list the tasks they notice themselves repeating. For each row, record:

• Task name: "Answering price questions that come in via WhatsApp"
• Who does it: Sales consultant
• Times per week: 120
• Minutes per occurrence: 3
• Where the input comes from: WhatsApp message
• Where the output goes: Reply to the customer, sometimes a quote form

After one week, you will typically have 15 to 40 tasks on the list. What surprises people is how large the total time is: a 3-minute task done 120 times a week adds up to 6 hours a week, or roughly one and a half working days a month.

Step 2: Score the tasks on five criteria

Score each task in the inventory from 1 to 5 on five criteria:

The total score ranges from 5 to 25. Tasks scoring 18 or above are strong candidates for the first pilot; tasks scoring 12–17 wait for the second wave; tasks below 12 are either not suitable for AI or need their data and process organized first.

Example: An auto repair shop's scorecard

The table below is a sample assessment we prepared for a 6-person auto repair shop (the figures are illustrative):

The takeaway from this table is instructive: although issuing invoices scores high, it is a calculation and compliance task; it belongs to accounting and service-management software, not to AI. Fault diagnosis requires expertise and carries a high cost of error; at most, AI can give the mechanic a summary of "what was done in similar cases." The right pilot candidates are vehicle status messages and the post-service work summary.

Step 3: Choose the right level

The same task can be solved at three different levels. The volume of work, the sensitivity of the data and the budget determine the level.

Level 1: Off-the-shelf tool + written rules

The business (enterprise) edition of a tool such as ChatGPT, Gemini, Claude or Microsoft Copilot is rolled out to the team. Employees paste in the text themselves, get a draft, edit it and use it.

• When it fits: Drafting, summarizing, translation, brainstorming; a few dozen tasks a day.
• Cost item: A monthly per-user subscription.
• Advantage: You can start in a day, with no technical setup.
• Limitation: Every task requires manual copy-and-paste; there is no connection to your own systems; if entering personal data into the tool is not controlled, risk arises.

Level 2: Connected automation

AI is connected to your existing tools through an automation platform (such as Make, Zapier or n8n) or a small integration. For example: a request submitted through your web form is automatically classified, summarized, written to Google Sheets, and the right person is notified.

• When it fits: When the same task is repeated dozens of times a day and the input arrives in digital form.
• Cost items: Automation platform subscription, AI API usage (per transaction), setup effort.
• Advantage: Copy-and-paste disappears; work keeps flowing overnight.
• Limitation: As workflows multiply, they become harder to manage; if error handling and logging are not designed well, you end up with automations that break silently.

Level 3: Custom integration

AI is embedded in the business's own software (CRM, ERP, service or booking system). Assistants that answer from your own documents and cite their sources, bots that collect requests via WhatsApp and create records in your system, and screens that automatically extract data from documents all belong to this level.

• When it fits: The work is business-critical, volume is high, the data is sensitive, or multiple systems need to work together.
• Cost items: Development, hosting, model usage, maintenance.
• Advantage: Permissions, logging, approval mechanisms and data security are built according to your rules; the team works in the screens it already knows without having to learn a new tool.
• Limitation: Requires a longer setup period and the right implementation partner.

Practical advice: build the habit with Level 1, measure with Level 2, and move proven, critical tasks to Level 3. Starting directly at Level 3 only makes sense if the process is already well defined and the volume is already high.

Implementation scenarios by industry

The scenarios below are drawn from the needs we encounter most often in conversations with businesses across different industries. For each one, we describe the input, what the AI does and what the human checks.

Auto repair and spare parts

• Vehicle status updates: Based on the work order status in the service record, a message draft is prepared for the customer, such as "your brake pads have been replaced, a test drive is under way, estimated pickup 4:00 PM." The mechanic or service advisor approves it with a single click.
• Work summary: The notes a technician enters in shorthand ("frt pads repl., disc measure ok, oil+fltr") are turned into language the customer can understand.
• Parts matching support: From a customer message like "left headlight lens for my 2016 [car model]," the make, model, year and part name are extracted and prepared for a stock lookup; the part number is verified in the inventory system.

In service-management software such as Garaj Defteri, features like these work most efficiently at Level 3, because the work order data is already in the system.

Car dealerships and used vehicles

• Listing copy: When the vehicle details (model, year, mileage, equipment, inspection notes) are entered, descriptions of suitable length are generated for different listing sites. Damage and replaced-part information is never left to the AI's interpretation; it is taken verbatim from the inspection report.
• Sorting incoming inquiries: Messages such as "What's your best price?", "Do you take trade-ins?" or "Can I get financing?" are classified, and serious buyers are prioritized.
• Pricing research support: Listing data for similar vehicles is collected into a table and summarized; the pricing decision stays with the dealer.

