A marketing agency I worked with had a problem most organizations using generative AI eventually have. When drafting content got faster but the reviewing of it was not as fast.
This ranged from campaign briefs, advertising copy and creative concept decks; someone still had to check every one of them for accuracy, for sourcing, for policy compliance, for whether client information was being handled the way it was supposed to be. The more AI-assisted content the agency put out, the more of that piled up on reviewer’s desks. This is where the problem started, the review became the bottleneck and in many situations something that should have been caught made it through anyway.
The agency did not need a tool that could produce more content but a better way to review the content it was already producing. That is the problem I built an agent I called SignOff to solve.
The problem and results
Before SignOff, only 34 percent of AI-assisted deliverables were reviewed before going out the door. After 90 days:
- Review coverage: 34% → 100%
- Average review time per flagged item: 24 minutes → 9 minutes a 62.5% reduction
- Corrections required per quarter: 11 → 2 an 82% reduction
- Monthly AI-assisted deliverable volume: 96 → 118 up almost 23%
The agency was not reviewing less to save time. It went from partial coverage to full coverage while handling more volume, not less.
Where did I start?
Before building anything, I spent time understanding how people at the agency were actually using generative AI, what they were producing and where the process was breaking down.
The review work was not simple proofreading; reviewers had to catch unverified performance claims, outdated advertising policy references, missing sources and potential problems with how client campaign data was being handled. They also had to check every deliverable against the agency’s internal guidelines, by hand, each and every time.
Employees kept hitting the same issues, the process took too long and it was not applied consistently. When reviewers had to reread every document without a focused system pointing them somewhere specific, things got missed.
That assessment told me what the agency actually needed. Not an AI that replaced reviewers, but one that could tell reviewers where to look first. I felt this was more logical and time saving.
Mapping the agent
Once I understood the problem, I mapped the agency’s workflow to figure out where an agent could add value without disrupting how people already worked.
What an agent can do is the easy part but the harder part is where the agent fits in this process. When does it get to act on its own?, when does a person need to step in for a decision? I sat down with the agency and just asked how things actually worked. How reviews happened, how something got approved, what made an issue serious enough to escalate instead of fixed on the spot. That conversation is what became the criteria and guardrails behind SignOff.
SignOff sat in the pre-delivery review step. It examined AI-assisted materials before they went to a client and evaluated them for specific governance risks, including:
- Unverified performance claims in campaign briefs
- Outdated advertising policy references
- Missing sourcing in creative concept decks
- Client-data handling concerns
- Other issues covered by the agency’s internal guidelines
It produced a risk score from 0 to 100, so reviewers knew at a glance how much attention a deliverable needed. It also generated a prioritized action list that separated what had to be corrected before sending from what needed to be escalated.
A score by itself was not going to be enough, people needed to know why something got flagged and what to actually do about it. SignOff flagged work to a named reviewer with a checklist explaining exactly what triggered the flag.
SignOff never sent anything to a client and never made the final call on whether a deliverable was approved, corrected or rejected. That stayed with the agency’s account team. Human approval was not a feature it was one of the most important guardrails, built into the system from the start.
From prototype to a working agent
After the assessment and the workflow mapping, I built the first prototype and ran the agency’s actual content through it, comparing the output against its guidelines. That told me whether SignOff was catching what mattered to this agency, not just applying a generic checklist.
Then I sat down with the people who would use it and gave them the prototype. I had them use it because I wanted to know what was useful, what was missing, what needed to change before this became part of their daily work.
The agent was operational in about two weeks. I spent another two weeks refining it, fixing what did not work and adjusting how the criteria, outputs and workflow functioned in practice. A prototype proving an idea works is different from people being able to use it consistently and that second part is where most of the refinement time went.
The final workflow came down to five steps:
- SignOff scans AI-assisted campaign briefs, advertising copy and creative concept decks.
- The material gets evaluated against the agency’s guidelines and governance criteria.
- A risk score comes back, along with the specific issues that need attention.
- Flagged items go to a named reviewer with a checklist built around what tripped the flag.
- That reviewer either fixes it, signs off on it or sends it back, before anything reaches a client.
The agency kept a governance log too, so leadership had an actual record of how AI-assisted work got evaluated and approved, not just a general sense that it happened.
What this project taught me
The biggest lesson was that a good agent starts with an operational problem, not a list of technical capabilities. I did not open with a wishlist of what the agent should do. I opened with the organization itself, where time was actually getting lost, which mistakes kept coming back, what people needed in front of them to make a better call. That is what ended up shaping SignOff’s purpose, its criteria, its guardrails and where it sat in the workflow.
Getting the actual users in the room early mattered just as much. Their feedback is what told me the output made sense, what I had missed, whether the thing fit how they worked day to day instead of how I assumed they did.
An agent can look great on paper and still be useless if nobody can act on what it produces. The risk score worked, the checklist worked; this together told reviewers where to look and why, instead of just announcing that something, somewhere, needed attention; I provided a focus.
Here is the part people tend to get wrong: oversight and efficiency are not a tradeoff. They only look that way when oversight sits outside the workflow instead of inside it. Once SignOff’s checks were built into the process, the agency reviewed work faster, and the agent still never made the final call; it went right back to reviewers.
Value
When I started out building this agent, I would always reference in my notes this is a review agent but as I started to build it, I realized that it was more than that. It turned an inconsistent review process into something defined and measurable, gave reviewers a clearer sense of where their attention belonged and left leadership with a documented history of how AI-assisted work got approved. SignOff never took anyone out of the process and to me that was an important; it fosters a partnership between AI and human judgment.
I am an executive communications strategist with experience in government, media and corporate organizations. I write about AI, the workforce and what responsible communication looks like when technology moves faster than people are ready for.