Products people open and talk to
Claude
ChatGPT
Gemini
Perplexity
✦ The AI Office · An interactive 7-minute guide
KNOW WHAT YOUR AI CAN SEE, DO, AND GET WRONG.
Follow one request through the app, model, context, tools, and human review. You’ll leave knowing what to ask, what to verify, and when to stop.
See the hidden layersMake safer choicesKeep a field guide
A useful metaphor—not a literal description. Source-backed · Reviewed 1 August 2026 · No live model is called.
THE AI OFFICE
Absolutely.
Literal? No.
APPFront desk
AI model
CONTEXT
TOOLS⌕▦↗
YOUapprove consequential actionsUseful map?Absolutely.
Literal? No.
FAMILIAR NAMES · DIFFERENT JOBS
Where do the brands fit?
One company can make an app, models, and integrations. Ask which role its product is playing in this request.
Examples for orientation—not endorsements or permanent categories.Companies that train model families
Anthropic- OpenAI
Google
Meta
Services an approved agent might use
GitHub
Slack
Notion
Figma
BEFORE WE ENTER · MAKE A PREDICTION
What would AI need to give you today’s train times?
01Who is answering?
The app runs the front desk. The model drafts the reply.
The app collects your request, settings, and available context. It passes that package to a model, then displays what the model generates.
AI APPpackages the request
AI model
DRAFTnot generated
The request has not reached the model yet.
02What can it see?
Put useful context on the desk—and keep secrets off it.
For this reply, the model combines patterns learned during training with the context the app places on its desk. App memory is stored elsewhere and must be retrieved back into that context.
CHOOSE WHAT REACHES THE DESKEvery choice changes the draft.
HOW THE DRAFT CHANGES
The model can draft something generic, but it still needs your budget and preferences.
03Where did that come from?
Current facts need a current source—not a confident guess.
The model can request a lookup, but the surrounding software performs the approved tool call and returns the result. You still need to inspect the source.
A traveler’s weekend guide
Useful for ideas and lived experience. Not reliable evidence for a current timetable or price.
Official rail operator
Use the operator’s current result for schedules and prices, retaining the link and lookup time.
VERDICTNo current source has been selected, so the draft must not invent exact times or prices.
04What can it do?
An agent is a loop with permissions and a stop sign.
An agent repeatedly asks a model what to do next, lets the surrounding harness perform approved actions, observes the result, and stops when the goal or a boundary is reached.
- 1
Waiting…
- 2
Waiting…
- 3
Waiting…
- 4
Waiting…
- 5
Waiting…
Ready. The loop has not started.
05What must I check?
The office is a useful map. It is not a responsible human.
The model can sound purposeful without understanding or accountability. Before trusting the draft, inspect the live facts, sensitive data, permissions, and trade-offs.
Your Amsterdam → Paris day trip
Choose an outbound service after confirming a total budget. Add meal preferences if they matter.
No current source has been selected, so the draft must not invent exact times or prices.
No booking made. Human approval is still required.OPEN EVERY REVIEW STAMPFluent is not the same as finished.
THE EXIT CHECK
Can you run the office safely?
Three quick decisions turn the metaphor into a habit you can use in any AI tool.
0/5room checks correct
REFERENCE SHELF · OPTIONAL
Go deeper after the lesson.
The core tour is complete. Open only the definitions you need, or inspect the primary frameworks used to maintain this guide.
01Training + weightsPrior training, not a filing cabinet
Training adjusts a model’s numerical weights so it learns statistical patterns. Those weights are not a searchable copy of the training documents, and they are different from the context supplied for one request.
02RAGTrusted research library
Retrieval-augmented generation finds relevant material in a chosen collection, adds it to the model’s context, and asks the model to answer using that material.
03APIBusiness phone line
An API gives software a defined way to exchange structured requests and responses. An AI application can use one to request data or trigger an action.
04MCPStandard office doorway
Model Context Protocol is a standard for connecting AI applications to approved tools and data. The surrounding application still controls what is exposed and executed.
05Embedding + vector databaseMeaning map + catalogue
An embedding represents aspects of meaning as numbers. A vector database stores those representations and can retrieve items that are close in meaning.
06Fine-tuningSpecialist training
Fine-tuning continues training a model on selected examples so its behavior becomes more consistent for a task, domain, or response style.