Adding AI to your business software without the hype

Most talk about AI in business software starts with the technology. The conversations that go somewhere start with a problem. Maybe your support inbox takes half the morning to sort, or nobody can find anything in the policy library, or someone rebuilds the same forecast by hand every Monday. Give AI a clear problem like that and reasonable data, and it can save your team real hours. Without them, you'll get an impressive demo and not much more.

We wrote this for operations directors, product owners and founders who want to add AI to software they already use, or to a new system they're planning.

Where does AI in business software actually help?

AI does well with work that involves lots of text or lots of numbers, where a sensible answer is useful even if it isn't perfect every time, and where a person can check the result. We see five kinds of work come up again and again.

• Search and answers over your own documents. Staff type a question in plain words and get an answer taken from your manuals, contracts or policies, with a link to the source. This is the AI assistant for business that most companies start with.
• Sorting and routing. Incoming emails, tickets, claims or orders get read, tagged by topic or marked urgent, and sent to the right team. People still do the work. They just skip the morning spent sorting it.
• Summarising. Long call notes, case histories or reports turn into a short summary a manager can read in a minute.
• Forecasting. Machine learning (software that finds patterns in past data and uses them to predict what comes next) can estimate demand, stock levels, staffing needs or delivery times from your own history.
• Spotting unusual activity. A model learns what normal looks like for your payments, logins or sensor readings, and flags anything that doesn't fit, so a person can take a look before it turns into a problem.

In every one of these machine learning use cases, AI speeds up a slow, repetitive step and your team still makes the call.

Where does AI not help much?

Knowing where AI falls short will save you money. Here are the situations where we'd steer you away from it.

• When a simple rule would do. If the logic is "orders over a set value need approval", just write that rule. It's cheaper, and it gives the same answer every time.
• When every answer has to be exactly right. Language models can say wrong things with total confidence. For tax calculations, dosages or legal commitments, AI can draft or suggest, but it shouldn't have the final word.
• When the data isn't there. A forecast needs history. A document assistant needs documents that are current and reasonably well written. No model can make up for information that's missing or out of date.
• When nobody owns the result. An AI feature with no clear owner and no measure of success tends to drift, and before long people quietly stop using it.

Why does data readiness come first?

Data readiness for AI means your information is in a state a system can actually use. It can be found, it's reasonably accurate, and you're allowed to use it for this purpose. Most of the effort in a successful AI project goes into this step, and it's also the one teams most often skip. Check these points before you choose a model.

• You know where the data lives. Name the systems, folders and databases involved, and who looks after each one.
• It's accurate enough. Duplicates, old versions and missing fields lead to poor answers. A short tidy up often helps more than a better model would.
• It can be reached safely. The data can be read through an API (an agreed way for one system to ask another for data) or a regular export, without anyone copying it onto laptops.
• You're allowed to use it. Check contracts, consent and privacy rules, especially for personal or customer data.
• Access rules are clear. If someone isn't allowed to see a document today, an AI assistant mustn't show it to them either.
• It stays up to date. Decide how new and changed information will reach the AI feature, and how often.

If several of these look shaky, your first project might be data work instead of AI. That could mean joining up systems, cleaning records or building reporting you can rely on. Our data analytics and AI development services cover both.

Which approach should you choose?

You can build AI into business software in three main ways. They differ in effort and cost, and in how closely the answers reflect your own information.

Use an existing model through an API. Several providers offer large language models as a service. These are AI systems trained on huge amounts of text that can read and write in plain language. Your software sends some text, and the model sends back a summary, a category or a draft. It's the quickest route, and it works well for general jobs like summarising and sorting.

Retrieval over your own data. The technical name is retrieval augmented generation, or RAG. Put simply, when someone asks a question, the system first searches your documents for the most relevant passages. It then gives those passages to the model along with the question and asks it to answer using only what it was given. Your documents stay the source of truth, and every answer can link back to where it came from.

Custom machine learning. Here a model is trained on your own past data to make one specific prediction, such as which deliveries will run late or which payments look unusual. It needs more data and more time, but for a narrow problem it can be very accurate.

Plenty of real projects mix them. A support tool might use retrieval to answer questions and a simple API call to sort tickets. Start with the lightest approach that could work, and only move to heavier ones once you have evidence you need them.

How do you keep AI use safe and accurate?

The same concerns come up in every AI project, and each one has a practical answer.

