Is AI worth it for a small business? Time saved, real costs and a 30-day plan

It can be, if you use AI for a few jobs you repeat every week and you measure the result. The studies so far show modest savings on average: about 3% of working hours in a large Danish study, 5.4% for users in a US survey. Bigger gains appear on narrow, repeated jobs, such as answering the same kind of customer question. So the honest way to decide is to count: the minutes a job takes now, the minutes with AI including your checking, and the cost of the plan. This lesson shows how, with a worked example and a 30-day plan.

This site is written with AI tools, including Claude by Anthropic. This page uses only what makers, public authorities and published studies say, with links.

This is lesson 10 of the free course “AI for small business”, and the last one. Lesson 9 kept confidential data on your own computer; this lesson asks whether it is all worth your time and money.

What do the studies say about time saved?

Each study below was opened on 3 November 2026. Two terms help. A study is peer-reviewed when other researchers checked it before a journal published it; a working paper is shared before that check. Self-reported means people said how much time they saved; nobody timed them.

Danish workers. Anders Humlum and Emilie Vestergaard, working paper 33777 of the US National Bureau of Economic Research (NBER), May 2025, revised March 2026. About 25,000 workers in 11 jobs where chatbots can do part of the work, from accountants and office clerks to legal staff, answered surveys at the end of 2024. Users said they saved about 3% of their working hours, and most said they spent the time saved on other tasks; in Denmark’s official records, earnings and recorded hours showed no measurable change two years after ChatGPT came out. Where the employer encouraged use, provided a business chatbot and trained people, 19% of workers said they saved more than an hour a day. Limits: time saved is self-reported; one country; a working paper, not peer-reviewed.
US workers. Alexander Bick, Adam Blandin and David Deming, for the Federal Reserve Bank of St. Louis, 27 February 2025, an article on the bank’s website, not a peer-reviewed paper. In a US survey of November 2024, built to reflect all US workers, workers who had used generative AI, the kind that writes text, in the previous week said it saved them 5.4% of their work hours, about 2.2 hours in a 40-hour week; over all workers, users or not, 1.4%. A third of users (33.0%) saved one hour a week or less. Limits: self-reported; the authors note that people may spend the time on “on-the-job leisure”. A note on the page points to an update of 13 November 2025 that revised the 2024 results; there, surveys from February to August 2025 put the time saved at 1.6% of all workers’ hours.
Customer support. Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work”, published in 2025 in the Quarterly Journal of Economics, a peer-reviewed journal (the link opens the working paper page, which names the published version). 5,179 support agents at one large software company, answering technical questions from US small business owners. With an AI assistant, issues resolved per hour rose 14% on average and 34% for new and less skilled agents, with minimal change for the most experienced. Limits: one company, one tool, one kind of job.

How many small firms use AI at all? In 2025, 17% of EU firms with 10 to 49 employees used some form of AI, against 55% of firms with 250 or more, according to Eurostat (page last edited 23 October 2026). In Italy, Istat counted 16.4% of firms with at least 10 employees (15 December 2025). A third of the Italian firms using AI (33.4%), mostly with 10 to 49 employees, named none of the business purposes listed in the survey: use that is “ancora poco strutturata” (“still unstructured”), Istat writes. Neither survey covers firms with fewer than 10 employees.

Together, the studies show savings that are real for some jobs and modest on average. In the Danish study, reported benefits were highest where the employer organised the use. That is the approach of this course: a few repeated jobs, chosen on purpose, and measured.

How do you work out whether it pays in your business?

Do it on paper. For each job you repeat every week, write five numbers:

Minutes a week the job takes now, all of you together.
The share AI can draft, as a guess at first: 0% for a phone call, perhaps half for standard emails.
Minutes the AI draft replaces: line 1 times line 2.
Minutes to give the material and check the result, every time. This is the cost that is easiest to forget.
Net minutes saved: line 3 minus line 4.

Add up the net minutes, turn them into hours a month, multiply by what an hour of work costs your business, and compare with the monthly cost of the plan. Take the real prices from lesson 3, not the round figure below.

A worked example: a three-person estate agency

This is a made-up example to show the method, not a promise and not a measurement. The shares and checking times are assumptions.

