Brief IA
›

AI Analytics: Experts Demand Verifiable Answers

Business & Startups·Tom Levy·

AI Analytics: Experts Demand Verifiable Answers

AI Analytics: Experts Demand Verifiable Answers
Key Takeaways
1Teresa Tung recommends being able to trace the sources and transformations of data.
2Megha Kumar advocates linking automated actions to the triggers that initiate them.
3Sonal Ramrakhiani calls for maintaining existing access restrictions.
Why it matters : a mistake can impact an automated process before an analyst verifies the result.

Specialists from Accenture, IDC, and LatentView Analytics recommend making the responses of conversational analytics tools verifiable, especially when they trigger automated actions. These interfaces give employees more freedom to query business data, but they also require them to judge results that they may not necessarily be able to control.

Companies must be able to reconstruct every analysis

Simply retaining the result is not enough to understand an error, warns Sonal Ramrakhiani, CEO of LatentView Analytics. She advocates for a history that allows a team to reproduce a calculation after modifying the data used. Data specialists must examine the analytical processing; business leaders, the decisions affected.

For an automated operation, Megha Kumar, Vice President of Research for Analytics and AI at IDC, recommends establishing acceptable sources and linking the action to the triggering response. The investigation should be able to trace back to the initial question, its requester, the produced query, the model version, the permissions, and the result. An erroneous calculation or a forgotten restriction can indeed affect a process before any human verification.

Before a significant decision, the user must also understand the meaning given to their request, the definitions used, and the freshness of the data. Teresa Tung, Global Head of Advanced Data at Accenture, recommends being able to trace their origin, transformations, and lineage.

Permissions must cover all sources

An interface that brings together multiple systems can reveal inconsistencies between accesses designed for distinct audiences. Being able to consult purchasing trends should not allow access to individual customer records solely because the tool utilizes both sets. A business response may, however, obscure the requester's rights to the mobilized information.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Sonal Ramrakhiani calls for existing restrictions on sensitive information to be maintained in conversational interfaces. Teresa Tung specifies that these do not inherently neutralize controls: the risk arises from inconsistent application of identities and permissions across systems. She urges data and security teams to verify the complete journey, from the queried sources to the returned elements.

A plausible figure can mask several errors

The number of customers from the previous quarter can vary without arithmetic error, explains Teresa Tung. The sales department may include anyone who made a purchase, while finance may count only accounts with an active contract. Without an explicit definition, this difference risks being perceived as a performance gap; an initial misinterpretation can also skew subsequent processing.

With a dashboard, teams can resolve these discrepancies before dissemination; a conversational interface may entrust this arbitration to AI with each request. A response regarding quarterly sales may thus omit selected transactions or the definition of revenue retained. Data managers must make these choices explicit during tool adoption, as some errors may only become apparent once the response is utilized.

Megha Kumar highlights another risk: an incorrect join between datasets can count orders multiple times and inflate revenues, or exclude unmatched records and omit part of the activity. The result may still appear credible. She also warns that a tool may reuse information collected for another purpose or apply a filter based on a sensitive attribute, without the plausibility of the total allowing for detection.

Nick Maynard, Vice President of Research at Juniper Research, emphasizes the particular difficulty for users who are not familiar with data to recognize these flaws. Lacking a good understanding of the sources, they may accept a response that confirms their preconceived notions without skepticism.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.