AI Integration.
Wiring models into the workflows that already cost you money, with guardrails and evaluation that keep them honest.
40 articles · See the AI Integration service →
Your staff are already using AI you didn’t approve
Banning it drives it underground where you can’t see it. A workable policy starts by accepting it is happening.
When not to use AI at all
A studio that sells AI integration telling you where it doesn’t belong. Five cases where a model is the wrong component.
Keeping an AI feature working for three years
Nothing about it stays still: the model, your data, your business, and the prompt. Four recurring commitments that keep it honest.
How to buy AI from a vendor
Every product has AI in it now, which makes the label useless. Six questions that reveal what you are actually purchasing.
How much of your data does an AI feature actually need?
Less than the vendor implied, and less than your first architecture assumed. Scoping the corpus is the real design work.
Where AI fits in an e-commerce business
Product data, support volume and returns are the three places it pays. Recommendation engines are further down the list than you’d think.
How to switch off an AI feature
Not every feature works. Removing one cleanly is a skill, and doing it badly is how a team decides never to try again.
When your AI feature has to explain itself
A customer disputes a decision, or a regulator asks how it was made. Explainability is an architecture choice made long before the question.
Where AI fits in a professional services firm
When you sell expertise by the hour, automation threatens the billing model. Where it helps anyway, and where it genuinely doesn’t.
How to decide which AI feature to build next
Everyone has a list. Ranking it by excitement produces a portfolio of demos. Four axes that produce a portfolio of returns.
Guardrails for when the model is confidently wrong
Fluent, plausible, and incorrect is the default failure mode. Designing for it is an engineering problem, not a prompting one.
Adding AI to an existing product without a rebuild
You don’t need a new architecture. You need one seam where a model can be inserted and its output checked.
Using AI for accessibility work, carefully
It can produce alt text, captions and plain-language versions at a scale nobody was going to fund. It can also produce confident nonsense for the people who most depend on accuracy.
Why your AI feature got slower and more expensive after launch
Nothing broke. The context grew, the retries stacked up, and the prompt accumulated instructions nobody removed.
What QA looks like for an AI feature
You can’t assert an exact output, so the usual test suite doesn’t apply. What replaces it, and what stays the same.
Build your AI feature in-house or hire someone?
The build is the small part. What decides it is who maintains the thing eighteen months from now.
Where AI actually belongs in your product
Bolting a chatbot onto a homepage isn’t AI strategy. The wins come from wiring models into the workflows that already cost you money.
Building on someone else’s model
Your feature depends on a system you don’t control, that changes without asking. Four ways to hold that dependency safely.
Permissions: what your AI feature is allowed to see
Retrieval ignores your access rules unless you build them into it. The question to ask in week one, not week twenty.
How to explain an AI project to a board
They are not asking how it works. They are asking what it costs, what it risks, and how you will know if it failed.
Extracting data from documents, reliably
Invoices, forms and contracts arriving as PDFs is one of the most common real problems AI solves. The trick is knowing when it’s wrong.
Building an AI assistant for your own team first
The lowest-risk deployment in your company answers your own staff. It also builds everything a customer-facing one would need.
What to do with the data your AI feature generates
Every interaction produces a record of what was asked, what was answered, and what a human did next. Most teams throw it away.
What happens to your AI feature under load
Model calls are slow, rate-limited and metered. A traffic spike behaves very differently to one hitting ordinary code.
How to run an AI proof of concept in two weeks
Long enough to learn something real, short enough that cancelling isn’t a defeat. The shape that avoids permanent pilot purgatory.
Translating your content with AI
Good enough to make translation affordable, not good enough to publish unreviewed. Where the line falls, by content type.
Handling personal data in an AI feature
Customers type things you never asked for. Four decisions to make before that becomes a problem you can’t unwind.
Should this run in real time or overnight?
Real-time is the default assumption and often the expensive one. Much AI work is genuinely happier in a batch at 2am.
Your prompts belong in the repository
A prompt is the specification of how your feature behaves. Keeping it in a dashboard where anyone can edit it untracked is a strange choice.
The three places AI belongs in a support workflow
Everyone puts the bot at the front door. The higher-return positions are further back, where customers never see it.
Choosing a model for a feature, not for a benchmark
Public leaderboards measure something other than your workflow. Four questions that pick the right model for the job in front of you.
Chatbot, copilot, agent: what the words actually mean
Vendors use them interchangeably, which makes buying hard. The distinctions that change what you get and what it costs.
The AI feature that should have been a form
Conversation is a slow interface for structured input. Sometimes the sophisticated answer is four fields and a button.
Should this be an agent or a workflow?
Agents are the exciting answer and usually the wrong one. The distinction that decides it is how much of the path you already know.
Write the evaluation before you build the AI feature
If you can’t describe what a good answer looks like, you can’t ship the feature — you can only demo it.
Do you actually need to fine-tune a model?
Almost certainly not yet. Three cheaper things solve the problem you think fine-tuning solves, and they’re reversible.
Meeting transcription: what to do with all that text
Recording every meeting is easy now. Most companies accumulate thousands of transcripts nobody reads and call it knowledge management.
The AI pilot that never ships
It demos well, everyone is pleased, and eleven months later it is still a pilot. The reasons are organisational, not technical.
Designing the interface for when the AI fails
Most AI interfaces are designed for the successful case. The failure states are where users decide whether to trust it again.
The questions to ask before any AI project
Eight questions that take an hour to answer and predict, more reliably than anything else, whether the project will ship.