Where AI helps a 20-person company, and where it doesn't
Which tasks in a small or mid-sized company AI tools handle well today, which ones they handle badly, and how to test a use case before paying for a project.
The management team has agreed the company should "do something with AI". Someone has tried ChatGPT for emails, someone else for summarising a contract. It is useful, sometimes. Now a vendor proposes an AI project and nobody can say whether the problem it solves is real, or what would count as success.
The short answer#
AI language models save time on work that involves reading, writing and reorganising text where a first draft is useful and a person checks the result: drafting replies from past answers, summarising long documents, extracting fields from unstructured emails and documents, classifying incoming requests, and searching internal knowledge in plain language. They are weak where the answer must be exactly right without checking, where the information is not in the text they are given, where rules are already precise enough for ordinary software, or where a mistake is expensive and hard to notice. The most practical results come from connecting a model to the company's own systems and data, for example through the Model Context Protocol, for a narrow task with a human in the loop. Before paying for a project, test the use case on real examples for two weeks and measure how often the output is usable without correction.
Where it works well#
Drafting from examples
Replies to common customer questions, quotes from a standard template, job descriptions, product descriptions. A person edits instead of writing from scratch.
Summarising
Long email threads, meeting transcripts, contracts, tender documents. Useful for deciding what to read in full, not as a replacement for reading what matters.
Extracting information
Pulling order details, dates, amounts or names out of emails and documents that do not follow a fixed format, and putting them into a structured form for a person to confirm.
Sorting and routing
Classifying incoming requests by topic or urgency and sending them to the right person or queue.
Answering from internal knowledge
Letting staff ask questions in plain language across manuals, policies and past tickets, with the answer linked to its source so it can be checked.
Where it does not#
When the answer must be exact
Calculations, legal deadlines, prices, stock levels. A model can produce a confident answer that is wrong. Ordinary software or a lookup in the real system is more reliable.
When the information is not available to it
A model does not know your current stock, your customer's contract terms or today's order status unless it is given access to them. Without that access, it guesses.
When a rule already works
If a process can be described as clear rules, such as "invoices from this supplier go to this cost centre", conventional automation is cheaper, faster and predictable. See what to automate in a small accounting team.
When mistakes are costly and quiet
Anything sent to customers or authorities without review, or decisions about people, need a person responsible for the result.
Connecting AI to your own systems#
The biggest step from "chat assistant" to useful tool is access to real data and actions: reading an order, checking a stock level, creating a draft ticket. The Model Context Protocol is an open standard for connecting AI applications to such tools and data sources in a controlled way, so the model works with current information instead of guessing. What MCP is and what to check before connecting an agent explains it in more detail. The design questions are the same as for any integration: which data it may read, which actions it may take, what requires approval, and what is logged.
Test before you buy#
- Pick one task that happens often and takes real time.
- Collect 30 to 50 real examples.
- Run them through the proposed approach.
- Have the person who normally does the task grade each result: usable as is, usable with edits, or wrong.
- Measure the time for each case, including checking.
- Decide based on that, not on a demonstration with hand-picked examples.
Before staff paste company information into public AI tools during such tests, set the rules in company data in ChatGPT.
Nevo builds AI integrations over the Model Context Protocol. The AI and MCP integration page describes how a project begins.