Integrating AI into a business does not mean buying “an AI” and hoping it handles everything. It means picking one specific, repetitive task, connecting a model to the data that task needs, and deciding upfront where a person checks the result. AI works well on loosely structured text: emails, documents, customer requests, notes. Where fixed rules are enough, it works poorly or adds nothing. Below are ten examples, all illustrative: for each one, what it does, which data it uses and where a human stays in control.
Ten practical AI use cases
The examples below are possible applications, not client cases. I describe them the way I would design them, starting from situations that are typical for small and mid-sized companies and tourism operators.
1. Sorting incoming email
- What it does: reads messages arriving in a shared inbox (info@, bookings@) and labels them by type: quote request, booking change, complaint, supplier invoice. It routes each one to the right person or queue.
- Data: the email text, the list of categories and a few already-labelled examples.
- Human control: whoever receives the email can correct the label; corrections show where the system goes wrong.
2. Extracting data from documents
- What it does: pulls amounts, dates, codes and line items out of supplier invoices, orders or confirmations, and prepares a record to load into the management software.
- Data: the PDFs or document images and the field structure the management software expects.
- Human control: the record never reaches the books on its own. A person reviews it next to the original document and confirms.
3. Searching company documents
- What it does: answers questions such as “what is the cancellation policy for groups?” by searching manuals, procedures and internal contracts, citing the document the answer comes from.
- Data: the document archive, kept up to date.
- Human control: the answer always shows its source so the reader can check it. If the document does not exist, the assistant must say so rather than improvise.
4. Customer support
- What it does: on the website or WhatsApp, answers recurring questions (opening times, meeting point, what to bring) and drafts replies for more complex cases.
- Data: current service information and, where needed, the status of the customer’s booking.
- Human control: sensitive requests (refunds, complaints) go to a person; drafts are approved before they are sent.
5. Bookings via chat
- What it does: a customer writes “do you have space on Saturday for four?”. The assistant checks real availability, offers the options and prepares the booking.
- Data: availability and prices read from the booking engine or management software, not typed by hand into a prompt.
- Human control: the booking is only completed with the customer’s explicit confirmation and under the sales rules the system already enforces.
6. Assisted quotes
- What it does: turns a free-text request (“group of 20, three days, transfers included”) into parameters and fills in a draft quote using the price list.
- Data: price lists, pricing rules and customer records.
- Human control: prices are calculated by the software with fixed rules, not by the model. The operator reviews the draft and sends it.
7. Asking the management software in plain language
- What it does: lets someone ask “how many open files do we have with this supplier?” without opening three screens and setting up a filter.
- Data: management software tables, reached through predefined, authorised queries.
- Human control: the assistant only sees what the person asking can see, and shows where each figure comes from.
8. Ticket triage
- What it does: assigns priority and category to support requests or fault reports and suggests who should handle them.
- Data: ticket history and routing rules.
- Human control: the proposed priority is a suggestion; whoever manages the queue can change it.
9. Summarising case files
- What it does: condenses a long booking or case file (emails, notes, changes) into a few lines for whoever picks it up, for instance when a colleague is away.
- Data: the file’s history in the management software.
- Human control: the summary helps people get their bearings, but decisions are made on the underlying data, which stays available.
10. AI plus API automations
- What it does: combines the points above. For example, a travel agency might receive a supplier confirmation by email, extract the details, find the matching booking and propose an update to the supplier payment deadline.
- Data: email, management software and the APIs that connect the different systems.
- Human control: every change to the data is proposed, confirmed by a person and logged.
Generative AI and automation are different
A classic automation follows rules: when a payment arrives, mark the invoice as paid. It is predictable, testable and cheap to run. A generative AI model interprets text and produces text: it understands that “I’d like to move the trip to Thursday” is a change request, even though nobody filled in a form.
In real projects the two work together. The model handles the ambiguous part (understanding a request, reading a document); rules and code handle everything that must be exact: calculations, availability, permissions, writes to the database. Asking a model to calculate a price or check availability is a design mistake.
Why connecting AI to business data matters
A general-purpose model knows about the world, not about your company. Ask it about Saturday’s availability and it will either say it does not know or, worse, make up a plausible answer. Connected to real data, through document search and calls to the management software’s APIs, it answers with current information and can show where it found it.
That is why “can I connect ChatGPT to my management software?” has a precise technical answer: yes, as long as the software exposes data in a controlled way (APIs or dedicated queries) and there is a layer of code in between that decides what the model may read and what it may ask to do. I cover how a business AI assistant connected to company data works in a separate article.
When not to use AI
- When deterministic rules are enough. If the process can be described as “when A happens, do B”, a classic automation is more reliable and cheaper.
- When errors are unacceptable and cannot be checked. A model can be wrong; if nobody reviews the output and a mistake is costly, the process is not a good fit.
- When the data is missing or messy. If information is scattered across outdated spreadsheets, sort that out first; custom management software often comes before AI.
- When volumes are too low. If a task happens twice a month, the design and testing effort is unlikely to pay off.
Privacy, permissions and human oversight
Three rules I apply to every AI project:
- Suitable providers. I use AI services (Anthropic and OpenAI via API) whose contracts exclude training on submitted data. This does not replace the company’s own data-protection assessment, but it is a necessary starting point.
- Minimum access. The system only reaches the data needed for that process, with the same permissions as the person using it.
- Confirmed, logged actions. Reading and changing data are separate permissions. Important actions require confirmation, answers show their source, and everything is logged so it can be reviewed later.
Start with one process
The most sensible approach is a pilot on a single process:
- Choose the process: repetitive, time-consuming today, with a result that can be checked.
- Identify the data it needs and where that data lives.
- Define permissions: what the system can read, what it can propose, what needs confirmation.
- Measure the baseline: how long it takes today and how many errors occur.
- Test with a small group, collecting the cases where the system gets it wrong or hesitates.
- Compare time and errors against the baseline.
- Roll out further only if the results justify it.
Starting narrow tells you early whether AI is the right answer, before you invest in more processes.
If there is a task your team repeats every day, let’s start there and assess where AI can reduce manual work. You can read how I work on the AI assistants page, or get in touch directly.
Frequently asked questions
Can AI read company documents?
Yes: PDFs, emails, document images and internal archives. The real question is not whether it can read them but how the result is checked. That is why extracted data goes through a confirmation step and answers cite the source document.
Can AI answer customers directly?
It can handle recurring questions and information read from up-to-date data. For sensitive requests it is better for it to prepare a draft or hand the conversation to a person.
Is company data safe?
It depends on how the system is built. Providers whose contracts exclude training on submitted data, access limited to the data actually needed, and a log of every action are the foundations. Legal and privacy aspects still need to be assessed by whoever handles them in your company.