Where does AI actually help a business?
AI, in the form most companies use today, is software built on language models that reads, writes, classifies and summarizes text. It pays off where three things meet: the work is repetitive, its inputs are written (emails, documents, tickets, records) and someone can say what a correct result looks like. When one of the three is missing, the result is usually a demo that never reaches daily use.
Smaller companies are behind, and Romania more so: 5.21% of Romanian enterprises with 10 or more employees used AI in 2025, against an EU average of 19.95% (Eurostat; all figures on our 2026 statistics page). Where companies do use it, the most common purposes are marketing or sales (34.70% of AI users) and business administration processes (31.05%). Those are the same areas where we see the clearest first projects. Buyers now also ask ChatGPT and Claude for suppliers, and what makes a company website quotable is mostly ordinary good structure.
The OECD found the same pattern from another angle. In its survey of over 5,000 SMEs in seven countries at the end of 2024, 31% used generative AI, but only 29% of those used it in their core activities; most used it for text, and for simple, one-off tasks more than for recurring, important ones (OECD, 2025). The opportunity for most companies is to move AI from ad hoc use by individuals into the processes that run the business.
How can AI help each department?
Below are the use cases we see most often in companies of 10 to 250 people. For each: the task, what AI does, what a person still checks, and the prerequisite without which the project should not start.
Sales
- Task: qualifying inbound requests, keeping the CRM up to date after calls and meetings, preparing first drafts of offers.
- What AI does: reads each inbound request and tags it by product, size and urgency; turns call notes into structured CRM fields; drafts an offer from your templates and price list.
- What a person checks: prices, discounts and anything sent to a customer. AI prepares, the salesperson decides.
- Prerequisite: a CRM the team actually uses, and offer templates that already exist in writing.
Customer support
- Task: answering recurring questions (order status, delivery, returns, how-to) and sorting the inbox or ticket queue.
- What AI does: classifies each message by type and urgency, attaches the relevant order or contract, and drafts a reply from your knowledge base with the sources it used.
- What a person checks: every reply at first, then only the categories that matter (complaints, refunds, anything legal). Customer-facing AI must also say that it is AI; see our article on AI agents for the EU AI Act rule.
- Prerequisite: written answers to the common questions (policies, product sheets, past replies) and access to order data.
Finance and accounting
- Task: capturing supplier invoices, matching payments to invoices, coding expenses, chasing overdue receivables.
- What AI does: extracts fields from PDFs and scans that do not arrive as structured data, proposes the expense account based on past bookings, matches bank lines to open invoices and drafts payment reminders. In Romania, RO e-Factura already delivers domestic B2B invoices as structured XML, so much of that flow is an integration job rather than an AI job; see our invoice automation guide.
- What a person checks: the booking itself, VAT treatment, anything unmatched and every payment. The accountant approves; AI removes the typing.
- Prerequisite: accounting or invoicing software that can export and import data, and booking rules that someone can write down.
HR
- Task: answering employee questions on leave, benefits and procedures; drafting job ads; preparing onboarding material.
- What AI does: answers policy questions from the employee handbook with a link to the source, drafts job ads from a role description, assembles onboarding checklists.
- What a person checks: anything about an individual employee, and every hiring decision. Using AI to screen or rank candidates is a high-risk use under the EU AI Act, a different project with legal obligations to check first.
- Prerequisite: an up-to-date handbook and a clear rule on which personal data may reach the model.
Operations and back office
- Task: entering orders that arrive by email or PDF, processing delivery notes and supplier confirmations, keeping master data clean.
- What AI does: reads the order or document, maps it to your products and customers, enters it into the ERP or prepares the import, and flags what it could not read with confidence.
- What a person checks: the flagged exceptions, and a sample of the rest until the error rate is known.
- Prerequisite: an ERP or order system that accepts imports or has an API, and product codes that are consistent.
Management reporting
- Task: the weekly or monthly report that someone assembles by hand from several systems.
- What AI does: gathers the figures, fills the report, writes a first commentary on what changed and answers follow-up questions about the data.
- What a person checks: the numbers against the source, and the conclusions. The owner of the report signs it.