Crane, heavy equipment and equipment rental

• Quote preparation: From a customer message like "Thursday, 2 tons of material to the 8th floor, the site is on a narrow street," the job type, date, load, height and site conditions are extracted, and the quote form is pre-filled.
• Reading site documents: Dates, times and signatures are transferred from photos of timesheets, job completion forms and delivery notes into a table; discrepancies are flagged.
• Summary for operator and vehicle planning: The day's jobs, together with vehicle and operator availability, are summarized in a morning report.

We saw the needs of this industry up close in the Vinç Sistem project: the real gain comes from turning scattered messages into a structured job record.

Clinics, dietitians and healthcare services

• Appointment and reminder communication: Drafts are prepared for messages that create, remind about and reschedule appointments.
• Organizing client notes: Scattered notes taken after a consultation are fitted into a standard template; the specialist reviews them.
• Informational content: Brochures and social media posts with general nutrition or care advice are prepared.

Important: Under KVKK (Turkey's Personal Data Protection Law, Law No. 6698), health data is a special category of personal data. Diagnosis, treatment recommendations and client-specific assessments must not be left to AI, and client data should only be processed on business-grade infrastructure where the contract, the processing location and the security measures are clearly defined.

Accounting firms and office services

• Document sorting: Images of invoices, receipts and bank slips sent by clients are classified by type, and date and amount information is extracted; the accounting entry is made by a professional.
• Draft answers to client questions: For frequently asked questions such as "When is withholding tax on rent due?", answer drafts are generated from information notes approved by the firm.
• Regulatory change summaries: Text taken from an official source is summarized and turned into a "how this affects you" note for each client group. The source text is always kept with its link.

Boutique manufacturing, food and retail

• Turning order messages into forms: From a message like "Saturday, for 15 people, chocolate, with a name written on top," the order form is filled in.
• Product and menu descriptions: Product descriptions are written from the ingredient list, with allergen information checked separately.
• Social media calendar: A one-month posting plan and draft copy are prepared; visuals are matched with the business's own photos.

E-commerce

• Product content: Original product descriptions, FAQs and category copy are generated from supplier technical data.
• Return and shipping questions: An assistant connected to order status answers "Where is my order?" from the shipping record.
• Review analysis: Recurring complaint and praise themes are extracted from hundreds of customer reviews.

Prompt templates you can copy and use

The secret to getting good results from AI is a well-written prompt. We use this five-part structure: Role, Context, Task, Rules, Output format. You can use the templates below by filling in the bracketed placeholders. Enter anything containing personal data ([customer name], phone number, etc.) only in a business account with clear data processing terms; anonymize it where possible.

1. Draft reply to a customer message

Role: You are a customer representative for [business name].
Context: Our services: [short list]. Business hours: [hours]. Pricing information: [approved price ranges].
Task: Write a draft reply to the customer message below.
Rules: Do not commit to an exact price; give a range. If you don't know something, say "our team will get back to you." Use a friendly but professional tone. Maximum 5 sentences.
Output: The reply text only.
Customer message: [message]

2. Translating technical notes into customer language

Role: You are an experienced service advisor.
Task: Summarize the technical notes below in language a customer with no technical knowledge can understand.
Rules: List the work performed as bullet points. Do not add any work that is not in the notes. Do not comment on warranty or price.
Output: Two short lists headed "Work performed" and "Recommended checks."
Notes: [technician notes]

3. Classifying an incoming request

Task: Assign the message below to one of these categories: QUOTE_REQUEST, INFO_QUESTION, COMPLAINT, APPOINTMENT, OTHER.
Also rate the urgency as LOW, MEDIUM or HIGH and write a one-sentence summary.
Rules: If you are not sure, choose OTHER. Do not invent information that is not in the message.
Output format (JSON): {"category": "...", "urgency": "...", "summary": "..."}
Message: [message]

4. Action items from meeting notes

Task: Extract the action items from the meeting notes below.
Rules: For each item, write the responsible person and the date if they appear in the notes; if not, write "not specified." List undecided topics under a separate heading.
Output: Table (Task | Owner | Date) + an "Open issues" list.
Notes: [notes]

5. Product or service description

Role: You are a copywriter who writes with a sales focus but without exaggeration.
Context: Target customer: [who]. Product/service details: [features, dimensions, materials, warranty].
Task: Write a 120–160 word description and a 3-point "highlights" list.
Rules: Do not add features beyond the information provided. Do not use unprovable claims such as "the best" or "number 1." Use correct spelling and grammar throughout.

