How to Add AI to Your Business Software

Useful AI features, not gimmicks

The AI projects that pay back are usually unglamorous:

• Document extraction. Reading supplier invoices, purchase orders or delivery notes (PDFs, emails, photos) and turning them into structured data for review. See AI document processing.
• Search over your own knowledge. Staff ask questions in plain English and get answers from your policies, manuals and past jobs, with links to the source. See RAG knowledge bases.
• Classification and routing. Sorting incoming emails or tickets by type and urgency.
• Drafting. First drafts of quotes, reports or customer replies for a person to check and send.
• Summaries. Turning long notes, calls or job histories into short summaries.

A chatbot on your website can help too, but only if it answers from accurate, current information and hands over to a person cleanly.

How it works technically

Most business AI features call a hosted large language model through an API, such as those from OpenAI, Anthropic, Google or Microsoft Azure. Your software sends a carefully written instruction plus the relevant data, and gets back text or, better, structured JSON that matches a schema your code can validate.

For answers from your own documents, retrieval-augmented generation (RAG) splits documents into chunks, stores them with vector embeddings (for example in PostgreSQL with pgvector), finds the most relevant pieces for each question and gives only those to the model. The answer can then cite its sources. This is usually better and cheaper than training a custom model.

Around the model you still need ordinary good engineering: queues for long tasks, logging, permission checks so users only see documents they are allowed to, and tests with real examples.

Risks and how to manage them

• Wrong answers. Models can produce confident errors. Validate outputs, show sources, and keep a person approving anything that affects money or customers.
• Data protection. Check the provider's terms on data retention and training, where data is processed, and whether a data processing agreement is in place. Our UK GDPR guide covers the basics.
• Prompt injection. Text inside an uploaded document or email can try to instruct the AI. Treat model output as untrusted, and never let it trigger sensitive actions without checks.
• Cost creep. Usage is billed per token. Monitor it, cache where sensible and use smaller models for simple tasks.
• Supplier change. Models are updated and retired. Keep the provider behind your own interface so it can be swapped.

A sensible way to start

1. Pick one task that happens often and is easy to check.
2. Collect 50 to 100 real examples with the correct answers.
3. Build a small pilot and measure accuracy against those examples.
4. Roll out with human review, and track how often people correct the AI.
5. Automate further only where accuracy is proven.

Measuring whether it works

AI features can feel impressive in a demo and still fail in daily use. Measure them like any other process change:

• Accuracy on your own examples. Keep a test set of real documents or questions with known correct answers. Run it whenever you change the prompt, model or data, and compare scores.
• Correction rate. Log how often staff change what the AI produced. A falling correction rate shows the system is improving; a rising one is an early warning.
• Time saved. Compare how long the task took before and after, including review time. A feature that saves typing but adds checking may not save anything.
• Coverage. What share of cases does the AI handle confidently, and what share goes to a person? Automating the easy majority and routing the rest is often the right design.
• Cost per task. Track model usage costs against the volume handled, so you notice if a change makes it expensive.
• User feedback. A simple thumbs up or down on answers, with an optional comment, gives a steady stream of examples to improve the system.

Decide the thresholds before you start: for example, the pilot moves to rollout only if accuracy on the test set passes an agreed level and staff report time saved. That keeps the project honest and makes the business case easy to explain.

When to talk to dijitul

dijitul adds AI features to existing systems and builds new ones, from invoice extraction feeding Xero to staff knowledge assistants and AI-assisted workflows. We focus on measurable tasks, data protection and keeping people in control. Start with a free chat about the task you have in mind, then get a fixed-price quote for a pilot.

Related