AI Search Module ⚠️ WARNING: A Cautionary Tale
Hello all 😁
We have long wanted to enable the AI Search module within our portal, and earlier this year we were excited to begin the process.
Our DAM currently holds approximately 1.1 million assets, primarily consisting of:
- Imagery (key art and still photography)
- Logo graphics
- Content spanning both Sports and Entertainment
Bynder initiated the setup process, which required the AI to read and evaluate every file within the portal. This took close to two weeks to complete. Once finished, the module was ready to be switched on.
Where Things Went Wrong
Unfortunately, this is where things started to go sideways.
In our environment, we make heavy use of APIs for:
- Metadata synchronisation and updating
- Asset identification and publishing to downstream products
As soon as the AI module was enabled, we experienced a critical surge in API activity and webhook (SNS) notifications.
- Approximately 800,000 update events were immediately queued (against our 1.1 million asset base), with numbers still climbing
- Processing throughput was roughly 6,000 events per hour, meaning full clearance would likely take close to a week
This spike was driven by AI-generated metadata updates, as the system applied tagging across a large volume of existing assets. Each update triggered webhook notifications (asset_bank.media.meta_updated), resulting in hundreds of thousands of events flooding downstream systems.
This quickly became a major issue, as it:
- Overwhelmed dependent systems
- Put BAU asset publishing at risk
- Occurred during a critical end-of-month period, where automated archiving based on metadata updates was already in effect
Immediate Response
Working closely with Bynder Support, we took urgent action to stabilise the situation:
- Disabled the primary metadata update webhook
- Paused key AI tagging features, including:
- Text-in-image recognition
- Facial recognition tagging
These steps successfully stopped the flood of notifications.
Internally, we also:
- Stopped API listeners and syncing/publishing services
- Purged processing queues
- Gradually reinstated services
This allowed systems to stabilise and resume normal operation. BAU processes were eventually restored.
Key Limitation Identified
A core issue we identified is that all metadata updates—regardless of source—trigger the same webhook (asset_bank.media.meta_updated).
This includes:
- AI-driven updates
- Automated system updates
- Manual (human) metadata changes
Currently, there is no way to distinguish between AI-generated updates and standard updates, making it difficult to manage or filter event traffic effectively.
Ideally, there should be separation between AI-driven metadata events and standard metadata changes, allowing for better control and system resilience.
The good news is that Bynder is aware of this limitation and is exploring improvements to this flow.
Current State
- AI Search is enabled
- Text-in-image and facial recognition features are currently disabled
- We are aiming to reintroduce facial recognition at a later stage, during a quieter and more controlled period
In Summary
If your organisation or company:
- Heavily relies on Bynder and internal APIs
- Has a large asset base
- Is considering enabling the AI Search module
Plan carefully.
Engage with your Bynder CSM early and ensure:
- Your systems are prepared for a significant spike in metadata updates
- You activate the module during a period where both your platform and business operations can absorb the impact
Hopefully this helps others avoid the very frantic 48 hours we experienced—and perhaps a few extra grey hairs along the way.
All the best,
