Softcat
AI features your customers use, not just try once
AI features add a specific capability to your website or product, such as search that understands questions, an assistant, or help filling in a complex form. Each one is built for a real user need and tested for accuracy.
When to call us
When AI should earn its place
If customers struggle to find answers or finish tasks, an AI feature can help, as long as it's built for a real need.
Competitors have launched AI assistants, and you need to know what would help your users.
People can't find the right answer or finish a complex task in your product.
Your chatbot deflects questions but frustrates customers, and can't say where its answers came from.
A prototype works in the demo, but isn't reliable or affordable enough to launch.
You need evidence a feature improves conversion or cost to serve before scaling it.
Why it matters
Customers notice when AI gets it wrong
Mistakes and data worries put people off, so accuracy, clear data boundaries and a route to a person matter more than the feature list.
-
51%
of people at organisations using AI say it has had at least one negative consequence for their organisation.
Source: McKinsey, The state of AI in 2025, November 2025, opens external site
-
64%
of customers surveyed would prefer companies didn't use AI for customer service.
Source: Gartner, Customer service survey, July 2024, opens external site
-
53%
of consumers fear their personal data will be misused when companies automate interactions with AI, up 8 points in a year.
What we do
-
Start with the user problem
The task people struggle with, and whether AI is the right fix or better search, content or design would do.
-
Prototype and test
A working prototype, tested with real users and a set of real questions before anything goes live.
-
Build it properly
Integrated with your product, content and systems, using the Claude API or whichever model suits the task.
-
Measure the quality
Accuracy, sourcing, refusals, speed and cost per answer, checked before launch and monitored after.
-
Plan the hand-off
When the feature should pass someone to a person, and what that person sees when it does.
Worth knowing
Before you add AI to your product
Three things that catch teams out.
An AI feature isn't just a chatbot
The most useful ones are often quieter: better search, summaries, recommendations, or improving your product experience.
Grounding reduces mistakes, it doesn't remove them
Answering from your own content makes errors much rarer, but you still need testing, monitoring and a way for people to flag a wrong answer.
A disclaimer doesn't move the risk
Saying "AI can make mistakes" is fair. It doesn't replace testing, data controls and a route to a person.
Self-check
Is this a good AI feature?
Five questions to ask about any AI feature you're considering.
- Which task does it help with, and how many people struggle with that task today?
- What's the cost of a wrong answer, and who would catch it?
- Can it show where its answer came from?
- What happens when it doesn't know?
- How will you know, three months in, whether it's working?
Questions
The things people usually ask
Plain answers about AI features for digital products, what's involved and what we need from you. Anything else, just ask.
Will our data be used to train the model?
No. We use providers and settings that don't train on your data, and we agree with you what the feature can see.
Do you only use Claude?
We build with Claude most often, and we choose the model for the task, its cost and where the data needs to stay.
Can it run on our website as well as our product?
Yes. Concierge is our ready-made version for B2B websites.
