AI for business: where to start when you don't know what it can do
You don't need to know what AI can do to get started. You need to know where your time goes. Four steps, with examples from eight trades.
Tuesday, 09:10. The head of accounts opens her inbox: 43 supplier invoices, half of them PDFs, the rest photos taken on a phone. By lunchtime she’ll have keyed every one into the ERP. Nobody in the company calls this “an AI opportunity”. They call it Tuesday.
If you hear about AI everywhere but can’t picture what it would change in your own business, you haven’t fallen behind. You’re in the majority. This guide assumes no technical knowledge. Only that you know where your time goes.
Why you can’t see what AI could do
Among firms with 10+ staff, AI use in Greece fell from 9.81% in 2024 to 8.93% in 2025, while across the European Union it rose from 13.48% to 19.95% (Eurostat, official statistics). Roughly 4 in 5 have never even considered using it: ≈79%, against ≈68% in the EU (our analysis of the same data).
Of those that considered it but didn’t go ahead, fewer than one in four said they had no use for it (22.0%). What stopped most of them was a lack of expertise (61.8%, and even more often across the EU: 69.9%) and, a distinctly Greek problem, cost: 54.3% against 38.2% in the EU (our analysis, Eurostat). In a survey of Greek SMEs, 92% named not knowing what AI can do as a main barrier (AI Readiness Index 2025, survey co-authored by a vendor).
The problem stays invisible for three reasons:
- It’s sliced thin. Five minutes here, ten there, across 30 people. No line in your accounts says “retyping”.
- It’s become normal. “We’ve always done it this way.” The workarounds turned into job descriptions.
- You picture a chat window. In practice AI works inside the workflow you already have.
What AI actually does in a business
AI for business, as we use the term here: software that reads, listens, sees or predicts inside an existing workflow, and leaves the decision to a person wherever a mistake is costly.
- It reads invoices, delivery notes and order emails, and posts the fields to the ERP.
- It listens and answers the phone at 15:01 or Viber at 23:40, saying in its first sentence that it’s AI.
- It sees every part on the line, not one sample an hour.
- It predicts demand and breakdowns.
- It drafts quotes and replies for a person to approve.
Every business is a factory: some produce parts, some produce room nights, some produce invoices. The logic is the same everywhere.
Eight trades, one question: what eats the hours?
Use cases. In every row, something stays with a person.
| Trade | Today | What the system takes on | What stays with a person | What we measure |
|---|---|---|---|---|
| 50-room hotel, Crete | WhatsApp at 23:40, answered at 08:30 | Replies in the guest’s language, with the transfer price from the PMS | Complaints | Time to first reply |
| Accounting office, Heraklion | Documents classified by hand | Checks VAT number, VAT and totals, suggests a classification | Approval | How many go through uncorrected |
| Law firm, 5 lawyers | Deadlines copied out of emails by hand | Picks out parties and deadlines, drafts the calendar entry | Confirmation | Time to first reply |
| Tutoring centre, 300 students | Every parent update written from scratch | A draft per student from attendance and marks | The teacher’s sign-off | Teacher hours a month |
| Dairy, Thessaly | The week’s quantities set by gut feel | A forecast per product, as a range | The production decision | Forecast error |
| Steel fabricator, Almyros industrial area | 3.1 mill certificates typed into Excel | Reads the heat number and chemistry, checks them against spec limits | Deviations | Minutes per certificate |
| Olive-oil mill and exporter, Crete | Three folders for every shipment | Documents from the ERP, a portal for the importer | What the importer sees | Shipment preparation time |
| Lift maintenance, Crete | A handwritten job sheet that reaches the office on Thursday | An app that works with no signal, the invoice in the ERP | Approval | Days to invoice |
Two of the eight don’t need AI at all: the exporter needs systems that talk to each other, and the lift firm needs an app that works in a basement. A serious diagnosis tells you that too.
Where to start: four steps
1. Write down an ordinary day, not the org chart
Take three roles, say accounts, sales and the warehouse, and follow their day from 08:00 to 16:00. Mark four kinds of time: retyping data that already exists digitally, searching, waiting for sign-off, and giving the same answer again.
2. Put a number on it: minutes × volume
hours a month = minutes each time × times a month ÷ 60
Example: 900 delivery notes a month × 6 minutes ÷ 60 = 90 hours. At €18.20 an hour, the average hourly labour cost in Greek manufacturing (Eurostat 2025, official statistics), that comes to about €1,640 a month. That’s what the task costs today, not what you’d save. How much of it comes back is in the full Workflow X-Ray formula.
