AI agents fail on messy company data, not weak technology

Companies rushing to use AI agents, programs that carry out tasks on their own, may be tripping over something dull: their own records. A sponsored opinion piece in Business Insider argues that messy, conflicting company data, not weak AI, is what sinks these projects.

The numbers cited

The piece cites Gartner, a research firm. It says 63% of organizations either lack the data management needed for AI or are not sure they have it. Gartner also predicts companies will abandon 60% of AI projects by the end of this year because their data is not ready for AI to use. The authors say the cause is not model performance, talent or cost.

Why agents make it worse

Ask a large company how many active customers it has, and sales, finance and support may each give a different number. None is necessarily wrong. They never agreed on what "active" means.

A person who spots that gap asks a colleague. An agent, the authors say, will pick an answer and act on it unless it is set up to flag doubt. That can mean misrouted shipments, duplicate outreach campaigns or unnecessary compliance flags. Errors can pile up before a human steps in.

Gathering data is not enough

A common reaction is to move everything into one big storage system. The authors say that only puts conflicting records in the same room. The harder work is deciding which record is right, how records connect, what terms mean and what an agent is allowed to do. They say this is ongoing upkeep, not a one-time cleanup.

They also point to Forrester research that names context management and error recovery as sources of agent failure. Their reading is that tweaking prompts or tuning the AI model is not enough.

Their example is a retailer that lets an agent approve refunds for delayed orders. Customer, order and contract data must agree first. The agent then refunds up to a set limit and passes the rest to a person.

Who is behind the piece

This is a sponsored post supplied by Reltio, an SAP company. The authors are Manish Sood, Reltio's founder and CEO, and Irfan Khan, president and chief product officer of SAP Data & Analytics. The piece ends by promoting Reltio's tools, so read the argument with that in mind. The material does not say how Gartner reached its figures.

What to do

• Ask each team what "active customer" means and whether the numbers match.
• Check that customer, order and contract records agree before an AI tool acts on them.
• Give any agent a clear limit, with a person handling anything above it.
• Plan to keep records current, since the problem returns as data changes.