Privacy. Know exactly what data leaves your systems. If you use an outside model, find out where your data is processed, whether the provider keeps it and whether they use it for training. Strip out or hide personal details the task doesn't need. Apply the same access rules as the rest of your software, and keep a log of who asked what. Our application security services include reviewing these data flows.

Accuracy. Put together a test set: a few dozen real questions or cases with known good answers, checked by someone who knows the work. Run it before launch and after every change, and you'll see straight away whether quality is getting better or worse. Ask the system to show its sources, and to say "I don't know" instead of guessing.

Keeping a person in the loop. Where a customer, money or safety is involved, AI should suggest and a person should decide. Make it easy to accept, edit or reject each suggestion, and record which one happened. Over time those records tell you better than anything whether the feature is working.

Treat AI output like work from a capable new colleague. It's often helpful, sometimes wrong, and worth a second look whenever it matters.

How do you measure value with a small pilot?

A pilot is a small trial with a fixed end date, run with real users and real data. It exists to answer one question: is this worth building properly?

How a small AI pilot runs

1. Pick one taskOne team, one clear problem.
2. Measure todayTime, errors, backlog now.
3. Before you buildAgree successWhat result would make it worth it?
4. Build and trialSmall group, real data.
5. DecideScale up, adjust or stop.

Before you change anything, measure the task as it runs today. How long does it take, how often does it go wrong, and how big does the backlog get? Then agree what result would justify going further, for example a clear drop in handling time with no rise in errors. When the trial ends, compare the numbers honestly. If the pilot didn't deliver, stopping it is still a cheap lesson.

You can apply the same thinking to any new software feature. Our guide on building a minimum viable product explains how to test an idea with the smallest version that's still useful.

What do AI features cost to run?

AI changes how software costs add up. On top of the build, you'll often pay an ongoing cost that rises with use. Many model providers charge by the amount of text sent and received, and custom models need computing power to train and run. So the same feature can cost very different amounts with ten users and with ten thousand. These are the things that move the bill most.

What drives the running cost of an AI feature

• How often people use itHigh
• How much text each request sendsHigh
• Which model you chooseMedium
• How often documents are refreshedMedium
• Testing and monitoring over timeSteady

This shows roughly where running costs come from. It isn't a price list, and real figures depend on your usage and provider.

You can keep costs sensible in a few ways. Use a smaller, cheaper model wherever it's good enough. Send only the passages a question needs. Save answers to common questions so you don't pay for them twice. And set spending alerts and limits from the first day. Before you approve the build, ask for an estimate of monthly running costs at low, expected and high usage. For the cost of the build itself, our guide to what custom software costs explains the main factors.

How do you roll AI out to your team?

Even when the pilot works, how you roll it out still matters.

1. Explain what it does and doesn't do. Tell people what the feature is for, where its answers come from and when to double check.
2. Open it up in stages. Go from the pilot group to one department, then to everyone else, fixing problems at each step.
3. Make feedback easy. A simple "helpful or not" button and a way to report a wrong answer give you a steady supply of new test cases.
4. Keep an eye on quality and cost. Rerun your test set regularly, track usage and spending, and review both every month.
5. Name an owner. Someone on your team decides what changes, what the feature can access and when it should be switched off.

Our shipment tracking case study shows the groundwork AI needs: clean, connected data and clear alerts, which on their own led to 40% fewer delivery delays for a logistics client.

How V6 Cube adds AI to business software

We begin with your problem, and only then think about models. What task do you want to improve, who does it today, and what data sits behind it? If a simple rule or a fix to your reporting would solve it, we'll tell you. If AI is a good fit, we'll suggest a small pilot with an agreed measure of success and a fixed estimate agreed before work begins. Here's what you get.

• A data readiness review of the systems and documents involved, listing the gaps to close first.
• The lightest approach that works, which could be an existing model, retrieval over your data or custom machine learning.
• Privacy and access built in from the start, with a person in the loop wherever decisions affect customers or money.
• A test set and monitoring, so you can follow quality and running costs over time.
• Your code and your data. You own all code and designs, and we're happy to sign a confidentiality agreement first.

AI features usually sit inside a bigger product, so this work often goes hand in hand with our custom software development. Read how our software development process runs, or tell us about your project and we'll reply within one business day. For examples by sector, see our pages on manufacturing software and insurance software.