6. Interpreting weekly figures

Role: You are a business analyst.
Context: Below are the sales / inquiry / return figures for the last 8 weeks.
Task: Describe the 3 most notable changes and their possible causes; suggest 3 questions to check next week.
Rules: Show your calculations step by step. Do not present a cause that is not in the data as fact; label it "possible."
Data: [table]

The "show your calculations step by step" rule in the sixth template matters: AI can make mistakes when working with numbers, and seeing the steps makes checking easier. Critical calculations should always be verified in spreadsheet software.

Calculating the return in numbers

The answer to "What will AI gain us?" should come from the figures in your inventory, not from guesswork. A simple formula does the job:

Monthly gain = (minutes saved per task × monthly task count ÷ 60) × hourly labor cost

Monthly net benefit = Monthly gain − (subscription + usage fees + maintenance allowance)

Sample calculation

At one business, the sales team answers an average of 60 WhatsApp messages a day, and each message takes an average of 3 minutes (the figures are illustrative):

• Monthly messages: 60 × 22 working days = 1,320 messages
• Current time: 1,320 × 3 min = 66 hours/month
• Assumption: The assistant answers 60% of frequent questions on its own, gives the employee a ready summary for the rest, and approval time drops to 1 minute.
• Answered automatically: 792 messages → employee time close to zero
• Left for the employee: 528 messages × 1 min = 8.8 hours/month
• Time saved: 66 − 8.8 ≈ 57 hours/month

Multiply this time by your own hourly labor cost, then compare the result with the system's total monthly cost. There are also gains that belong in the calculation but are hard to put into numbers: instant responses to messages that arrive after hours, fewer lost inquiries, and the time the team can devote to more valuable work.

For a detailed breakdown of cost items and how to calculate the payback period, see our article how to calculate the cost of an AI project.

Don't start without measuring

Before the pilot begins, measure the current state for two weeks: average response time, minutes per task, error or return rate, number of missed inquiries. At the end of the pilot, take the same measurements again. A feeling that "it seems faster" is not enough to justify the investment.

Data security: the traffic light rule

The biggest risk in AI use is not technical; it comes from habits: an employee pasting a customer list or a contract into a free personal account to get their work done faster. The easiest way to prevent this is a rule that sorts data into three colors and that everyone can understand:

The difference between business and personal accounts

Data processing terms differ between the personal and business editions offered by AI providers. Business editions typically include features such as not using the data you enter to train models, administrator controls, user management and logging. Check the current terms of the tool you use in its contract, and make sure the team uses only accounts managed by the company.

Three critical points under KVKK

KVKK (Turkey's Personal Data Protection Law, Law No. 6698) is broadly similar to the EU GDPR in concepts such as transparency, data minimization and cross-border transfer rules. Three points matter most here:

• Privacy notice: If you process customer data with an AI-powered system, your privacy notice should cover the purpose of this processing and the parties the data is transferred to.
• Cross-border transfer: A significant share of AI services process data on servers abroad. An amendment in 2024 reworked Article 9 of KVKK: where there is no adequacy decision, data can be transferred with appropriate safeguards such as a standard contract, and the standard contract must be notified to the Personal Data Protection Authority within five business days of signing. Clarify with your provider which safeguard is being used.
• Data minimization: Give the AI only the data it needs to perform the task. Classifying a message does not require the customer's phone number.

We cover this topic in full detail in our article AI and data protection: a guide to working with customer data. Where a situation requires legal assessment, we recommend seeking advice from a KVKK specialist.

A 90-day implementation roadmap

The plan below is designed so that it can be run by a single "AI lead" with a few hours of support per week from management.

Days 1–15: Preparation

Days 16–45: First pilot

Days 46–75: Connecting and a second process

Days 76–90: Embedding

At the end of 90 days, the goal is not to be able to say "we use AI"; it is to reach a setup where two or three processes are measurably faster, the team uses AI while following the rules, and the cost is known.

