AI Integration for Existing Software
Adding AI where it genuinely pays for itself - and saying so when it does not.
Most businesses do not need an AI product. They have a specific task that consumes staff time - reading documents, classifying enquiries, searching an archive nobody can search, drafting repetitive text - and language models are now good enough to take a meaningful share of it.
That is the work I do: adding well-scoped AI capability to systems you already run, measured against whether it actually saves time.
Where AI actually earns its keep
Tasks with high volume, tolerable error rates and a human in the loop.
- Document extraction - invoices, forms, contracts
- Semantic and natural-language search over your own content
- Classification and routing of enquiries or tickets
- Summarisation of long documents and threads
- Retrieval-augmented answering over internal knowledge
- Drafting assistance inside existing workflows
- Data cleaning and normalisation at scale
- Transcription and meeting note processing
- Structured output from unstructured sources
- Evaluation harnesses to measure whether it works
How an AI integration runs
Small, measured, and abandoned quickly if the numbers do not support it.
1Find the task worth automating
We identify a specific, high-volume, well-bounded task and establish what it currently costs in time. Without that baseline there is no way to tell whether the result was worth building.
2Prototype against real data
A prototype on your actual data, not a curated demo set. Real data is where accuracy claims meet reality.
3Measure it
An evaluation set with known correct answers, so accuracy is a number rather than an impression. This step is skipped in most AI projects and it is the reason so many of them quietly fail.
4Integrate with a human in the loop
The output goes into the existing workflow with review where the cost of an error justifies it. Confidence thresholds decide what gets checked.
5Monitor and control cost
Token usage, latency and accuracy tracked in production, with caching and model selection tuned so the running cost stays proportionate to the benefit.
A realistic view
Language models are extremely good at a certain shape of task and unreliable at others. They are strong at transforming, summarising, classifying and extracting from text. They are weak wherever a confidently wrong answer is expensive and nobody is checking.
The engineering that matters is mostly around the model rather than in it: getting the right context in front of it, constraining the output to a structure your system can consume, handling failures and rate limits, caching aggressively, and measuring accuracy continuously rather than once during the demo.
I will tell you when a piece of work does not need AI. A well-written rule, a database index or a fixed form is often faster, cheaper, and correct every time - and reaching for a model instead is a way of spending money to make a system less predictable. Where AI is the right tool, it gets added to your existing application rather than replacing it; see app development and API development for the surrounding work.
Related services
App Development
Web, mobile and desktop applications for Windows, macOS and Linux, built with a real focus on the end user.
API Development
REST and GraphQL APIs, integrations and microservices that are documented, versioned and pleasant to consume.
Data Migration & ETL
Moving data between systems without losing rows, breaking references or taking the business offline.
Frequently asked questions
Do we need to rebuild our system to add AI?
Almost never. In most cases AI capability is added as a service alongside your existing application and called from the workflows that need it.
What about our data - is it safe?
That depends on choices we make together: which provider, what data leaves your systems, what is retained, and whether a self-hosted model is warranted. It is a design decision made up front, and I will lay out the trade-offs plainly.
How do we know if it is actually working?
Because we measure it against an evaluation set with known correct answers, and against the time the task used to take. If the numbers do not justify the running cost, the honest recommendation is to stop.
How much does it cost to run?
Per-token costs are usually small relative to staff time, but they scale with volume and can be surprising if nothing is caching. Cost control is part of the build rather than an afterthought.
Wondering whether AI would help?
Describe the task that is eating your team's time. I will give you a straight answer on whether it is a good candidate.
Tell me about your project