Human-in-the-loop marketing strategy automation
Automated HitL* marketing strategy workflow for Japanese global player eliminated 95% effort** enabling their experts to serve clients better.
We have experience how Japanese companies operate and can match while improve velocity and client's internal capability.
*: HitL: human-in-the-loop **: From baseline weeks long research to ~4 hours until viable cuts.
What the business got
Challenge
A global marketing service/agency player from Japan wanted to improve their way of working to serve more clients with same quality.
Our KPI: drop the weeks long single strategy engagement down to days.
Implement the 11 steps of complex strategic decisions, from market landscape to bias-breaking. Preserve the secret sauce of the client: their expert's judgement.
Strategy
Pilot to launch in-house tool in under 3 months and knowledge to operate and adapt it to stay in-house at the client.
High-bandwidth cadence between remote Hungarian team and Japanese colleagues, with 2 engineers augmenting and 6 Japanese on-site to deliver together until client fully owns after roll-off.
Our delivery
Implemented and deployed onto client infrastructure:
- while the researched problem evolved during engagement: 10 dedicated agents for analytical tasks had to be built.
- Real-time streaming UX keeping the experts in the loop as the content evolved by the agents.
- Architecture is LLM provider-agnostic, model is a config choice rather than coded around.
- Production scale operations, full document vectorizations, Redis sessions, async processing, event logs, CI/CD and quality validation services delivered as part of the engagement.
Outcome
4 hours instead of weeks of strategy work per client engagement. KPI overshot with the 4 hours instead of the days long expectation.
11 complex steps automated, experts stayed in control of quality.
Application operates in production for over a year now, roll-out ongoing for multiple continent's offices of the marketing agency.
The container architecture
Adaptable workflow implemented by chaining action and guardrails on a configuration-first approach enabling fast roll-out over participants.
Simplified engine for agentic automation with self-orchestration workflow steps and an event bus loop.
Workflow steps: pub-sub event triggered Kotlin services specifically designed for one of the marketing research flow step with three stages: pre-condition checks, task core and output verification.
Workflow configuration: durable stored linkage and parameters on how each step is invoked, from-to directions chains and verification conditions.
Domain stores: accumulated or integrated/ingested data to enrich for lead research, scrapes, factual stores for company knowledge to retrieve during agentic steps.