How do machine learning and artificial intelligence technologies help businesses with 10 to 250 staff? Mostly not in the way the usual lists suggest. Fraud detection, demand forecasting and recommendation engines need millions of records that a 60-person firm does not have. The help that actually reaches a firm your size comes from two places: language models that read and write the text your business runs on, and machine learning already built into the tools you use. This article answers which of the two to put to work first, and why.
Two different technologies under one label
Machine learning is the practice of training a model to spot patterns in data and make predictions: which lead is likely to buy, which email is spam, which time of day a contact tends to open messages. It needs examples, usually many of them, and its output is a score or a category.
Artificial intelligence is the wider field. The part that has changed small and mid-sized businesses since 2023 is generative AI, the large language models that read, summarise and write text and speech. They are themselves built with machine learning, but trained by someone else on enormous amounts of text. You do not train them. You give them context and instructions.
That distinction matters for planning. Machine learning on your own data needs volume. Language models need good instructions and good source material. Most firms in the 10 to 250 range have plenty of the second and not enough of the first.
What most articles get wrong
The standard answer lists the most impressive uses: predictive maintenance, fraud detection, dynamic pricing, image recognition on production lines. All real. Almost none of them apply to a consultancy, an agency, a mortgage brokerage or an estate agency with flat revenue.
The error is treating machine learning as something you build. For a firm your size, building a custom model is a data science project with a long payback and a high chance of failure, because the data set is too small and too messy to train on. A model trained on 400 past deals will find patterns, but many of them will be noise.
The useful question is narrower. Where does your business lose time or revenue to work that is mostly reading, writing, sorting or replying? That is where these technologies help, and they help immediately.
Where they help a firm your size
First replies to enquiries. An AI agent reads an incoming chat, SMS or WhatsApp message, answers from your own content, asks qualifying questions and books a call. This is language work, and language models are good at it.
Answering the phone. AI voice agents answer calls outside office hours or when the team is busy, take details and book appointments. Missed calls stop being lost leads.
Writing. Drafting follow-up emails, proposals, social posts and articles from your notes. The output still needs a human edit, but the blank page disappears.
Summarising. Turning a 40-message conversation or a call recording into five lines a salesperson can read before picking up the phone.
Sorting and scoring. Here machine learning appears, but embedded in the tools: spam filtering, send-time optimisation, lead scoring based on behaviour. You switch it on rather than build it.
Routing. Reading what an enquiry is about and sending it to the right person or pipeline, without someone triaging an inbox each morning.
Suppose the firm lists every task that takes more than two hours a week, and the top three are replying to web enquiries (about 6 hours), writing follow-up emails after discovery calls (about 5) and preparing a weekly pipeline summary (about 3). All three are language tasks. An AI agent could take over first replies and booking, AI content could draft follow-ups for review, and a workflow could compile the pipeline summary. If that saved around 10 of the 14 hours and cut reply times from a day to a minute, none of it would have needed a trained model.
The trade-off of buying rather than building
Using machine learning that is embedded in your platform means you accept the vendor's model. You cannot tune the lead scoring algorithm beyond the settings provided, and you cannot see inside it. A firm with a data team and a large, clean data set could build something more tailored.
For almost every firm in the 10 to 250 range, that trade is worth making. The embedded version works on day one, costs nothing extra to maintain and improves when the vendor improves it. The tailored version costs months and may never beat a sensible rule such as "contacts who visit the pricing page twice are hot".
The trade-off with language models is different. You give up some control over exact wording. An AI agent will phrase answers in its own way, and occasionally it will phrase them badly. You manage that through instructions, a knowledge base and regular review of transcripts, not through code.
How to tell if a use case will pay
Three tests separate useful uses from impressive ones.
First, is the task mostly reading, writing, sorting or replying? If yes, language models will probably help. If it needs prediction from your own historical numbers, check whether you have thousands of examples. If not, use a rule instead.
Second, does someone already do this task every week? A task that happens weekly has a measurable cost in hours. A one-off analysis does not need AI.
Third, can a wrong answer be caught before it does harm? First replies, drafts and summaries are easy to review. Regulated advice, contract terms and pricing commitments are not, and should stay with a person.
Start with lead scoring by machine learning. With around 1,500 past cases, too few for a trustworthy model, it fails the first test, so a simple rule based on deposit size and timescale would do the job instead. First replies on WhatsApp would pass all three tests and suit an AI agent. Suitability commentary fails the third test and stays with advisers.
Setting up the first use on the levelpartner.ai platform
The quickest place to see these technologies help is the first reply to a new enquiry. It is language work, it happens every day and every reply can be reviewed.
Before you build anything, list the five questions new enquiries ask most and write down the answer your best person gives to each. Those answers become the agent's first training material, ahead of any page on your website.
- Go to AI Agents > Conversation AI and click + Create Bot.
- Select Guided Form Setup.
- Click Create New Bot.
- Choose the bot type: General Q&A or Appointment Booking.
- Select or create a Brand Voice.
- Configure Bot Settings.
- Configure Bot Goals.
- Save the bot.
- Choose Suggestive as the mode, so a person approves each reply.
Run it in Suggestive mode for a week, then move to Auto-Pilot and read ten transcripts every Friday.
After a month, add the second use: AI-drafted follow-up emails, or an AI voice agent for missed calls. Add one at a time, and measure the hours each one returns before adding another.
What this means across the four industries
For agencies, the biggest gain is usually speed of first reply and drafting client reports. For consultants, it is qualification, so senior people only take calls with a fit. For mortgage brokers, it is gathering case details before the adviser speaks to anyone. For real estate, it is answering portal and website enquiries out of hours and booking valuations and viewings.
The pattern is the same in each. Language tasks first, embedded machine learning switched on where it exists, custom models only if you later have both the data and a specific prediction worth making.
Frequently asked questions
Do I need a data scientist to use machine learning?
Not for the uses that pay at your size. Embedded features and language models need someone who understands your process, not someone who builds models. Hire for data science only when you have a large data set and a defined prediction problem.
Is my data used to train someone else's model?
Check the terms of any tool you use. On the platform, your knowledge base and conversations are used to answer your contacts, and usage is billed as used. Keep sensitive client data out of free public assistants.
What is the difference between AI and automation?
Automation follows rules you write: when a form is submitted, send this email. AI handles the parts rules cannot, such as understanding a free-text question. Most useful setups combine both, as the AI business process automation guide explains.
Where should I read next?
For where to start with your own team, see how to use AI for business. For AI working inside customer records, see AI in CRM examples, and for a first agent, AI agents.