The short answer: delegate bounded tasks with reliable inputs and results you can verify before use. Preparing materials, finding information and reviewing recurring enquiries are useful starting points. Decisions affecting money, commitments and customer relationships require separate human oversight.

Break the work into individual tasks first

“Hand marketing over to artificial intelligence” sounds compelling but offers little guidance. Marketing includes budgeting, understanding customers, writing copy, approving offers and publishing. These activities demand different levels of accuracy and carry different consequences when something goes wrong.

A useful unit is smaller than a department: draft three service descriptions from an approved brief; categorise enquiries; explain an already verified report. Each candidate needs clear answers: what goes in, what should come out, who checks it, and what happens if it is wrong?

Imagine a hair salon with two locations. The owner runs the business, receptionists manage appointments, stylists deliver services, and an accountant maintains the books. This hypothetical example illustrates task selection; it is not a client case study. Every proposed scenario requires testing in a specific business. There are no measured implementation results for this salon.

Editorial diagram · practical synthesis

Delegate the draft, retain the decisionThe degree of automation depends on the cost of error and whether the output can be checked.
  1. 01AI proposesDrafts, options, summaries
  2. 02A person reviewsFacts, context, promises
  3. 03A person decidesPublication, commitments, exceptions

How to identify a suitable candidate

Build your initial list from recent work. Where did someone repeatedly search for the same information? What needed rewriting for another audience? Which messages needed rereading to identify a shared problem? Then examine four characteristics.

  • Frequency. The task occurs often enough to justify setup and training.
  • Available material. Current prices, documents, records or other permitted sources exist.
  • Verifiability. Someone can check accuracy and completeness without repeating the entire task.
  • Reversibility. An incorrect draft can be fixed before publication, sending or another consequential action.

These are practical guidelines, not a universal formula. A frequent task with expensive checks may be less suitable than an occasional, well-defined task. For our salon, drafting parking information seems more reasonable than autonomously resolving complaints. However, if the parking question is always identical, a template is simpler still.

Consider who inherits additional work. If a receptionist prepares a summary faster but the owner must reconstruct missing context, one employee’s convenience becomes another’s burden. Prefer an initial task with an identifiable end user. That person should confirm that the output genuinely helps with the next activity, rather than merely looking tidy.

Marketing starts with an approved offer

Suppose the owner wants to promote available daytime appointments at one location. AI receives an accurate service description, address, permitted offer terms, dates and examples of the desired tone. Customer photographs are used only with the necessary permissions.

The model can draft headlines, a social post, a shorter newsletter version and answers to likely questions. The marketer or owner selects a version and checks prices, dates, address and promises. A receptionist confirms that appointments remain available before publication. The model must not invent a discount, guarantee a colouring result or extend the offer to the other location.

The output is an approved set of materials. Measure time to publication-ready content, editing effort and factual errors. Enquiries and bookings matter, but changes also depend on the offer, season and distribution channel. AI-generated copy alone does not explain increased sales.

Another task is turning customer questions into ideas for future posts. AI helps identify gaps in explanations; the business still chooses its offer and positioning.

What research says about writing tasks

In Noy and Zhang’s 2023 Science experiment, 453 professionals completed occupational writing assignments. ChatGPT access reduced average completion time by 40% and improved assessed quality by 18%. Tasks lacked full company context, and factual accuracy was not comprehensively checked. These findings support investigating writing tasks but do not forecast savings for a particular salon. Science paper; MIT explanation and limitations.

Those limitations matter when selecting tasks: if gathering facts and securing approval consume most of the time, faster drafting may barely change the overall process.

Finance needs verifiable calculations

The owner wants to understand differences between locations: revenue, utilisation, material costs and cancellation rates. Inputs come from booking and accounting systems. Before analysis, someone reconciles reporting periods, removes duplicates, checks refunds and agrees metric definitions. A deposit received and revenue from a completed service, for example, can concern different events.

Calculations belong in a verified spreadsheet with explicit formulas. An accountant or another competent employee checks formulas and reconciles totals against source records. AI can explain validated figures, flag unusual changes and prepare questions for the owner. Every numerical statement should point to its corresponding metric or cell.

The output is a management note: what changed, possible explanations and what remains uncertain. If revenue falls at the second location, the model must not blame a receptionist without additional evidence.

Measure time to a verified report, discrepancies and corrections. Payments, refunds and sensitive disclosures require an authorised person’s approval. Confident wording is insufficient grounds for delegating financial decisions to a model.