In-house AI usage policy: 10 points

There is no need for a lengthy regulation. The points below fit on a single page and answer most of the team's questions:

1. Only AI tools approved and managed by the company are used for work.
2. Red-category data (ID, health, banking, passwords, personnel files) is never entered into any general-purpose tool.
3. Yellow-category data is used only in business accounts, with personally identifying details removed where possible.
4. Every AI output that goes to a customer, supplier or the public is read and approved by an employee.
5. For texts involving pricing, quotes, contracts, health or legal matters, AI only prepares drafts.
6. Information provided by AI (figures, dates, regulations, sources) is verified before use.
7. In automated systems that talk directly to customers, it is made clear that the person is chatting with an AI assistant, and they are handed over to an employee on request.
8. AI-generated images and texts do not use brands, people or works belonging to others.
9. Errors, unexpected outputs or suspected data sharing are reported to the AI lead immediately.
10. The tools in use, their monthly costs and the processes they are used for are kept on a list and reviewed every three months.

Getting the team on board

A significant share of AI projects fail not for technical reasons but because the team doesn't use them. Three things make a difference:

• Appoint a lead. They don't need to be a technical expert; someone curious, organized and well liked by the team is enough.
• Talk openly about the "it will take my job" concern. Show, by sharing the measured gains, that the goal is to reduce repetitive work so the team can spend more time on customers and quality.
• Keep training short and hands-on. A 45-minute session in which each employee writes their own template using an example from their own work is more effective than hours of generic presentations.

When are off-the-shelf tools not enough and custom software needed?

If two or more of the following signs apply, it is time to consider moving to Level 3:

• The same data is manually moved between more than one system every day.
• For the AI to answer, it constantly needs up-to-date information (stock, appointments, order status).
• A customer-facing assistant needs to record who saw what and who approved what.
• The data is sensitive, and you need to determine on which server and under what terms it is processed.
• Automation workflows have multiplied to the point that you only find out one is broken when a customer complains.
• You have a workflow unique to your business, and off-the-shelf tools can only handle it with workarounds.

This decision is broadly similar to the general choice between off-the-shelf and custom software; for a detailed comparison, see our article custom vs. off-the-shelf software, and if you still run your operations in Excel, take a look at our guide to moving from Excel to software. If you need an assistant that works with your own documents, our article what is RAG will guide you, and for assistants that talk to customers over WhatsApp, see our WhatsApp AI assistant setup guide.

7 signs that quietly sink projects

Spotting the following signs early prevents wasted budget and team motivation:

1. No success metric. If the verdict at the end of the pilot is "it wasn't bad," nothing was actually measured.
2. Tool collecting. Every employee is trying a different tool; there is neither cost control nor data control.
3. A 100% automation target. Because AI is expected to answer every message, customer satisfaction drops on difficult cases. The right goal is for easy tasks to be automated and difficult ones to reach a person quickly.
4. The knowledge base isn't updated. Prices have changed, but the assistant still quotes the old price. Nobody owns the information source, and there is no set update frequency.
5. Dependence on one person. When the person who built the workflows leaves, nobody knows what's going on. Every automation should have a short description and an owner.
6. Usage costs aren't monitored. With services billed per transaction, the bill grows as volume grows; a monthly cap and an alert should be defined.
7. Nobody reads the outputs. After the first few weeks, no one looks at what the AI is producing. Random samples should be read at least once a week.

Pre-launch checklist

• ☐ A one-week process inventory has been compiled and the tasks have been scored.
• ☐ A single task and a single success metric have been selected for the first pilot.
• ☐ The current state (time, volume, errors) has been measured.
• ☐ Business accounts have been opened for the team; use of personal accounts has been stopped.
• ☐ The traffic light data rule and a one-page usage policy have been shared.
• ☐ A human approval step has been defined for every output that goes to customers.
• ☐ An owner and an update frequency have been assigned for the information source (prices, FAQs, rules).
• ☐ A monthly cost cap and a tracking method have been defined.
• ☐ The privacy notice and the status of cross-border data transfers have been reviewed.
• ☐ A review meeting for day 30 has been added to the calendar.

Conclusion: Start small, measure, embed

For SMEs, AI is not a fad; applied to the right task, it is a layer of workforce that saves hours every day. What makes the difference is not using the newest tool, but choosing the right task to apply it to, managing its risks with rules, and measuring the results. Start with a one-week inventory, pick the single highest-scoring task, and see the results in 30 days. Embed what works, then move on to the next one.

Sources

• NIST — AI Risk Management Framework (AI RMF)
• Turkish Legislation Information System — Law No. 6698 on the Protection of Personal Data (official text, in Turkish)
• Meta for Developers — About the WhatsApp Business Platform
• Google Search Central — Creating helpful, reliable, people-first content