Potential consequences are severe when AI agents lack context

Investments in AI continue to rise

In January, research and advisory firm Gartner predicted that worldwide spending on AI will total $2.5 trillion in 2026, up 44% from $1.8 trillion in 2025, and increase to $3.3 trillion in 2027.

Simultaneously, some enterprises are putting at least some agents into production as they try to benefit from AI's potential.

With good context, an agent can map vague natural language requests to exact corporate data. Without it, the agent can only act generically and may hallucinate what it does not know.

ISG Software Research's year-end 2024 survey found that at the time, only 17% of AI initiatives reached production, according to ISG analyst David Meninger. One year later, the firm's year-end 2025 survey showed about one-third of AI projects making it into production.

Similarly, Deloitte's State of AI in the Enterprise report, published in January, found that one-quarter of the 3,200 business and IT leaders surveyed reported having put 40% of their AI projects into production. In addition, Deloitte predicted that about half of the organizations it surveyed would reach that 40% threshold by midyear.

"We are making significant progress with agents -- we have more enterprises putting agents into production -- but that still leaves a lot that aren't in production," Menninger said. "If we are doubling the number of agents going into production, we're paying more attention to the issues surrounding production quality agents."

David Menninger

In other words, many organizations are paying more attention to context to realize agentic AI's potential. However, it's imperative that they get the context aspect of development correct, Menninger continued.

If agents are asked to execute straightforward tasks such as drafting a document, accuracy isn't imperative. Mistakes can be caught, and the consequences aren't catastrophic. But if multi-agent systems are managing entire supply chains or making recommendations to medical workers, government employees or other people doing critical work, the consequences of agents acting without proper context can be severe.

"Now that we're starting to do more agentic activities … we need to get it right, and the risks are higher," Menninger said.

Real-world consequences

Enterprises are spending vast amounts of money and devoting significant time and effort to building agents because the potential benefits of deploying networks of AI agents are massive.

For example, authors from Johns Hopkins and MIT reported that teams comprised of humans and agents were 60% more productive than teams made up of humans alone. Meanwhile, a separate study conducted by Rand Group found that agents can help enterprises reduce spending by 40%.

But there are repercussions when AI tools lack appropriate context.

"Context is given to the agent to ensure accuracy," Cindi Howson, chief data and AI strategist at ThoughtSpot, told TechTarget. "In the absence of context, you get hallucinations. The agent takes its best guess, but it is just a guess."

Real-world examples of AI hallucinations causing harm include the following:

• Air Canada suffered reputational harm and was forced to refund money when an AI chatbot didn't know the company's policy related to bereavement travel and misled a passenger.
• Google's parent company, Alphabet, lost $100 billion in market value after incorrect AI-generated content was included in a promotional video.
• OpenAI's Whisper, used by tens of thousands of medical professionals to transcribe patient visits, made up entire sentences, invented medication names and attributed racially charged statements to patients with aphasia, a disorder that impairs the ability to speak.
• A California court found that agents from Perplexity may be violating state and federal laws by accessing Amazon accounts without prior authorization.

Beyond public consequences, there are private ones when agents don't have the context they need to deliver accurate outputs, Howson noted, pointing out that organizations are charged for LLM use by the number of tokens they consume.

Cindi Howson

"Happening in parallel are the token costs," she said. "The more context you give an LLM, you're going to get a better answer on the first attempt, whereas if the LLM gets it wrong, you're going to run up your token costs and it's not going to be efficient."

When an agent cannot be trusted in production, the time and money spent building it will be wasted. Not including all the tools it takes to create an AI infrastructure, the cost to develop a single agent can range from around $1,000 for simple applications to over $100,000, according to IT consulting firm Triple Minds.

Differences come down to the scope, autonomy level and depth of integration with enterprise systems, including other agents. A basic agent that responds to frequently asked questions might cost around $15,000 to build, while a custom-trained, fully autonomous agent capable of working with other agents could cost over $100,000.

"Enterprises want agents that are reliable, with responses that are grounded, properly attributed and consistent in how long they take to run," Michael Bendersky, director of research at Databricks, told TechTarget. "Speed and cost matter as much as accuracy. An agent that [takes] unnecessary steps before reaching the right answer is expensive and unpredictable at scale."

Success stories

While failure to connect agents with appropriate context can have consequences, the potential benefits of agentic AI outweigh the risks. Improved efficiency and cost savings are real when agents can instantaneously access the relevant data and business logic they need to be accurate.

Over the past few months, as data management and analytics providers have hosted their annual user conferences, they've put on a parade of companies that are among those having early success building, deploying and benefiting from agents.

In late March, during one of the spring's first conferences, analytics vendor Domo introduced Nichole Gunn, CEO of channel marketing specialist Extu. Aided by a specialist from Domo, Gunn used capabilities launched by the vendor during its conference to build an agent in less than 30 minutes that takes on time-consuming operational tasks related to Extu's customers.

Gunn, who wanted an agent that eliminated multi-step onboarding tasks that must be repeated for each Extu customer, estimated that the agent would save Extu hundreds of thousands of dollars in operational costs in its first few months in production and millions within six months.

The agent not only handles onboarding but provides updates on customers and suggests campaigns to keep customers and partners engaged, according to Gunn.

"It went from this one thing that I wanted to a multi-faceted way of saving our company a lot of money," she told TechTarget in March. "We will be using the AI agents immediately when I get home."

Throughout April and May, vendors including Alteryx, Informatica, Qlik and Tableau held events to showcase success stories. And just last month, AWS, Databricks, Microsoft and Snowflake closed out the spring conference season with events where they highlighted customers successfully building contextually aware agents and realizing some of agentic AI's potential.

One of the customers Snowflake highlighted was Accenture, a global IT services and consulting firm that is using Snowflake's platform to help its own customers manage their data and build insight-generating analytics and AI tools.

For example, a European utility company is using combined capabilities from Snowflake and Accenture to reduce queries from weeks to seconds, according to Accenture chief strategy and services officer Manish Sharma, who spoke with Snowflake CEO Sridhar Ramaswamy during the Snowflake Summit keynote address on June 1. Similarly, a U.S. manufacturer is using tools from Snowflake and Accenture to create a unified data layer for analytics and AI after its data was previously fragmented across myriad systems.

"There's a simplicity and beauty to all of this," Ramaswamy said. "We succeed when we help our customers make more money or spend less money."

Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.