Teams jump straight to tooling. They pick a platform, wire up a model, and only later discover they automated the wrong step, optimized a process that should have been deleted, or built something nobody trusts enough to use. A model gives you a shared order of operations so that the expensive decisions happen before the expensive engineering.
The SENSE model is a five-stage structure for taking a workflow from "this is tedious" to "this runs reliably without us." The stages are Scope, Examine, Normalize, Synthesize, and Evaluate. They are sequential, but you will loop back as you learn. The point is not ceremony. The point is to stop building automations whose foundation was never checked.
Below, each stage names what you produce, the question it answers, and the signal that tells you it is safe to move on. You can run the whole model in an afternoon for a small workflow or stretch it over a sprint for something that touches client money.
Scope: Decide What Is Worth Automating
Pick a workflow with real frequency and real pain
Automation pays back through repetition. A task done twice a year rarely justifies the build and maintenance cost. Look for work that is frequent, rule-heavy, and currently eating senior time on low-judgment steps. Scope is where you say no to tempting but low-value candidates.
Set a boundary around the workflow
Define where the process starts and ends. Automations sprawl when their edges are vague, absorbing adjacent steps until they become unmaintainable. A tight boundary keeps the build shippable.
- Name the trigger that starts the workflow and the output that ends it.
- List the steps you are explicitly leaving manual for now.
Examine: Understand the Process as It Really Runs
Trace the actual path, not the official one
The documented process and the real one usually differ. Sit with the people doing the work and trace what they actually do, including the workarounds. Hidden judgment calls are where automation most often breaks, so name them now.
Separate mechanical steps from judgment steps
Mechanical steps follow rules and automate cleanly. Judgment steps require context and tolerate exceptions. Sorting them tells you which parts a model handles and which parts should stay with a human or need a review checkpoint.
Normalize: Clean the Inputs Before You Build
Standardize the data the workflow consumes
Automation amplifies whatever you feed it. If inputs arrive in five formats, normalize them to one before the model sees them. This stage is unglamorous and is also where most reliability comes from.
Decide how unexpected inputs are handled
Real inputs include the cases your sample missed. Define a quarantine path for anything outside the normal shape so the workflow flags oddities instead of guessing. The scaling implications of this are explored in Building AI Workflow Automations That Actually Scale for Clients.
Standardize before, not inside, the model
It is tempting to ask the model to clean and interpret messy inputs in one step. Resist it. Normalizing inputs in a deterministic step before the model sees them makes the whole workflow more predictable and easier to debug. The model should reason about clean data, not spend its effort untangling formatting nobody fixed upstream.
Synthesize: Build the Smallest Working Version
Assemble the thinnest end-to-end slice
Build the narrowest path that produces a real result, then widen it. A thin vertical slice proves the whole chain works before you invest in edge cases. It also gives stakeholders something concrete to react to early.
Keep humans in the loop where stakes are high
For outputs touching contracts, money, or client communication, route the result to a person for approval first. You remove the checkpoint later, with evidence, not on faith. This staged trust mirrors the approach in Using AI Internally to Run Your AI Agency More Efficiently.
- Ship the slice to a small audience before a full rollout.
- Instrument it so you can see what it does in production.
Evaluate: Prove It Works and Decide What Is Next
Compare against the baseline you recorded
You measured how the manual process performed during Examine. Now compare. If the automation is not clearly better on time, error rate, or cost, find out why before expanding scope. Honest evaluation is what keeps a model from becoming a rubber stamp.
Set a review cadence and a retirement trigger
Automations drift. Decide when you will review this one and what condition would justify retiring it. A workflow that no longer matches the business should be retired, not maintained forever. The operational version of running these systems appears in How to Automate Your Own AI Agency Operations.
When to Apply Each Stage
Match the depth to the stakes
A low-risk internal automation can move through SENSE in a single sitting, spending most time in Synthesize. A high-stakes, client-facing automation should spend most of its time in Examine and Normalize, where the costly mistakes hide.
Loop back without restarting
SENSE is not a one-way march. Discovering a new edge case in Evaluate sends you back to Normalize. The model gives you a place to put each lesson rather than starting over. The first pass through is sequential because each stage feeds the next; later passes jump to whichever stage a new lesson belongs in.
A Worked Example of SENSE in Action
From tedious task to reliable automation
Imagine a team drowning in inbound intake forms that someone reads, categorizes, and routes by hand. In Scope, they confirm the task is frequent and rule-heavy and draw a boundary: from form submitted to ticket routed, leaving the actual resolution manual. In Examine, they trace the real path and discover the person doing it applies an unwritten rule about urgent keywords that never made it into any document.
In Normalize, they standardize the form fields and decide that anything missing a required field goes to a quarantine queue rather than being guessed at. In Synthesize, they build the thinnest slice: categorize and route, with a human approving each routing decision for the first two weeks. In Evaluate, they compare against the baseline they recorded and find the automation matches the human's categorization closely enough to remove the checkpoint for the common cases while keeping it for the rare ones.
Why the structure paid off
The valuable insight, the unwritten urgency rule, surfaced in Examine, where it was cheap to incorporate. Had the team skipped to building, they would have discovered it as a stream of misrouted urgent tickets in production. The model did not make the work clever; it made sure the cheap discovery happened before the expensive one.
- Each stage produced an artifact the next stage needed.
- The costly lesson surfaced early, where it was inexpensive to handle.
Frequently Asked Questions
How is SENSE different from just building and iterating?
Iteration without structure tends to skip the unglamorous stages, Examine and Normalize, that determine reliability. SENSE forces those decisions to the front, where they are cheap, instead of discovering them after launch, where they are not.
Do I have to do the stages in order?
The first pass should be in order, because each stage feeds the next. After that, you will jump between stages as you learn. The order matters most the first time through a new workflow.
Which stage do teams most often shortchange?
Normalize. It is the least exciting stage, so people rush it, and then their automation produces unreliable output because the inputs were never standardized. Investing here pays back across the whole build.
Can SENSE work for a tiny one-person automation?
Yes, scaled down. Even a solo build benefits from naming the trigger and output, sorting mechanical from judgment steps, and recording a baseline. The model collapses to a checklist at small scale.
How does SENSE relate to ROI analysis?
The Scope and Evaluate stages produce the numbers an ROI case needs: frequency, baseline cost, and measured improvement. Run SENSE and you have most of the business case already written.
What if Evaluate shows the automation is not worth it?
That is a successful outcome, not a failure. Killing a weak automation before it ships saves the maintenance cost and the trust you would lose when it underperforms. A model that helps you say no is doing its job.
Key Takeaways
- SENSE moves a workflow through Scope, Examine, Normalize, Synthesize, and Evaluate so the expensive decisions happen before the expensive engineering.
- Scope kills low-value candidates; Examine separates mechanical steps from judgment steps.
- Normalize is where most reliability comes from and where teams cut the most corners.
- Synthesize builds the thinnest working slice with human checkpoints on high-stakes outputs.
- Evaluate compares against a recorded baseline and sets a review and retirement cadence.
- Match the depth of each stage to the stakes, and loop back as you learn.