A process goes in.
An AI workflow comes out.
We take a process people run today and redesign it so software can understand what is happening, decide what comes next and take action.
The same process, before and after
BEFORE
A person does every step
AFTER
One workflow does the steps
A process goes in, an AI workflow comes out: INPUT, then UNDERSTAND, then GET DATA, then DECIDE, then ACT, then CHECK. One case travels the whole chain. The AI does every step and stops at DECIDE, where a person approves — not sure? a person decides. The case ends as: Done.
The AI checks it did what it should
done
Read, think, act
Read
Understand emails, documents, messages, images, calls and business data.
Think
Use context, rules and company knowledge to decide what should happen next.
Act
Update systems, create records, send messages, launch actions or ask a person for help.
The pieces behind an AI workflow.
Not a wall of logos. This is how software can actually run a business process — six layers, and what each one is for.
01
Understand
LLMs
They understand language, reason about information and create outputs.
Claude · GPT · Gemini · other models
Layer 1, Understand — LLMs. They understand language, reason about information and create outputs. Shown as: structured information comes out.
structured information comes out
02
Know
RAG + business knowledge
The AI needs your information, not just what a model learned during training.
RAG gives the workflow access to the right company knowledge when it needs it.
Layer 2, Know — RAG + business knowledge. The AI needs your information, not just what a model learned during training. Shown as: the workflow can read it when it needs to.
the workflow can read it when it needs to
03
Connect
APIs · MCP · Webhooks
AI becomes useful when it can reach the systems where work happens.
MCP gives AI applications a standard way to connect with tools and data.
Layer 3, Connect — APIs · MCP · Webhooks. AI becomes useful when it can reach the systems where work happens. The animation shows eight separate custom connections between the AI and eight systems being replaced by one standard connector that reaches all of them.
one way in, many systems
04
Act
Agents · APIs · Computer Use · RPA
The workflow has to do things, not only generate text.
When there is an API, the workflow connects directly. When there is not, it uses the application the way a person does.
Layer 4, Act — Agents · APIs · Computer Use · RPA. The workflow has to do things, not only generate text. Shown as: and by using the application, when there is not.
and by using the application, when there is not
05
See the physical world
Voice · OCR · Vision · Sensors
A process does not need to start on a website.
Calls, forms, scans, photos, barcodes and other physical events can start or continue a digital process.
Layer 5, See the physical world — Voice · OCR · Vision · Sensors. A process does not need to start on a website. Shown as: all on the same workflow.
all on the same workflow
06
Control
Guardrails
We define what an agent can see, decide and do — and when a person must take over.
Layer 6, Control — Guardrails. We define what an agent can see, decide and do — and when a person must take over. Shown as: autonomy needs boundaries.
autonomy needs boundaries
These tools help us build the automation.
They are not the automation itself.
A workflow breaks into technical work packages, and coding agents take one each. AI helps us build AI — faster iteration, more testing, more of the work automated. What ships is the workflow, not the tools that made it.
- CRM connector
- ERP connector
- knowledge layer
- agent logic
- tests
- observability
- dashboard
Claude Code
Agentic software development.
We use coding agents to understand codebases, create integrations, write and test software and speed up technical implementation.
Codex
Agents that build and operate software.
Useful for implementation, testing, debugging, research and repeatable technical work.
Google Antigravity
A workspace for orchestrating development agents.
Several agents can work on different parts of a technical problem at the same time.
Show the work — every figure, with its source
- Model Context Protocol — an open protocol for connecting LLM applications with external data sources, tools and capabilities, with consent, privacy and tool safety written into the specification. This page does not use the phrase the brief asks us to avoid; the animation makes the point instead.
- Retrieval-augmented generation combines retrieval from company information sources with generative models, so the workflow can act on relevant retrieved context rather than only on what a model learned in training.
- Intelligent document processing and OCR are established ways of turning documents and scanned material into machine-readable information.
- Computer-use systems can operate graphical interfaces, which makes browser, desktop and legacy-system work another automation surface alongside APIs and RPA.
- Anthropic positions Claude Code as an agentic coding environment with tool integrations including MCP; OpenAI describes current Codex models as capable of long-running agentic tasks across code, tools, research and deployment; Google describes Antigravity as an agentic development platform for launching and monitoring multiple agents.
- Deloitte's 2026 enterprise research reports that mature governance for autonomous agents remains uncommon — which is why layer six exists. Figures on this site come from the studies named beside them and were not independently verified.
Start with one process.
Show us how it works today. We will find what should stay human and what can be automated.