AI automation means using AI models to handle parts of a task that used to need a person reading, writing or sorting: drafting a reply, pulling figures out of an invoice, summarising a call, routing an email to the right person. For a small business, the best first projects are narrow and repetitive, happen many times a week, and keep a person checking the result before anything reaches a customer.
Since OpenAI released ChatGPT as a free research preview on 30 November 2022, almost every business owner has tried it, and many are asking the same question: what can this actually do for us? This guide answers that with specific starting points, plus the limits you need to plan around.
What has changed, and what has not
The models behind ChatGPT are large language models. They are trained on huge amounts of text and are very good at producing fluent, relevant language: summaries, drafts, rewrites, classifications and answers to questions phrased in plain English.
What changed in the last few months is access. ChatGPT put that ability in a chat window anyone can use. Developers have also had paid access to OpenAI’s GPT-3 models through an API for some time, and Microsoft made its Azure OpenAI Service generally available this week, which matters to businesses that already run on Microsoft’s cloud and want enterprise contracts and controls.
What has not changed:
- These models can be confidently wrong. OpenAI itself warns that ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers. Anything factual needs checking.
- ChatGPT’s knowledge has a cut-off. It was trained on data up to 2021 and does not browse the web, so it knows nothing about your business, your prices or recent events unless you tell it.
- The free preview is not a business system. It can be busy or unavailable at peak times, and OpenAI notes that conversations may be reviewed to improve its systems. Do not paste customer data, contracts or anything confidential into it.
So the opportunity is real, but it is about speeding up skilled people, not replacing judgement.
Where AI automation fits in a small business
Look for work that has these traits:
- Lots of text in, structured result out. Emails, forms, PDFs, call transcripts.
- High volume, low variety. The same kind of task many times a week.
- A clear “good enough” standard that a person can check quickly.
- Low cost of a mistake if it slips through, or a review step that catches it.
Then avoid work where an error is expensive and hard to spot, such as legal advice, medical guidance or financial figures going straight to a client.
Seven practical first projects
1. Email triage and first drafts
Many small teams lose hours to a shared inbox. An AI step can read each incoming email, label it (sales enquiry, support, invoice, spam, job application), pull out key details such as name, company and request, and draft a reply for a person to edit and send.
The person still decides what goes out. The win is that they start from a sorted inbox and a reasonable draft instead of a blank screen.
2. Turning web form enquiries into clean CRM records
Website forms produce messy data: free-text job descriptions, inconsistent phone formats, budgets written as “around 5k?”. An automation can send each submission to a language model with instructions to return set fields (service type, location, budget range, urgency), then create the lead in your CRM through its API. Sales staff get tidy records, and reports start to mean something.
3. Pulling data out of documents
Invoices, purchase orders, delivery notes and application forms all follow rough patterns. Text recognition (OCR) has existed for years, and AI models make the next step easier: finding the supplier, date, totals and line items even when every supplier lays the page out differently. The extracted data goes into a spreadsheet or accounting system for a quick human check before approval.
4. Call and meeting summaries
OpenAI open-sourced a speech recognition model called Whisper in September 2022. Combined with a language model, it can turn a recorded sales call or project meeting into a transcript, a short summary and a list of actions. Always tell people a call is being recorded and follow the consent rules where you operate.
5. Support answers from your own content
A language model can draft answers to customer questions using your help articles, policies and product sheets as source material, rather than its general training. The safest version keeps a person in the loop: the system suggests an answer and links the source, and a team member approves it. Fully automatic replies need much more testing and clear limits on what the system may say.
6. Product descriptions and listings
Online stores with hundreds of products can generate first drafts of descriptions from a spec sheet: dimensions, materials, use cases. A person then edits for accuracy and brand voice. This is especially useful when manufacturers supply the same copy to every reseller, which does nothing for your search visibility.
7. Internal writing jobs
Job adverts, policy drafts, meeting agendas, rewrites of technical notes for customers, first versions of proposals. These are low-risk ways for staff to get comfortable with the tools, as long as no confidential details go into public services.
How to run a first project well
Pick one process and measure it. Before you automate anything, record how long the task takes and how often it happens. You need a baseline to know if the project worked.
Map the steps. Write down each step a person takes today. Mark the ones that need judgement and the ones that are just reading and copying. AI usually helps with the second group.
Start with “draft and review”. Let the AI prepare the work and a person approve it. Only remove the review step once you have evidence the output is reliable for that specific task.
Use the API, not copy and paste. For anything beyond personal experiments, connect to a model through an API inside your own workflow. Automation tools such as Zapier, Make and Microsoft Power Automate can call APIs and webhooks, or a developer can build a small service. This keeps data flowing between your systems instead of through someone’s clipboard.
Write clear instructions and test with real examples. The instructions you give a model (the “prompt”) work like a job description. Test them on a few dozen real past cases, including awkward ones, and adjust until the results are consistent.
Log everything. Keep the input, the AI output and what the person changed. That record shows where the system is weak and gives you evidence if a customer asks how a decision was made.
Data, privacy and cost
- Read the data-use terms of any AI provider before sending business data, and check where data is stored and for how long.
- Strip or mask personal data where the task does not need it.
- Check your obligations. If you handle personal data of people in the UK or EU, GDPR rules on processing and third-party processors still apply when the processor is an AI service.
- Watch usage costs. API pricing is based on how much text you send and receive. Costs are usually small for light use but can grow with volume, so set limits and alerts.
FAQ
Can ChatGPT run parts of my business on its own?
Not safely today. It is a free research preview with no service guarantee, it can be wrong, and it does not connect to your systems. Treat it as a capable assistant, not an employee.
Do I need a developer to automate with AI?
For simple jobs, no-code tools that connect apps and call APIs can be enough. Anything touching customer data, several systems or high volumes benefits from proper development and testing.
Will AI replace my staff?
The realistic gains for a small business are in removing repetitive reading, sorting and drafting so people spend more time on customers and decisions.
What is a sensible first budget?
Start with one process, a short pilot and clear measures. If the numbers do not improve, you have lost little and learned a lot.
Want help choosing a first project?
We help small businesses find the repetitive work worth automating, connect AI models to the tools they already use and keep a person in control of the results. See how we approach it at IWS AI and automation.
