Artificial intelligence agents are not only changing the way organisations deploy new technology, they are also throwing new challenges to data managers as the autonomous structures place new demands on their infrastructure.
It’s not enough that AI needs good-quality data to generate and perform the most accurate and safest outputs; agents are also prompting a rethink of practicalities such as data storage and accessibility.
At A-Team Group’s AI in Capital Markets Summit in London last week, delegates heard how these and other requirements of agentic AI are causing data chiefs to go back to their pipeline drawing boards.
In a panel discussion entitled How to Achieve AI Data Readiness, leading practitioners looked at how data quality affects the quality of AI and agentic AI outcomes. The panel, which was moderated by Tina Salvage, independent data, AI and governance adviser, also looked at solutions for achieving those outcomes.Reasons Why Good Data Quality is Important
To prevent error loops across agents
Nicolas Hourcard, co-founder and chief executive of QuestDB, opened a discussion on the impacts of stale data. Old foundational data is likely to generate errors that will be felt all the way along the data pipeline. This could create a domino effect of inaccuracies that would worsen the more the data was utilised and the further it progressed along the data pipeline, potentially rendering large swathes of work unusable or – worse – unlawful.
When agents are working at speed, as they need to in order for organisations to capture the productivity benefits they offer, any errors will be looped at very low latency, making intervention extremely difficult and potentially more expensive.
To ensure regulatory compliance
Stéphane Rio, chief executive and founder of Opensee, reminded the panel that financial institutions operate within a highly regulated environment and that compliance with a wide range of rules and guidelines was not only obligatory but enforceable by law.
Risk management and other processes must have iron-clad guardrails to ensure what’s reported to regulators is accurate. To do otherwise risks regulatory censure, possible criminal action and reputational damage.
The panel agreed that organisations couldn’t afford to be satisfied with being “kind of happy” with their data quality – they need to be 100 per cent confident in its trustworthiness.
Speed and accessibility aids accuracy
This is where storage becomes important. Agents work quickly and in bursts, the panel discussed, and consequently, the data has to be stored in a way that makes it easily and quickly accessible.
Speed bumps in that process can come from many points along the pipeline but the most pressing impediment to fast delivery is a fragmented tech and storage stack, the panel agreed.
Solutions for Shaping AI Data Readiness
Formal data contracts reduce data product confusion
Duncan Cooper, founder of Duruedma Consulting and Fractional CDO, initiated a discussion on data contracts as an aid to quality assurance of data products and what they do. By insisting on detailed written data contracts instead of what have been described as “arbitrary ranking systems”, organisations can detail precisely such concepts as the definition of data ownership, usage rights and provenance.In this way, the panel heard, organisations can establish what a good data product should do and, with continual monitoring of its performance, ensure consistency in its delivery. The latter could be achieved through a “flight recorder”-style system of oversight in which all interactions with AI are logged. This will give managers visibility into what the product is achieving and offer signals on how to improve shortcomings.
Conversion standardisation ensures better data quality
For unstructured data to be ingested into systems and AI applications, it has to be transformed into a structured format. Matthew Cheung, chief executive of Ipushpull, led a discussion on the benefits of standardising that transformation early in the data process and, with heed to its metadata, and how the resulting structured data can be put to valuable use such as analytics. It can also be better monitored for anomalies, the panel was told.
This enables organisations to create “real-time lineage and audit”, the panel heard, a characteristic that will be essential as T+1 trading becomes more commonplace.
Sensitise agentic autonomy level to use case criticality
The panel looked at the concept of an “autonomy slider” in which the level of AI agents’ self-guidance is matched to the desired level of human involvement. This would see more autonomous agents applied to processes that didn’t need humans to take a prominent role in their execution, and vice versa.
The workflow example given in the discussion was a financial trade. In the case of booking simple and block trades, the level of autonomy might be taken to a high level, whereas complex OTC trades would be less automated and involve greater human input.
Subscribe to our newsletter
