What we are learning as we build with AI.
Notes from real work, written for the people who run operations.
Longer reads on how AI fits into real work.
Each one starts from a problem we have worked on and ends with what we would do again.
The model isn't the product anymore.
What differentiates useful AI systems is everything built around them: context, data, tools, memory, permissions, evaluation, interfaces and operating workflows.
What we learnedThe competitive advantage is shifting from having AI to engineering systems around intelligence.
Read the perspectiveStart with the decision. Then choose the AI.
A practical way to turn a broad AI ambition into a first build that someone can use.
Read the perspectiveCompany knowledge needs context, not just search.
What to decide before connecting internal documents to an AI assistant.
Read the perspectiveDesign the handoff before the voice.
A useful voice agent needs a clear task, a clear boundary and a good way to involve a person.
Read the perspectiveShort lessons from building.
Two or three minutes each. One problem, what we saw, and what we do now.
Agents need state, not just conversation history.
Keep a small, structured record of where the work stands. Let the conversation be a log, not the memory.
Top-K is not a retrieval strategy.
Choose fewer, better passages. Filtering and ranking matter more than how many results you pass to the model.
When agent frameworks help, and when they don’t.
Start with plain code. Add a framework when you need branching, pausing for approval or rolling back, not before.
Long AI tasks don’t belong in a web request.
Put long-running work on a queue, show progress to the person waiting, and make every step safe to retry.
“The answer looks good” is not an evaluation.
Write down what a good answer must do, collect real examples, and re-run them every time the model, prompt or data changes.
Agents should not inherit every user permission.
Give each tool only the access its task needs, and ask a person before anything sensitive is changed.
Not everything we try works. Here is what we changed.
Failed assumptions are where a lot of the useful learning is.
More agents did not make the system smarter.
We expected a team of cooperating agents to beat one. It mostly added coordination problems.
What we do nowDefault to one well-equipped agent with clear tools. Split into several only when there is a real reason.
Read what we changed Didn’t workGiving the model more context made answers worse.
Large context windows are tempting. Loading every document at once cost more and picked up outdated rules.
What we do nowA large context window does not replace choosing the right information. Select first, then send.
Read what we changed Didn’t workAutomation you cannot see becomes a risk.
An agent kept retrying a failed step for hours and nobody knew. What we now require before anything touches live data.
What we do nowIf you cannot see what an automated system did and why, you cannot trust it with real work.
Read what we changedWhat we’re watching, and whether it’s ready.
A short, honest read on newer technology. We update it as our view changes.
- Ready to useWe would use it in real work today.
- Building withWe use it in prototypes and pilots.
- TestingWe are running small experiments.
- WatchingPromising, but not ready for most teams.
- Ready to use
Small, specialised models
Not every task needs the largest model. Smaller ones can run closer to your data.
Our takeWorth using now for high-volume, narrow tasks such as sorting, extraction and routing, especially where data must stay in-house.
- Building with
Shared standards for agents and tools
Open standards such as the Model Context Protocol let one tool work with many AI systems.
Our takeUseful today for exposing your own systems to AI safely. Open agent-to-agent networks are still early.
- Building with
AI that operates existing software
Models that can use screens, keyboards and forms could automate older software with no API.
Our takePromising for back-office work. Keep it in a sandbox and have a person approve anything that commits a transaction.
- Testing
Long-term memory for agents
Agents that work over weeks need structured memory that can be reviewed and corrected.
Our takeImportant for multi-day workflows, but the patterns are still settling. Keep memory visible and editable.
- Testing
Interfaces built around the task
Software that assembles the right screen for the job, instead of fixed menus and dashboards.
Our takeWorth trying in internal tools. Keep customer-facing screens predictable for now.
- Watching
AI in machines and sensors
Vision and sensor models are moving AI from screens into factories, sites and vehicles.
Our takeCamera-based inspection and safety checks are usable today. Robotics is still one to watch.