Practical artificial intelligence for smaller teams
Where language models genuinely save money in a small company, where they waste it, and how to add them to software you already run without a risky rebuild.
Most smaller companies are in an awkward position with artificial intelligence. They are told constantly that it will transform their business, they have tried a chat interface a few times, and they cannot see the connection between that experience and their actual operation.
The pattern worth looking for
Ignore the technology for a moment and look for a specific shape of work. Something a person does repeatedly. Something involving reading or writing text rather than exact numbers. Something where an occasional imperfect result is inconvenient rather than serious. Something where a human will see the output before it matters.
Work with that shape is where language models pay for themselves quickly. Work without it is where artificial intelligence projects become expensive disappointments, however impressive the demonstration was.
Four things that reliably work
The first is drafting. Anything your team writes repeatedly, such as quotes, standard replies, job descriptions, report summaries or follow up messages, can be produced as a first version for a person to edit. Editing is far faster than composing, and the quality stays under human control.
The second is reading documents. Invoices, applications, forms and reports arriving as files that somebody retypes into a system can be read automatically, with the extracted fields presented for confirmation. Anything the system is unsure about gets flagged rather than guessed at silently.
The third is sorting. Incoming enquiries, support messages and applications can be classified and routed to the right person immediately, with borderline cases sent to a human. This removes a surprisingly large amount of low value attention from senior people.
The fourth is answering questions from your own material. An assistant grounded in your policies, manuals and previous replies turns a document library that nobody reads into something people actually use. Internal assistants of this kind are frequently the highest value and lowest risk place to start.
Three things that reliably disappoint
Anything requiring exact arithmetic. Calculations should be calculated by software that calculates, not predicted by a model that produces plausible text. If a number must be right, a formula is the correct tool.
Anything requiring perfect consistency. If the same input must always produce precisely the same output, or if a decision must be defensible with a rule a person can inspect, a language model is the wrong instrument regardless of how well it performs in testing.
Anything replacing judgement where the cost of being wrong is high. Unsupervised automation of consequential decisions is how organisations end up explaining themselves publicly. Keep a person in the loop wherever the consequence is real.
What it actually costs to run
Charges are based on the volume of text processed, which means cost scales with usage rather than being a fixed licence. For most small business features the monthly figure is modest, and it becomes significant when a feature processes large documents repeatedly or is used far more than expected.
The estimate is worth doing before building. Take the number of times the feature will run per month, multiply by the approximate amount of text involved each time, and you have a usable projection. Caching repeated answers and using smaller models for simpler steps reduce that figure considerably, and both should be designed in rather than added after the first surprising invoice.
Adding it without a rebuild
The good news for smaller companies is that adding artificial intelligence to existing software is usually straightforward. The work involves connecting to a provider, retrieving the relevant information from systems you already have, structuring the request, handling the response and presenting it in a screen your staff already use.
None of that requires replacing your platform. We routinely add features of this kind to applications that have been running for years, and we keep the integration behind an internal boundary so the provider can be changed later without disturbing anything else, which matters in a field that moves this quickly.
How to start sensibly
Choose one repetitive task with a low cost of error. Build it properly, including the logging that lets you review what it produced. Run it alongside the existing method for a few weeks and measure whether it genuinely saves time. Expand only if it does. Companies that begin with a large speculative programme tend to spend a great deal and learn very little, while companies that begin with one small honest experiment usually end up with three or four features that quietly work.
What does your team do over and over?
Tell us the task and we will tell you honestly whether artificial intelligence would help or whether something far simpler would do the job better.