Job	Minutes a week now	Share AI drafts	Minutes replaced	Minutes to check	Net saved
Property descriptions, 6 a week at 40 minutes	240	70%	168	60	108
Replies to standard enquiries from property portals, 50 a week at 4 minutes	200	50%	100	40	60
Updates to owners after viewings, 8 a week at 15 minutes	120	50%	60	32	28
English versions of listings, 3 a week at 30 minutes	90	80%	72	24	48
Total	650		400	156	244

The agency saves about 244 minutes a week, roughly 4 hours, or about 16 hours a month counting four weeks. The three people work about 120 hours a week together, so this is about 3.4% of their time, close to the 3% reported in the Danish study.

Now the money, with two assumptions to replace with your own:

An hour of work is worth 30 euros to this agency. 16 hours a month are worth about 480 euros.
The plan costs 25 euros per person a month before VAT, near the top of the range in lesson 3. Three people cost 75 euros a month.

The plan pays for itself if the office saves at least 2.5 hours a month (75 divided by 30), about 40 minutes a week. Here it saves far more. But time saved is only worth money if it goes somewhere useful: one more viewing, a quicker reply to a buyer, an evening off.

What are the hidden costs?

The monthly price is the easy part. Five costs hide behind it:

Checking time. Every draft must be read before it reaches a customer, because AI can write things that sound right and are false. If checking takes as long as writing, that job does not pay.
Training. Lesson 2 asks you to help everyone who uses AI understand it, and to note it in the training log. In the example, two hours each at the start cost 6 hours, about 180 euros of time, once. In the Danish study, training on its own went with smaller gains; benefits were highest where encouragement, a business tool and training came together.
Mistakes. A wrong floor area in a listing, an old price, a promise your policy does not allow: one mistake that reaches a customer can cost more than a year of the plan. List every error you catch.
Forgotten subscriptions. A free trial that turns into a paid plan: lesson 3 notes that Microsoft’s one-month trial converts automatically. A seat for a colleague who stopped using it. An automation service from lesson 7 with a bill of its own. Keep every AI subscription on one list, with its renewal date.
Usage limits. Each plan allows a certain amount of use, and jobs that run on their own use it up. OpenAI says a scheduled ChatGPT task can use your plan’s allowance even when it has nothing to report (lesson 6). Anthropic says multi-step tasks use more than a quick question, and Microsoft bills its Copilot Cowork agent by use (lesson 8). If you keep hitting the limit, the cheap plan in your sums may not be enough.

What has the course covered? A short recap

Lessons 1 and 4. AI drafts, you decide: you give it the facts, with placeholders such as [CLIENT] instead of names on a personal plan, and check every name, number and price.
Lesson 2, the rules. The AI Act asks you to help your staff understand AI. The European Commission says no certificate is needed: you can keep an internal record, such as the training log. Customer data falls under the privacy law.
Lesson 3, the plans. Most business plans keep your work out of training by default, and add an administrator and a data protection contract. The AI itself is often the same.
Lesson 5, the instructions. Write your tone and rules once, and keep the price list in a dated file in a project or custom assistant, replaced when prices change. No customer data in it.
Lesson 6, scheduled tasks. An assistant can repeat a job at a time you choose, such as a list of deadlines every Monday, and leave the result for you. Read every result, check dates on the official page, and never schedule anything that writes to customers, pays, deletes or changes accounts.
Lesson 7, email and spreadsheets. A connector, or an automation tool such as Zapier, Make or n8n, links the AI to your inbox, calendar and sheets. Give each link only the access the job needs. Let it save a draft or add a row, never send or pay: a person reads the draft and presses Send. Remove access you no longer use.
Lesson 8, the agents. An agent is an AI that opens web pages, clicks, types and works on files for you. Give it slow, repetitive jobs with little risk, such as comparing suppliers’ published prices or sorting a copy of your receipts; keep confirmations on, check the result and press the last button yourself. Never let it pay, sign, delete originals or send to customers. The makers say the responsibility stays with you: Google writes “You’re responsible for Gemini’s actions during a task, including mistakes and unexpected results like purchases” (Google), and Anthropic “You remain responsible for all browser actions taken by Claude performed on your behalf” (Anthropic). ChatGPT’s agent mode is retired: OpenAI’s help page, in a copy read on 31 October 2026, says “ChatGPT agent is no longer available” and points to ChatGPT Work.
Lesson 9, local AI. A free program such as LM Studio or Ollama runs a model on a computer in your office, so confidential text stays there instead of going to an outside service. The trade-off is usually a smaller, slower model that makes more mistakes, and a computer with enough memory, 16 GB as a starting point. Keep the online extras off and encrypt the disk in case the computer is stolen.

