Frida — Reviewing campaign work without turning people into full-time reviewers
Marketing work often looks finished before it is actually ready.
A poster has been designed.
A landing page has been built.
A campaign is almost ready to go live.
But someone still has to check the details.
Is the date right?
Is this the current version?
Does the CTA point to the right place?
Does the material match the approved source?
What changed since the last review?
This work matters.
It is also repetitive, easy to miss, and difficult to carry across versions.
Frida is our attempt to understand how more of that checking work can move to AI — while the final judgment stays with people.
01 —The work
Campaign review usually combines two very different kinds of work.
One is mechanical:
- finding the right source
- comparing facts
- checking repeated details
- tracking versions
- collecting evidence
The other requires judgment:
- deciding whether an exception is acceptable
- judging visual quality
- understanding business context
- deciding whether something is ready to go live
These jobs are often treated as one manual review step.
We wanted to separate them.
Let the system do more of the checking. Keep people where judgment actually matters.
02 —Start with the source
One of the earliest lessons from Frida was simple:
Checking only works when the system knows what it should check against.
So Frida does not start with the material.
It starts with the source.
That might be:
- an approved campaign brief
- a confirmed document
- a current version of a requirement
- a piece of information the user has explicitly verified
The important part is not the format.
The important part is that the system does not quietly decide what should count as true.
This became one of our broader practice patterns:
Source before output.
03 —What AI takes on
Once the source is clear, Frida can take on more of the repeatable work.
Depending on the material, that can include:
- extracting visible text
- comparing facts
- checking approved terminology
- identifying differences
- preserving evidence
- keeping checks tied to specific versions
- rechecking only what has changed
The goal is not to automate every possible judgment.
It is to remove work that people should not have to repeat manually.
04 —Made does not mean ready
A useful result from Frida was not another feature.
It was a distinction:
Made ≠ Ready
A material can exist without being ready to use.
A page can load and still contain the wrong information.
A poster can look finished and still use an outdated date.
An AI check can complete and still lack enough evidence to support a confident conclusion.
So creation and readiness need to remain separate states.
This became a broader principle in how we think about AI-assisted work.
05 —Evidence before confidence
A system saying something is wrong is not enough.
People need to understand why.
Frida therefore tries to make the important evidence visible:
- what was observed
- what the source says
- what appears different
- which version was checked
- where uncertainty remains
The goal is not to expose every technical log.
It is to give the person making the next decision enough evidence to understand the claim.
What would I need to see to trust this conclusion?
That question changed how we thought about review.
06 —The human gate
Frida does not make the final launch decision.
That remains human-owned.
People still decide things such as:
- whether an exception is acceptable
- whether the source itself needs clarification
- whether visual quality is good enough
- whether business context changes the interpretation
- whether the material is ready to move forward
The useful question is not:
Where should we put a human approval step?
It is:
What judgment is this person making here?
If the answer is unclear, the gate may not need to exist.
07 —What did not work as simply as expected
Real workflows made several things harder than they first appeared.
Batch checking was not simply a matter of adding a button.
Once multiple materials and versions were involved, questions appeared around:
- which source version every check used
- partial failures
- duplicate work
- uncertain execution
- recovery
- cost
- which results should remain valid
We also learned that interface state can create false confidence.
At one point, some materials appeared to have no saved version even though the underlying data was still there.
And when an external check fails, “failed” is not always the same as “nothing happened.”
These are small implementation details.
But together they reinforced a larger lesson:
Reliable AI work needs explicit state, evidence and boundaries.
08 —Continuity matters
A useful AI workflow should not force the user to rebuild the same context every time.
Frida therefore treats versions, previous checks and prior evidence as part of the work.
The goal is:
Continue, don't restart.
A new version should not erase the old one.
A previous result should remain understandable.
Unchanged work should not need to be rechecked without reason.
This reduces repeated setup and makes handoff easier.
But we do not call this learning yet.
09 —What we still do not know
Frida has given us stronger evidence for continuity than for learning.
The harder question is:
Does the next cycle actually get better because the previous one happened?
That would mean things like:
- less repeated context
- fewer repeated mistakes
- faster review
- better defaults
- fewer unnecessary human checks
- clearer exception handling
We do not yet have enough evidence to make that claim.
That is what we want to measure next.
In practice
People establish the source, goal and context.
AI takes on more of the repeatable checking work.
The system surfaces differences, evidence and uncertainty.
People retain the judgment required to move the work forward.
What Frida is teaching us
Frida started as a way to make campaign checking more reliable.
It has gradually become evidence for a broader idea:
AI can do more of the work without owning the final judgment.
The interesting part is not how many checks can be automated.
The more important question is:
Can we reduce total human effort while keeping the work understandable, reviewable and human-led?
That question is still open.
And that is why Frida remains part of the practice.