What an AI agent is, in plain words
An AI agent is a piece of software that takes one routine task and carries it from start to finish: it receives an input, checks it against your rules, does the work in your systems and hands the result to a person where a decision is needed. A chatbot mostly talks. An agent also acts: it creates a booking, drafts a reply, fills in a purchase request or sends a report.
What matters most is the rules: what the agent may do alone, what goes to a manager and when to call a human.
How to pick the first task
Start with a task that meets four conditions: 1) it happens often, every day or every shift; 2) it repeats in the same shape; 3) it follows clear rules that you can write down on one page; 4) its result can be measured in minutes, errors or response time. If a task fails even one of these, leave it for later.
A quick test: ask the person doing it now to describe it step by step. One page with few exceptions means a good candidate.
Five typical tasks for a restaurant or bar
Replies and bookings in messengers: the agent answers questions about hours, menu and location, checks free tables and creates a booking, while unusual requests such as a private event go to the manager.
Review replies: the agent reads new reviews and drafts a reply in your tone, and the manager approves or edits it before it is published.
Stock minimums: the agent compares balances with the minimum levels you set and prepares a purchase request that the purchaser confirms.
Loyalty and cashback messages: the agent reminds guests about points or a birthday offer according to fixed rules and only to guests who agreed to receive messages.
Daily reports: every morning the agent collects revenue, average check, cancellations and write-offs into one short message for the owner.
Where a human must approve
Three areas should always pass through a person. Money: refunds, discounts, deposits and purchases above a set limit. Promises: anything that commits the venue, such as a guaranteed table, a special menu or compensation. Public replies: reviews and posts that everyone can see.
The agent prepares a draft and a person approves it in one click, so the time saving stays and responsibility stays with the team. The NIST AI Risk Management Framework describes the same idea for any organisation: define who oversees the system and how its decisions can be checked.
Data and privacy basics
Collect only the data the task really needs: for a booking that is usually a name, a phone number, the date and the number of guests. Ask for consent before sending marketing messages and make it easy to unsubscribe. Keep API keys and passwords on the server, never in a browser page or a shared chat. Give the agent the narrowest access it needs and keep a log of its actions.
How to measure payback
Use three simple measures. Hours saved multiplied by the cost of an hour of the person who did the work. Error rate before and after: wrong bookings, missed orders, late purchases. Response time to a guest, because a fast answer often decides whether the booking happens at all.
An illustrative example, not a real case: a manager spends 1.5 hours a day on messenger replies and bookings. If the agent takes over two thirds of that, you save about one hour a day, or roughly 30 hours a month. Multiply 30 hours by the hourly cost of that manager, and compare the result with what the agent costs to run.
Returns depend on how the work is organised: McKinsey's State of AI survey notes that organisations getting the most value redesign workflows rather than adding AI on top of old processes.
Common mistakes
The first mistake is automating chaos: if bookings are recorded in three places and nobody knows which is correct, an agent will only speed up the confusion. The second is having no owner: someone on the team must be responsible for the agent, check its work and update its rules. The third is having no fallback to a human: a guest who is stuck with the agent should reach a person quickly, and the agent should know when to hand over.
Generative AI is already common among small businesses: an OECD survey of more than 5,000 SMEs in seven countries found it in use in 31% of them.
A 30-day pilot plan
Week one: choose one task, describe it step by step, write the rules and agree on what the agent may do without approval. Record the starting numbers: time spent, errors and response time.
Week two: launch the agent in draft mode, where every action is checked by a person. Collect the cases where it was wrong and adjust the rules.
Week three: let the agent handle simple cases on its own and keep approval for money, promises and public replies.
Week four: compare the numbers with the starting point and decide whether to expand, fix or stop. If you want to estimate your routine losses before starting, the site has a free calculator, and the AI systems page describes how such agents are built.
FAQ
Will an AI agent replace my staff? Usually it takes over repetitive work such as typical questions and reports, leaving people more time for guests. Control over money, promises and public replies stays with the team.
Do I need a new POS or booking system? Usually not. An agent works with the tools you already use, as long as they allow data exchange. If they do not, start with a task that does not depend on them, such as review drafts or a daily report.
How quickly will I see results? A narrow pilot on one task shows first numbers within a month. Compare them with the baseline you recorded in week one, and decide on the next task only after the first one works reliably.
Sources and scope
OECD — Generative AI and the SME Workforce: New Survey Evidence ↗
Survey of more than 5,000 SMEs in seven countries; generative AI is used in 31% of them.
McKinsey — The State of AI ↗
Annual survey: organisations that get the most value from AI redesign workflows rather than layering tools on old processes.
NIST — AI Risk Management Framework ↗
Voluntary framework for managing AI risks, including human oversight and accountability.