How do you start? A 30-day plan

Week 1: choose and measure.

List the jobs you repeat each week and time them for one normal week. Without this starting point you cannot compare.
Pick two or three jobs that are written, repeated and low risk (lesson 1).
Write a one-page rules sheet and start the training log (lesson 2).
Choose the plan (lesson 3), and note the renewal date of anything you sign up for.

Week 2: try it by hand.

Use the requests from lesson 4 on the chosen jobs.
Save your tone, rules and price list as fixed instructions (lesson 5).
Time each job again, including the checking. Write down every mistake you catch.

Week 3: automate one safe step.

Set up one scheduled task, such as the daily reminder list or the Monday list of deadlines from lesson 6, or one automation from lesson 7 that turns a website form request into a draft reply, or invoice emails into rows in a sheet. It saves drafts and rows only; a person reads and sends.
If a job needs clicks on websites, try an agent on something harmless, such as comparing three suppliers’ published prices, with confirmations on (lesson 8). It does not pay, sign, delete or send.
If confidential data holds you back, try local AI on one document first (lesson 9).

Week 4: decide.

Fill in the table with your own minutes, the real price and the training hours. Then choose:

Keep if the net saving is clearly above the cost, no serious mistake reached a customer, and people use it.
Change if one job pays and the others do not: keep that job, drop the rest, and check whether fewer seats or a cheaper plan would do.
Stop if checking takes as long as doing the work yourself, or mistakes reached customers. Cancel before the next renewal, turn off scheduled tasks and automations, remove the access you gave to your email and files (lesson 7), and delete project files you no longer need.

Whatever you choose, put a date three months ahead in your calendar to look at the numbers again.

Where to go next

The course index: AI for small business lists all ten lessons.
AI for everyone is the complete guide to today’s AI, from the start.
Will AI take my job? sets out what the big studies say about AI and work, with their limits.

Quiz: check what you learned

Pick one answer for each question: you see at once whether it is right, and why. Your answers are not sent anywhere.

Under the EU AI Act, does your staff need a paid certificate to use AI at work?
Show the answer
Right answer:B. No: you help them understand AI and can keep an internal record, such as the training log
The European Commission says there is no need for a certificate and that a business can keep an internal record of the training; lesson 2 gives a log to copy.
What does a business AI plan add compared with a personal one?
Show the answer
Right answer:B. On most plans, no training on your work by default, an administrator and a data protection contract
The AI itself is often the same; a business plan changes the contract and the controls around it.
An automation or an agent has prepared a reply to a customer and a payment to a supplier. What should happen next?
Show the answer
Right answer:B. A person reads the reply before it goes out and makes the payment personally
Lessons 6 to 8: automations and agents prepare drafts and lists; paying, deleting, signing and sending to customers stay with a person, and the makers say you are responsible for what their agents do.
Why would a law office run an AI model locally, on its own computer?
Show the answer
Right answer:B. Because what is typed stays on a computer in the office instead of going to an outside service
Local AI keeps confidential text on your own machine; the trade-off is usually a smaller, slower model and a computer with enough memory.
When you work out whether AI pays in your business, which cost is the easiest to forget?
Show the answer
Right answer:A. The minutes spent giving the AI the material and checking its result, every time
Line 4 of the method in this lesson: every draft must be read before it reaches a customer, and if checking takes as long as writing, that job does not pay.

Common questions

How much time does AI really save at work?

Less than headlines suggest, on average. In a large Danish study, chatbot users said they saved about 3% of their working hours; in a US survey of November 2024, users said 5.4%, and a third saved one hour a week or less. Bigger gains were measured on narrow, repeated jobs, such as customer support.

How do I know if AI is worth it for my business?

Measure it. List the jobs you repeat every week and the minutes each takes, then try AI on two or three of them, counting the time to give the material and check the result. If the minutes saved, valued at what an hour of work costs you, are clearly more than the plan, keep it.

What are the hidden costs of AI for a small business?

The time spent checking every draft, the hours of training at the start, the cost of a mistake that reaches a customer, subscriptions nobody uses any more, such as a free trial that turned into a paid plan, and tasks or agents that use up your plan's allowance.

Topics: AI for small business, AI course, time saved, costs, productivity, studies, work