If you want to cut staff costs, which department should start using AI first?

1. Why many companies start AI in the wrong place

Ask in a meeting which department should go first and the answer usually comes from whoever speaks best or is struggling most, which is not necessarily where the payoff is greatest. A fairer approach is to collect the same figures from every department and let the data decide.

Many companies start by buying tool accounts for everyone or building a chatbot for the website, because the results show up quickly. They never ask which parts of the work actually drive cost. The result is that employees use AI to help them write now and then, while quotations are still slow, orders are still keyed in twice and managers still build reports by hand. The project looks modern but doesn't change the cost per unit.

Another mistake is starting with the most important processes, such as pricing, credit or production planning. These have high impact, complex data and many exceptions, so testing takes a long time, even though the company could save time sooner on lower-risk work such as preparing documents or checking data.

2. A business owner's formula for ranking projects

Ask four questions about every candidate task: how often it happens, how long each instance takes, how easy it is to check whether the result is right, and how much damage a mistake would do. Work that happens often, is easy to check and can be fixed when it goes wrong is the safest place to start.

Gather processes from every department and score each one from 1 to 5 on six criteria: volume, time per instance, how standardized the data is, how easy it is to check the answer, the risk if it goes wrong, and how ready the process owner is. Add up the first four criteria plus readiness, then subtract the risk score. The tasks with the highest scores are your candidates for a pilot.

Don't assess by gut feeling. Collect data for 5 to 10 days. The numbers often reveal that the work everyone thinks is big happens only a few times a month, while five-minute document searches or bits of data copying happen thousands of times and cost far more.

3. Sales: a good fit when the team spends more time preparing than selling

Sales is often a good starting point because the results connect directly to revenue. AI can help read leads from several channels, sort out what each prospect needs, check company information, prepare questions, draft proposals, summarize meetings, update the CRM and schedule follow-ups. The system doesn't need to negotiate for anyone. Its job is to make sure each salesperson starts every conversation with complete information and never forgets a follow-up.

A good project is one where there are plenty of leads but the team responds slowly, or where quotations follow a repeating pattern and take a long time to build. The metrics should be time from receiving a lead to first response, quotations per person, the rate of complete follow-ups and conversion. The number of messages AI can produce is the wrong thing to count.

4. Customer service: a good fit when questions repeat and the answers already exist

If customers keep asking about order status, prices, how to use the product or the same policies, customer service is a place where time comes back quickly. AI can find answers in the knowledge base, pull status from systems, summarize history and draft replies for agents. Only at a later stage should it be allowed to reply automatically, and then only to low-risk questions.

Before you start, check whether the knowledge is scattered or contradictory. If each employee gives a different answer, your first problem is the service standard, and technology comes second. Write standard answers, name an owner for the content and set review dates. Once the knowledge base is sound, AI can help service grow without the team having to grow in step with the number of customers.

Measure the share of questions resolved on first contact, average time per case, the number of cases escalated and satisfaction scores. Never measure only the number of cases the bot closes, because a system can close cases quickly and still have customers coming back to reopen the same issue.

5. Marketing: good for speed, but not always the place to reduce headcount

AI lets a team produce drafts of ads, articles, images, campaign ideas and test variations much faster. Producing more content, however, doesn't mean the business makes more profit. Without working sales channels or clear customer data, a company may simply publish more posts and get the same results.

Something worth more than “writing faster” is linking sales data to campaign planning: summarizing why customers didn't buy, building message sets by industry, turning one set of product information into material for many channels, and bringing test results back for the team to decide on. The tool should cut the time from brief to publication while keeping the brand voice and the approval process intact.

Choose marketing as the first department when the company has a clear product, sales channels that work, and a bottleneck in producing campaigns. If the problem is a lack of Product-Market Fit, producing content faster won't fix the root cause.

6. Accounting, procurement and back-office work: savings that are often overlooked

Back-office work involves many documents and rules, which makes it well suited to a mix of automation and AI: reading invoices, checking mandatory fields, matching purchase orders, categorizing expenses, flagging missing documents, summarizing cash flow or collecting quotations from suppliers. The system should prepare the work and flag anomalies, while payment approvals and important entries stay with the people who hold the authority.

The upside is that time and errors are easy to measure, but financial data and access rights need care. Don't send important documents into public tools without a data agreement. The system must separate permissions by role, keep a history and stop AI from changing source data without an approval step.

7. Operations: high returns, but you need to understand the real process

Operations may gain the most from AI, because every minute saved hits costs directly. Examples include turning orders into work orders, checking raw material availability, sequencing jobs, summarizing problems on the floor or warning of late deliveries. The data, however, usually sits in several systems and the floor throws up many exceptions, so this can't be designed from a meeting room alone.

Have the development team follow the real work from order to delivery, noting where people wait, where data is keyed in twice and where employees rely on experience to decide. Then start by having AI prepare data or send alerts. Don't rush to let the system schedule work or issue instructions automatically until you have quality data and a clear way to stop it.

8. A 14-day decision schedule

1. Days 1 to 3: ask every department head for a list of repetitive tasks, no more than five per department, with their approximate frequency and time
2. Days 4 to 7: collect samples of real data and time the high-scoring tasks. Don't use made-up data in the assessment
3. Days 8 to 10: score the six criteria, calculate hours and cost per month, and identify the risks
4. Days 11 to 12: choose the top process and a backup, and set the owner, the KPIs and the things AI must not do
5. Days 13 to 14: write a one-page goal, for example cutting the time to prepare a quotation from 45 minutes to 10 within 30 days, with no rise in the error rate

For most companies, a good starting point is sales, customer service or back-office paperwork, because the data is available and the results are easy to check. There is no fixed rule, though. The numbers from your own company's processes have to decide.

In short: Start AI in the department where the work is “ready,” which takes more than “executives are interested.” The first project needs high volume, measurable time, enough data, answers that are easy to check and risk you can control. Once you win in one place, use the lessons, the data and the confidence you gained to expand into more complex processes.