Buy a tool or build a feature?
- Buy when the task is generic (meeting notes, writing help) and your data doesn't need to be connected.
- Build when AI needs your data, your rules or your systems, or when it should sit inside a workflow your team already uses.
Choosing the first use case
| Good first use case | Risky first use case |
|---|---|
| High volume, repetitive | Rare, complex judgement calls |
| Output easy for a person to check | Output hard to verify |
| Mistakes are cheap to fix | Mistakes cost money or trust |
| Data is already digital | Data is on paper or in people's heads |
What it takes to build
- Access to your data, safely: documents, databases or business systems. Connections like MCP integrations make this practical.
- Instructions and rules written for your business.
- A test set of real examples, to check quality before and after changes.
- A review step where a person approves or corrects outputs at first.
- Logging of what was asked and answered.
What it costs to run
Most AI features are paid by usage: every request to a model has a cost based on how much text goes in and out. Costs grow with volume, document size and the model you choose. Plan for model usage, hosting and the time to monitor and improve quality. Our post on monthly running costs covers the rest of the stack.Risks to plan for
- Wrong answers stated confidently. Ground answers in your data and show sources.
- Data leaving your control. Choose providers and settings that don't train on your data, and limit what's sent.
- Over-automation. Keep a person approving actions that affect customers or money.
- Costs creeping up. Set limits and watch usage.
How to know it's working
Measure the task before you start: time per item, error rate, backlog. Measure again after a month. If the team still checks every output in full, the use case or the setup needs work. See where to start with automation.What we've built
SpeakSpace is a voice recorder for Android and iOS that turns spoken thoughts into transcripts, summaries and to-dos. Park Clear, built with Codenovos Digital, reads a photo of a parking sign and tells you whether you can park. We also connect business systems to AI assistants through our MCP integration service.FAQ
Do we need our own AI model?
Almost never. Most business features use an existing model connected to your data and rules.Will AI replace staff?
In most businesses it removes repetitive work so the same team handles more. Start by freeing time, not cutting roles.How long does a first AI feature take?
A focused feature on data that's already digital often takes a few weeks to build and test.Want to find the right first AI use case? Book a free scoping call and we'll look at your workflows together.



