Case studies are more useful than abstractions because they carry the texture of a real situation: the constraints, the politics, the moment a plan met reality. This is the account of a mid-sized creative agency that decided to replace much of its stock photography spend with AI image generation over a single quarter.
The names and specifics are composited from common experiences rather than tied to one company, but the arc is faithful to how these transitions actually unfold. The interesting part is not that it worked. It is where it nearly didn't, and what the team changed to keep it on track.
What follows is the situation they started in, the decision they made, how they executed it, the outcome they could measure, and the lessons that generalized beyond their walls.
A note on why this format is worth your time. Abstract advice about image generation tends to collapse into platitudes: be specific, select carefully, mind the rights. All true, all forgettable. A case study makes those platitudes concrete by showing what they cost and what they bought in a real sequence of decisions. The value is not in copying this agency's exact moves but in seeing the shape of a transition that worked, so you can recognize the same shape in your own.
The Situation
A Stock Budget Under Pressure
The agency spent a meaningful share of its production budget on stock licenses and occasional commissioned shoots. Finance flagged it as a line item that kept growing while client budgets stayed flat. At the same time, designers complained that stock libraries felt stale and that competitors were using the same recognizable images.
The Trigger
A new client wanted a visual identity that felt distinct and slightly surreal, something stock simply could not provide without an expensive custom shoot. A designer prototyped a few options with an image generator over a weekend, and the client picked one. That accidental win forced the question: could this become a real part of the workflow rather than a one-off trick?
The Decision
Scope It Narrow First
Rather than declaring an agency-wide mandate, leadership scoped the experiment to two channels: blog and social imagery, where the stakes were lower and volume was high. High-fidelity client deliverables and anything involving real products or real people stayed off the table for the trial.
Define What Success Meant
Before starting, they agreed on what they would measure: cost per published image, turnaround time from brief to approved asset, and a simple quality rating from art directors. Without those numbers, the debate would have collapsed into taste arguments.
The Execution
Building a Small Workflow
The team established a lightweight pipeline. A designer wrote a structured prompt, generated a batch, selected candidates, and ran light retouching and compositing where needed. They standardized on a couple of tools and built an internal library of prompt patterns that produced their house style.
The prompt library turned out to be the most valuable artifact of the whole trial. Early on, every designer reinvented the same prompts and got inconsistent results. Once the team captured the patterns that reliably produced their look, new work started from a known-good base rather than a blank prompt. Onboarding a new designer to the workflow went from days of trial and error to an afternoon of reading the library and adapting examples. The lesson generalized cleanly: the institutional knowledge in image generation lives in documented prompt patterns, and a team that writes them down compounds its skill while a team that does not relearns the same lessons person by person.
Handling the Failure Cases
Early on, they hit the predictable walls: garbled text, inconsistent characters, and uncanny hands. They responded with rules rather than frustration. Text always added in design software. Recurring subjects handled with reference images. Anything that needed literal accuracy routed back to traditional methods. Codifying these rules turned a flaky tool into a dependable one.
The Outcome
What the Numbers Showed
Over the quarter, cost per published image for the targeted channels dropped substantially, and turnaround time shrank from days to hours for routine assets. The art-director quality ratings started below stock and rose steadily as the prompt library matured, eventually matching stock for the channels in scope.
What the Numbers Hid
The quality average concealed variance: the best outputs beat stock decisively, while a long tail of mediocre generations needed rejection. The real efficiency came not from generation being free but from generating many options cheaply and selecting hard. The selection discipline was the actual product.
This surprised the finance side, which had expected the savings to come from cheap images. In reality the tool fees were a rounding error. The savings came from speed and from avoiding external vendor cycles, while a meaningful share of designer time shifted from sourcing to prompting and curating. The team learned to describe the change accurately: not cheaper images, but faster iteration and more creative control, paid for with a different mix of human effort. Framing it honestly mattered, because the next budget conversation depended on the numbers matching the story told the quarter before.
The Near-Miss
Where It Almost Went Wrong
The trial nearly collapsed in week three. A designer, excited by the speed, used a generated image of a recognizable celebrity-like figure in a client mockup. It never shipped, but it surfaced in an internal review and triggered a panic about likeness rights and brand safety. For a moment, leadership considered killing the experiment entirely.
What saved it was the narrow scope. Because the trial was confined to low-stakes channels with a review step, the mistake was caught internally rather than in front of a client. The team treated the scare as data: they added an explicit rule about likeness and protected styles to their growing rulebook, and the review step that caught it became a permanent fixture.
Why the Near-Miss Mattered
The episode taught the team something the smooth weeks could not: that the tool's worst risks are not aesthetic but legal and reputational, and that those risks hide inside attractive output. A beautiful image with a rights problem is more dangerous than an ugly one, because it sails through casual review on its looks. The discipline they built afterward was a direct response to having almost learned this the hard way.
The Lessons
Scope, Measure, Codify
The transition worked because it stayed narrow, defined success in numbers, and turned failures into documented rules. Teams that skip those steps tend to either over-promise and crash or dismiss the tool after one bad batch.
Know the Boundary
The clearest lesson was the boundary. Generation owned conceptual, atmospheric, high-volume work. Traditional methods kept literal, high-fidelity, brand-critical work. Drawing that line explicitly prevented both the failures and the internal arguments.
What Generalized
Three things from this account transfer to almost any team. First, the gains came from selection discipline, not from generation being cheap; the team that selects hard wins, regardless of tool. Second, the risks that nearly sank the trial were legal and reputational, not aesthetic, so the controls worth building are about rights and review, not just quality. Third, a measured pilot with an explicit boundary beats both the over-eager mandate and the dismissive one-batch rejection. A team of any size can copy that posture, and the agency's specific numbers matter far less than its specific habits.
Frequently Asked Questions
Why start with blog and social instead of client deliverables?
Lower stakes and higher volume. Mistakes there are cheap and easy to reject, and the volume gives you enough repetitions to build prompt patterns before anything important rides on the output.
What metrics actually mattered?
Cost per published image, turnaround time from brief to approval, and an art-director quality rating. Without agreed numbers, the decision degrades into subjective taste fights that never resolve.
Did quality match stock photography?
Eventually, for the channels in scope, once the prompt library matured. The average hid wide variance, and the gains came from generating many cheap options and selecting aggressively rather than from any single perfect render.
What stayed off-limits?
Anything needing literal fidelity: real products, real people, exact brand assets, and high-stakes client deliverables. Drawing that boundary explicitly was the decision that kept the program credible.
How long did the transition take?
The meaningful results came within a single quarter, but the early weeks produced unstable output until the team codified rules for text, consistency, and accuracy. The payoff followed the discipline, not the tool alone.
Could a smaller team replicate this?
Yes, and arguably more easily, since fewer stakeholders means less politics. The core moves, scope narrow, measure honestly, codify failures, scale with any team size.
Key Takeaways
- A narrow, measured trial beats an agency-wide mandate for adopting image generation.
- Define cost, turnaround, and quality metrics before starting or the debate stays subjective.
- Real efficiency comes from generating many cheap options and selecting hard, not from free renders.
- Convert recurring failures into documented rules to make a flaky tool dependable.
- Draw an explicit boundary between conceptual work and literal, high-fidelity work.