Marketers were pasting the same 400-word prompt into a chat box forty times a day with slight edits, then pasting the output into six other tools. The AI worked; the surface around it did not exist.
STUDY 04 / 30 · ANONYMISED · NDA · AI TOOLS · MUNICH · SERIES C
AI TOOLS · MUNICH The prompt is not the interface. The workflow is.
RoleLead designer
Timeline16 weeks
Team3 designers, 2 PMs, 1 ML engineer, 2 frontend
VerticalAI content ops platform · 1,200 marketing seats
Replaced the chat box with a workflow canvas where the prompt is a reusable configured node, variants are compared side by side, and brand constraints are enforced rather than requested.
The prompt disappeared as a user-facing artefact. Output volume rose while revision rounds fell, because constraints were structural rather than pleaded for in prose.
340ASSETS SHIPPED / WEEKfrom 84
1.4REVISION ROUNDS PER ASSETfrom 4.1
2%BRAND-COMPLIANCE FAILURESfrom 31%
THE ARGUMENT
Why the obvious solution was wrong.
The study matters because the product problem was reframed before the interface was polished.
Chat interfaces are a reasonable first surface for an unknown capability and a poor permanent one for a known workflow. Our diary study found the median marketer's prompt was 94% identical between uses; the 6% that changed was product name, audience and channel. They were hand-editing a template forty times a day because the product had never noticed the template existed.
The canvas makes the stable part a configured node and the variable part a set of inputs. Brand constraints - tone, banned claims, mandatory disclosures, reading level - are attached to the workspace and enforced at generation, so a non-compliant output cannot be produced rather than being caught in review. Variants generate in parallel and compare on a single screen with the differences marked. Revision rounds fell not because the model improved but because the thing being reviewed had already passed the checks review used to be for.
THE INTERFACE CRAFT
The interaction, rendered as a working product surface.
The specimen below is code-native and uses the study's own design logic. The client interface remains protected.
AI TOOLS
CONFIDENCE 82%
Action preview ready
DESIGN DECISIONS
Positions we would defend.
Each decision names the principle and the product consequence, not a stylistic preference.
Kill the chat box
A chat box in a workflow product is an admission that the workflow was never mapped. We mapped it.
Compare, don't scroll
Sequential variant review makes users pick the last one they read. Parallel comparison with marked diffs makes them pick the best one.
Constraint over instruction
Asking a model to obey brand rules in prose is a request. Enforcing them at generation is a guarantee.
PRODUCT LEADER READOUT
What transfers, and what should remain specific to this product.
A case study is useful when its operating principle travels without turning the original interface into a template.
Read the operating condition
For AI content ops platform · 1,200 marketing seats, the transferable lesson is not a copied screen. It is the condition the interface had to make legible: The stable prompt becomes a saved node with typed inputs. Users fill three fields rather than editing four hundred words. Rebuild that visibility for your own roles, risk, terminology, and operating cadence.
Protect the design rule
A chat box in a workflow product is an admission that the workflow was never mapped. We mapped it. Keep that rule in the acceptance criteria, component states, and production QA record so later visual cleanup cannot erase why the interaction exists.
Measure behaviour after ship
The evidence record is 340 for assets shipped / week, from 84. Recreate the baseline and outcome window before rollout, segment the result by role and context, and state clearly what the measure cannot prove.
RESEARCH RECORD
The work behind the interface.
These artefacts connect the final interaction back to the evidence and product model that produced it.
Prompt diary study
Collected 1,400 real prompts across 22 marketers; token-level diff analysis showed 94% median stability.
Workflow map
Traced the full path from brief to published asset across six tools; the AI occupied 11% of elapsed time and 0% of the surface design.
Constraint taxonomy
Worked with brand and legal to convert 40 pages of guidelines into 18 machine-enforceable constraints.
Variant comparison test
Tested sequential vs parallel variant review for selection quality against a blind expert ranking.
“Nobody writes prompts here any more. They fill in three fields and review four options.”
Director of Content Ops, enterprise SaaS · under NDA
NEXT STUDY · 05 / 30 · AI TOOLS