3. Run each task through five questions
- Does it happen dozens or hundreds of times a month?
- Is the input already digital (email, PDF, Viber, ERP)?
- Is there a rule you could explain to a new colleague?
- If it goes wrong, would a check catch it before it reaches the customer?
- Is there someone in the company who wants to own it?
Five yeses: you’ve found your first candidate. Four: worth measuring. Three or fewer: the process itself probably needs fixing first.
Run one task through the filtera self-check, not a diagnosis
Name one task your team repeats, then answer five questions.
Answer all five to see the result.
This tool needs JavaScript. Everything it shows is also in the text.
Your answers stay on this page and disappear when you close it.
4. A pilot with targets signed in advance
One process, 4–8 weeks, in live operation. The target is written down before the code, for example “7 in 10 invoices go through with no correction”. You see the numbers every week. At the end, you decide with the numbers in front of you.
What not to start with
- A chatbot on the homepage because everyone has one. If unanswered calls and messages aren’t where you lose customers, you’ll be counting conversations, not hours.
- The tool before the workflow. Only 10% of AI’s value is the algorithm itself and 20% is data and technology; the other 70% is redesigning the work, the people and the rules (BCG, consultancy).
- An in-house pilot with no targets. Around 95% of generative AI pilots showed no measurable effect on the bottom line. Projects built with outside specialists reached deployment twice as often: 67% against 33% (MIT NANDA, self-reported, mostly large US companies).
- The idea of “cutting jobs”. 69% of Greek firms struggle to find the right staff (EY, 2026, survey). The hours you free go where you need people.
Data, the law, and a person on the line
Since 2 August 2026, every chatbot or voice assistant must say at the first contact that it is an AI system (Article 50 of the AI Act, Jones Walker). In Greece, Law 5321/2026 puts the Hellenic Data Protection Authority in charge of Article 50 (Lawspot), the same body that oversees GDPR. Ask whoever proposes such a system for a data processing agreement, a list of sub-processors and a one-step route to a human. Ours are written down in the AI Trust Standard.
If you’d rather do it with us
The Workflow X-Ray runs the four steps for you, stopwatch in hand and on your own data. It ends with an Hours-Leak Map, five opportunities with the payback worked out, and a live demo on your own material. The fee is credited in full against a pilot signed within 60 days.
Cost, excluding VAT: €890 for one process in one week (X-Ray Compact) and €1,900 for up to three departments in two weeks.
If you run a plant, read in what order AI belongs on the factory floor.
Frequently asked questions
What can AI do for a small or mid-sized business?
In practice it reads documents and posts them to the ERP, answers calls and messages (saying up front that it’s AI), checks products with a camera, forecasts demand and breakdowns, and drafts quotes. It pays off first on high-volume tasks with digital input and a clear rule, with a person keeping the decisions that carry real cost.
Do we need a lot of data to start?
Not for most first projects. Invoices, emails, call logs and ERP exports are enough to measure the workflow and set up a pilot. A long history matters mainly for forecasting demand or breakdowns. If yours is missing, step one is to start recording it properly.
How much does a first trial cost?
The diagnosis costs €890 for one process and €1,900 for up to three departments, excluding VAT, and is credited in full against a pilot signed within 60 days. A single-process pilot starts at €6,000, runs for 4–8 weeks and has targets signed before development begins.
Will it work with the ERP we already have?
Usually, yes. Most Greek ERPs, such as SoftOne, Entersoft and Epsilon, connect through web services or exports, and myDATA has a public API. The integration is often the biggest part of the work, which is why we size it during the diagnosis, before any proposal.
Sources
- Eurostat, Artificial intelligence by size class of enterprise (isoc_eb_ai), 2024–2025 data · official statistics · firms with 10+ staff, excluding the financial sector. The ≈79% and the barrier shares are our analysis of the same data.
- Eurostat, Labour cost levels (lc_lci_lev), 2025 · official statistics
- AI Readiness Index 2025, ENTERSOFTONE and AUEB/ELTRUN, via powergame.gr (in Greek) · survey co-authored by a vendor
- BCG, Artificial Intelligence · consultancy
- MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, via Fortune · research, self-reported data
- EY Entrepreneurship Barometer 2026, via OT.gr · survey
- Jones Walker, AI Act: what applies from 2 August 2026 · legal analysis
- Lawspot, Law 5321/2026 implementing the AI Act (in Greek) · legal news