- Prerequisite: the indicators are already defined. If the figures still live in personal spreadsheets, the first step is bringing them into one system.
Which AI use cases should a small business start with?
The table ranks the use cases above by effort and by what has to be in place. Effort is our qualitative judgment for a typical company of 10 to 250 people; the real effort depends on how many systems are involved and how many exceptions the process has.
| Department | Use case | Effort | Prerequisite |
|---|---|---|---|
| Customer support | Ticket and email classification | Low | A shared inbox or ticketing tool |
| Customer support | Drafted replies from a knowledge base | Medium | Written answers to common questions |
| Sales | CRM updates from call notes | Low | A CRM the team uses |
| Sales | First drafts of offers | Medium | Offer templates and a current price list |
| Finance and accounting | Data capture from non-structured invoices | Medium | Software that can import the extracted data |
| Finance and accounting | Payment matching and reminders | Medium | Bank and invoice data in one place |
| HR | Policy questions from the handbook | Low | An up-to-date handbook |
| Operations | Order entry from email and PDF | Medium to high | An ERP with imports or an API, consistent product codes |
| Management | Recurring report and first commentary | Medium | Defined indicators, data in systems |
A good first project sits in a low or medium row, touches one or two systems and runs often enough that a week of use gives you real numbers.
Where do I start with AI in my company?
Starting with AI means choosing one process, not a tool. These are the four steps we follow, in this order:
- Pick one process with volume. One sentence, one owner, a task that repeats every day or week. "Customer service" is too broad; "answering the order-status emails to the shared inbox" is a process.
- Measure the baseline. Before anything is built, count how many times the task runs, how long it takes and how often it goes wrong. Without this number, nobody will be able to say whether AI helped.
- Pilot with a person in the loop. Build the smallest version that works on real data, and have the person who does the task today check every output. Keep the examples where AI was wrong; they become the test set.
- Scale what the numbers support. When the pilot meets the target agreed in step 2, reduce review to the cases that matter, connect the remaining systems and move to the next process.
In Eurostat's 2025 survey, 70.89% of EU enterprises that had considered AI but did not use it named lack of relevant expertise as the reason. The four steps are designed for that situation: they need decisions and a process owner more than in-house AI specialists. Our AI readiness checklist covers the prerequisites in twelve questions, and what AI implementation costs explains how to estimate a pilot before committing to it.
What is AI not good for yet?
- Open-ended judgment. Deciding whether to extend credit to a difficult customer or how to handle a sensitive dispute depends on history and relationships the model cannot see. It will answer confidently anyway.
- Rare tasks. A process done a few times a year has too few examples to test against and too little volume to repay the build.
- Tasks without a correct answer. If two experienced people in your company would disagree on the right output, there is nothing to measure AI against.
- Exact calculation and the final word on numbers. Models can misread or misstate figures; totals, VAT and payments are checked against the system of record.
- Irreversible actions. Payments, contracts, price changes and deleted records may be prepared by AI. A person approves them.
- Fixing a broken process. If the steps differ from person to person, AI makes the confusion faster. Start with process optimization, and check whether a simple rule-based automation would do the job with no model at all; the difference is explained in process automation vs digital transformation.
How does Murarum help companies start with AI?
Murarum is a business transformation studio that helps companies of 10 to 250 people find where AI pays and build it into the systems they already use. We start with The Diagnosis: a scoped step that maps your processes, checks the data and ranks the use cases, with a written go or no-go the client keeps whether or not a build follows. Sometimes the honest answer is that the first project is not AI. When it is, our AI integration service builds it with an evaluation set, human approval points and cost limits from the first week.
Summary
- AI helps where work is repetitive, written and has a clear correct result, at enough volume to matter.
- Every department has candidates: inbound requests in sales, replies in support, invoice capture and matching in finance, policy questions in HR, order entry in operations, recurring reports in management.
- In each case AI prepares and a person checks what touches money, customers, employees or legal documents.
- Start with one process: pick it, measure the baseline, pilot with a person in the loop, then scale what the numbers support.
- Leave open-ended judgment, rare tasks, final numbers and irreversible actions to people.
