Implementation guide
The four levels of AI implementation
A practical way to choose the right level for a workflow, without starting with technology or trying to automate the whole company.
Start with the workflow, not the model.
The right level depends on frequency, clarity, available information and the consequence of an error. More autonomy requires more control, observability and ownership.
01 / 04
From assistance to operation
Each level increases scope and responsibility. This is not a race to level four. The goal is to use the minimum level that produces a useful and safe result.
-
01 Assist
A person uses AI to perform one task better.
- Best when
- Inputs vary, judgment remains important, or the team is still learning where AI helps.
- Examples
-
- Prepare responses
- Summarize documents
- Extract information for review
- What to measure
- Time saved, acceptance rate and corrections required.
- Main caution
- Do not treat suggestions as final decisions.
-
02 Automate
A bounded workflow performs repeatable steps.
- Best when
- There is a clear start and finish, exceptions are recognizable, and a person can review the output.
- Examples
-
- Request triage
- Document processing
- Recurring reporting
- What to measure
- Cycle time, exceptions, rework and output quality.
- Main caution
- Starting with a boundary that is too broad.
-
03 Integrate
The workflow reads and updates several systems.
- Best when
- The process depends on data in different applications and there are clear owners for access and changes.
- Examples
-
- CRM updates
- Cross-team onboarding
- Operational exception handling
- What to measure
- Handoffs removed, data quality, integration failures and recovery.
- Main caution
- Excessive permissions and changes without a trace.
-
04 Operate
The system is monitored, maintained and improved continuously.
- Best when
- Volume and impact justify alerts, service levels, cost management and permanent operational ownership.
- Examples
-
- Agent-assisted operations
- Continuous routing and resolution
- Internal systems with human escalation
- What to measure
- Outcome quality, failures, cost per run, service level and escalations.
- Main caution
- Launching without an owner, observability or recovery plan.
Fit
Five signals of a strong first workflow
-
01
Frequent
The improvement repeats often enough to become visible.
-
02
Expensive in attention
It consumes time, concentration or coordination across several people.
-
03
Bounded
It has a boundary the team can explain and test.
-
04
Observable
You can compare time, quality, exceptions or outcomes.
-
05
Recoverable
An error can be detected, reviewed and corrected without disproportionate harm.
Test
A first cycle in four moves
-
01
Define the boundary
Choose one trigger, one output and the exceptions that remain outside.
-
02
Observe the starting point
Record volume, time, errors and reviews before building.
-
03
Test the narrowest version
Use real cases, human supervision and an output that is easy to verify.
-
04
Decide with evidence
Expand, change or stop based on observed results.
Canvas
One-workflow canvas
Before discussing tools, be able to fill these six fields on one page.
Now apply the guide to a real workflow.
The assessment takes about three minutes and gives you a directional reading of the likely level, value and caution required.