Operations and preparing for a shift

Appointments, stylist availability and material requirements change daily. AI could prepare a start-of-shift summary. Inputs include the current schedule, confirmed changes and permitted work notes. The output is a short list of issues for the receptionist to check before opening.

The model flags possible clashes, incomplete bookings and contradictory messages. The receptionist checks them against the booking system. Moving an appointment or changing a shift requires separate approval; a suggestion in a conversation must not become an accepted decision.

Similarly, AI could draft a purchasing list by comparing staff requests, the current product range and stock records. Accurate inventory must come from the inventory system, while quantities should follow verifiable rules. If a simple “stock below threshold” alert suffices, conventional automation is easier.

Useful measures include preparation time, missed important items and false alerts. A lengthy summary nobody reads gains little value from being generated quickly.

Customer support needs clear boundaries

Receptionists answer questions about prices, service duration, addresses and rescheduling policies. AI can draft replies from approved reference material. Inputs should include the customer’s message, applicable terms and only the booking details necessary to answer.

The employee checks the location, service, current price and whether the response addresses the question. Availability is confirmed in the booking system. Missing information should produce a clarification question or referral to a person. An invented appointment slot creates extra work and undermines the customer’s experience.

For complaints, the model can assemble the history and prepare a neutral draft. An authorised employee decides on compensation, refunds and exceptional promises. Conversations should not be uploaded indiscriminately to any service: first check permission to share, retention settings and contractual restrictions.

Measure time to resolution, repeat contacts, incorrect promises and customer satisfaction. A quick reply that forces another enquiry poorly represents actual value.

Why support results do not transfer automatically

Generative AI at Work, published in QJE in 2025, examined 5,172 support agents at one company. An assistant increased successfully resolved issues per hour by 15% on average. Employees retained responsibility for conversations and could edit suggestions. The study describes a specific working arrangement; it does not establish the value of fully autonomous support for every business. Brynjolfsson, Li and Raymond, QJE.

For the salon, this supports testing receptionist assistance on a limited category of enquiries. Its own reference materials, question types and checking time will influence results.

Reviews help identify questions worth asking

The owner could ask AI to group anonymised reviews by waiting time, booking convenience, price clarity, atmosphere and service results. Outputs should include themes with examples from the original messages and a separate list of ambiguous cases.

A person checks whether rare serious complaints were lost or messages about different locations were combined. Mentions in a review sample cannot represent the proportion of all dissatisfied customers: reviewers may not represent the entire customer base.

Practical measures include time to a verified summary and misclassified messages. This work helps select a question for investigation. Changing prices or evaluating an employee requires additional context.

When it is better to avoid AI

If an exact rule describes the task, first check your existing software. Appointment reminders, mandatory form fields and arithmetic generally need no text generation. A good instruction can also remove a recurring problem without introducing another tool.

AI is harder to justify without reliable sources or someone able to check its answers. Automation cannot supply missing management information. Sometimes a consistent price list or reliable cancellation records are the most useful first step.

Costly mistakes also change the choice. Publishing personal details, handling a customer dispute or committing to a payment requires clear approval. A convenient interface does not reduce the consequences of an incorrect action.

Who should select and check the tasks

An internal employee familiar with the process can often organise a bounded use case. Receptionists understand customer questions, accountants recognise accounting errors, and marketers assess content. They need allocated time, access to sources and authority to stop an unsuccessful experiment.

An external specialist may help when comparing tasks across functions, assessing data handling or when the team lacks testing time. Judge the contribution by the task selected, clarity of checks and the team’s ability to continue independently. The cited studies did not test the return on hiring a consultant.

General-purpose AI tools change quickly, so capability lists should not be treated as permanent. Choose using your own examples and current terms. Repeat checks after significant updates; a successful previous test does not guarantee another.

Where to start

Collect several recurring tasks and choose one with permitted sources, an identifiable reviewer and a safe draft output. Compare it with a simpler alternative. Decide beforehand what improvement and quality would justify continuing.

For our hypothetical salon, standard replies or review summaries could be suitable starting points. Actual workload determines the choice. A useful first experiment establishes where AI helps the team and where existing methods remain more convenient.

Sources

  • Noy, S., Zhang, W. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023. Paper.
  • MIT News. Study finds ChatGPT boosts worker productivity for some writing tasks. 14 July 2023. Explanation and limitations.
  • Brynjolfsson, E., Li, D., Raymond, L. Generative AI at Work. The Quarterly Journal of Economics, 2025. Paper.

Related reading