Bring the process
An interview transcript, an SOP, a runbook, or a description of where work jams. No prior AI design required.
Workflow Building Agent
The Workflow Building Agent, powered by Compass AI, turns how work is actually done into a structured portfolio of automatable tasks. Point it at your existing process documentation and it returns discrete tasks, each with a business objective and measurable KPIs already attached.
The blank page is the real blocker
Most AI programmes do not stall on technology. They stall because nobody can say which work should be automated first, or how anyone would know it worked.
Teams either automate whatever is easiest to describe, or spend a quarter in workshops before a single agent exists. You can start decomposing real workflows immediately rather than waiting for an audit to conclude.
What it does
Captures tacit knowledge
Upload a domain-expert interview or existing process documentation. The agent reads how the work is really performed, including the judgement steps that never made it into the written procedure.
Decomposes the workflow
Identify the data sources and application endpoints, then work towards the middle to find the tasks between them. Each task should be a single model operation; chaining them forms the workflow.
Marks the human checkpoints
Decomposition also identifies where a person must authorise the next step, so semi-autonomous execution is designed in rather than bolted on.
Sets business-aligned objectives
Each proposed task arrives with a stated business objective rather than a technical description, so the value case is legible to whoever signs off.
Attaches KPIs from day one
Success metrics are generated with the task, which makes evaluation possible from the first run instead of being retrofitted after deployment.
What a task actually contains
Decomposition produces more than a list of steps. Each task is a specification that Endeavor Studio can prototype directly.
| Element | Purpose |
|---|---|
| Input and output schemas | Define what the task accepts and returns, so tasks can chain reliably |
| Attachment definitions | Declare any images or documents that travel with the input |
| Model selection | Choose the provider and model per task, on cost and performance |
| Model parameters | Temperature and sampling controls that set determinism versus creativity |
| Prompt templates | Generate prompts dynamically from structured input |
| Tool capabilities | The MCP tools the task is permitted to call |
| Evaluation metrics | How success is measured before the task is trusted |
From friction to a testable agent
Get a task portfolio
A ranked set of candidate tasks with objectives and KPIs, ready to hand to Endeavor Studio for prototyping.
Keep the expert in the loop
The people who know the work shape the decomposition, so what gets automated matches what actually happens.
Knowing when not to use a model
Part of good decomposition is recognising which tasks should not be a model call at all. Work that must produce the same output for the same input, or that runs at a volume where token cost dominates, is better served by deterministic code or a purpose-built model — named entity recognition, for instance, is often better handled by a dedicated extraction model than a general one.
Surfacing that choice early is what keeps a programme economical.
Why this stage decides the outcome
- Tasks defined with KPIs can be scored; tasks defined without them can only be argued about.
- Small tasks make governance approval fast rather than fraught.
- Business objectives attached early mean the ROI case is ready before deployment.
- Starting from real process documents avoids automating an idealised workflow nobody follows.
- It is the first stage of the Accountability Cycle, and it feeds every stage after it.
Frequently asked questions
What does the Workflow Building Agent do?
Powered by Compass AI, it reads your existing process documentation and domain-expert interviews and decomposes the workflow into discrete, automatable tasks, each with a business objective and KPIs attached.
What do we need to provide?
An interview transcript, a standard operating procedure, a runbook, or a description of where work currently jams. No prior AI design is required.
Why does decomposition matter so much?
A task should be a single model operation with defined input and output schemas. Tasks at that granularity can be tested, scored, versioned, and approved individually - which is what makes governance fast rather than fraught.
How does it decide what should not be automated with AI?
Not every task should use a model. Work that must be deterministic, has strictly defined input, or runs at a volume that makes model calls uneconomic is better implemented as code or a purpose-built model. Compass AI surfaces the choice rather than defaulting to a model for everything.
What happens after decomposition?
The task portfolio moves into Endeavor Studio for prototyping, where each task gets model selection, prompt templates, tool capabilities, and evaluation metrics, and multiple versions compete.