AI for small businesses: a four-week plan

Small businesses start with one defined task

A small business needs a manageable first trial. Our recommendation is one recurring task, one already approved tool and one person responsible for the process. Set a decision date four weeks ahead. The number of generated texts is not a measure of success.

Consider a fictional retailer with five employees. It wants to turn approved product facts into short descriptions. AI prepares drafts; a person checks the facts and publishes through the existing process. Customer data, automatic shop access and new integrations are outside this first trial.

The plan also works for other bounded writing tasks with checkable source material. If the job merely inserts fixed details into a template, try that template first. Our comparison of rules and AI helps with this initial choice.

Week 1: establish a baseline from current work

The business first measures how it works today. For ten typical descriptions, record everyone’s working time from assembling the facts to finishing the checked text. Keep text length comparable and note unusual cases. Add person-minutes: five minutes of writing plus three minutes of checking equals eight working minutes. Apply the same requirements later. Germany’s Mittelstand-Digital centre for future working culture also recommends recording the time spent on individual work steps.

The owner sets the task and spending limit. One person prepares drafts and a second checks product facts. The other three receive a short introduction: which data is allowed, who resolves uncertainty and where approved texts are stored. Nobody shares account credentials.

Copy these five completed lines into your starting checklist. The time threshold is a chosen business requirement, not a general performance promise:

• Task: a German product description from an approved fact sheet, approximately 80 words; shorter if the facts do not support more.
• Tool: one business-approved AI chat, without automatic access to the shop.
• Responsibility: purchasing supplies the facts; sales prepares and checks the copy.
• Metric: total working minutes per fact-checked, approved description, including unsuccessful attempts.
• Decision on day 28: at least two minutes saved per text, affordable ongoing costs and no unresolved factual errors.

One approved tool is enough for the first month

An existing account takes priority over another subscription. Use a business-approved chat that can work with text. Microsoft, for example, describes Copilot Chat as a drafting and rewriting tool. That capability alone does not establish which data your business may enter.

Before starting, ask the responsible person to confirm the account type, permitted data, storage and additional services involved. Microsoft explains protections for Copilot Chat prompts and responses; web search queries receive separate handling. A provider commitment is not permission to enter any business document.

Start with fictional or already public product facts. If no tool is approved yet, keep week 1 focused on the baseline and artificial examples; real business data must wait. Removing names is not always enough, as our guide to masking customer data explains.

Week 2: use a consistent instruction

The instruction stays the same across ten practice cases. Microsoft recommends setting out the goal, context, source material and expectations. A prompt is the instruction you give the AI chat. Begin with fictional cases, including one missing a material specification and one with conflicting dimensions.

Reusable prompt: “Write a German product description of no more than 80 words using the approved facts below. Audience: online shoppers. Use only the supplied information. Do not add properties, suitability claims, certifications or delivery promises. Return: draft, facts used, open questions. If information is missing or contradictory, flag it without inventing a substitute. Facts: [insert details].”

Completed practice case: B12 storage box; capacity 12 litres; external dimensions 40 × 30 × 15 centimetres; lid included; grey; material unspecified. The expected check is straightforward: those five known facts may appear, while material and suitability for food remain open. This is an editorially defined expected result, not an observed model output.

Ask purchasing to resolve the missing material. Polite wording does not make an invented plastic type more accurate.

Week 3: checking is part of the working time

Measurement ends only when the result is approved. Process twenty new, comparable cases within the approved data boundaries. Start timing when you assemble the facts; add everyone’s working time for prompting, reading, questions, corrections and approval. Include waiting whenever it prevents the person from doing other work.

A discarded AI draft still consumes time. If somebody then writes the description manually, record both efforts in the same row. Otherwise, the worst cases disappear from the result. Record unusual difficulties rather than removing them afterwards.

The reviewer matches every property to the original fact sheet. Microsoft explicitly recommends checking facts and figures against source material; another AI answer does not replace that check. Use these fields for each case:

• Case ID | comparable task | total working minutes | facts checked: yes/no | approved: yes/no | error or question.
• Fictional example row: P17 | box description, five facts | 6 minutes | yes | yes | invented material removed.
• Fictional example row: P18 | box description, conflicting dimensions | 11 minutes | no | no | awaiting clarification from purchasing.

Week 4: repeat the comparison

The final week tests the process on another twenty cases. Keep the same instruction and quality requirements. New products of similar difficulty are better than repeating the ten practice descriptions, which people may now know by heart.

For each week, divide all working minutes by the number of fact-checked, approved descriptions. Record any unresolved cases and finalise the comparison only after they have been resolved. Compare the result with week 1 and review the error notes together.

Agree on this stopping rule in advance: do not expand on day 28 if the saving falls below the agreed two minutes per text, ongoing costs consume the benefit or factual errors remain unresolved. Too few comparable cases do not establish success either. End the trial for now or explicitly agree on another bounded test. Letting it run unnoticed is not a decision.

Three minutes saved can leave a small benefit

The calculation needs working time and ongoing costs. All the following figures are chosen planning assumptions, not measured savings or provider prices: eight minutes per description before, five with AI including review, and sixty suitable descriptions each month. That gives 60 × 3 = 180 minutes, or three hours.

Subtract one hour a month for maintaining the instruction and reviewing errors. The remaining two hours at an assumed internal value of €35 an hour amount to €70. Allow another €30 for the tool and the calculated benefit is €40 a month. This values freed working time; it is not automatically additional cash.

If setup, initial training and practice together take three person-hours, that adds €105. A projected full operating month of sixty descriptions including that initial effort therefore shows a €65 loss; with unchanged assumptions, setup breaks even after roughly three months. If each description saves only one minute, monthly maintenance consumes the entire time saving. This business would stop the trial in that case.

Three questions settle the next step

The business decides from its completed starting checklist. For this retailer, we would continue only the one writing task if the repeated comparison supports it. Other departments and automatic publishing require separate decisions.

Must all five people use the tool immediately? No. Everyone learns the rules; two people handle cases initially. Do we need an agent already? No. A chat without shop integration is enough for this drafting process. What remains if we stop? The checked fact template, time comparison and documented errors also help with manual writing.

Today, put the task, responsible person and decision date on one page. Tomorrow, complete the first two cases without AI and start the time log. If you need help choosing and setting up the workflow, bring that checklist to our free introductory call.