What does Kraft Heinz's $208 million fraud say about agentic AI in finance?
Daniel Nakamura
Kraft Heinz: process governance and the risk of agentic AI in finance
In 2017, Kraft Heinz, an American multinational food company (kraftheinzcompany.com), needed its EBITDA, operating profit before interest, taxes, depreciation and amortization, to grow 6.1% for its teams to receive bonuses. The company delivered 3%, less than half the target.
In this article
Who benefited from Kraft Heinz's improper savings, why a number reaches consolidation with no origin, what process governance is, why only 11% of companies have put agentic AI into production and where the error shows up first.
Under pressure from the chief operating officer for savings the budget did not support, the procurement team built misleading supplier contracts to book as savings discounts that did not yet exist. The head of procurement approved the contracts. Controllership, the function that closes the company's numbers, added up what it received. When the company restated its financial statements, it corrected $208 million in improper savings, from 295 transactions made between 2015 and 2018.
The fraud was deliberate. It passed through procurement, through the executive who approved it and through the controllers who added it up, because the process had no step that showed where each number came from. That step is the core of process governance, the rule that records who booked each number, on what basis and who checked it.
The case matters now because finance teams are starting to use agentic AI, AI systems that carry out tasks on their own, such as adding up and reconciling the numbers in a budget. The risk of agentic AI in finance lies in process governance, because AI speeds up whatever the process already does, right or wrong.
The incentive behind Kraft Heinz's improper savings
Before talking about character, the useful question is who gained what. The teams' bonus depended on EBITDA growing 6.1%. The chief operating officer pushed for savings that would bring the company to that target. The pressure reached procurement, where every discount booked as a saving moved the result closer to the target and, with it, the bonus.
On the other side was consolidation, the step where every department's numbers are added into the results. There, nothing traced the origin of each saving. Whoever booked the saving gained from it. Whoever added it up had neither a reason nor a way to doubt it.
3G Capital, the investment firm that controls Kraft Heinz, used zero-based budgeting, in which every expense must be justified from scratch each cycle. The practice is legitimate. At Kraft Heinz, it raised the pressure on whoever had to show savings.
The bill for the fraud came later. The company corrected $208 million in its results. In 2021, it paid a $62 million penalty to the SEC, the agency that oversees US securities markets. The chief operating officer paid a $300,000 penalty. The head of procurement paid $100,000 and was barred for 5 years from serving as an officer or director of a public company. In 2023, the settlement with shareholders cost another $450 million.
Incentives explain why the number was inflated. They do not explain why the number went through controllership exactly as it arrived.
A number reaches consolidation with no origin
Every number in a budget has an assumption, the reason to believe it. The number is "$1 million in discounts." The assumption is "the supplier agreed in writing and will deduct it from the invoice." Whoever adds things up only sees the number, unless someone sends the assumption along with it.
At Kraft Heinz, what reached controllership was the number, backed by a contract that described the discount differently. There is no record that anyone checked the assumption behind it before adding it up.
The extreme case hides the common one. Silent error tends to be more frequent than deliberate error: a discount a supplier promised over the phone, a price increase someone assumed and never wrote down. Almost every company tends to have a number everyone uses and nobody can explain. No one needs bad intentions for it to reach the results.
Now replace the person adding things up with an AI agent. It receives the number the same way, without the assumption, and adds it up. What is missing in both cases is the same process rule.
Process governance: who booked the number, on what basis and who checked it
Process governance is the rule that records, for every number in the budget, who booked it, on what basis and who checked it before it was added to the result.
It is different from approval. At Kraft Heinz there was approval: the head of procurement approved the contracts. What was missing was the record that would let someone else, in consolidation, see the assumption and check the number.
With agentic AI, the three questions stay the same. What changes is who books the number and how fast it reaches the result.
| Governance question | Kraft Heinz (2015 to 2018) | Agentic AI without process governance |
|---|---|---|
| Who booked the number? | Procurement, under pressure from the chief operating officer | The agent, with no record of what it changed |
| On what basis? | Contracts that described the discount differently | A rule or a conversation nobody kept |
| Who checked it before adding it up? | The head of procurement approved. Controllership added up what it received | Nobody, if the agent consolidates on its own |
| Pace | At the pace of people, over three years | At the pace of the agent, with more changes in less time |
The first three rows show the same failure in both columns. The difference is in the last one: AI repeats the failure faster. If the thesis is right, the lack of governance should show up in AI adoption numbers.
Only 11% of companies have put agentic AI into production
A 2025 survey, cited by Neurons Lab, an agentic AI engineering consultancy (neurons-lab.com), found that 99% of companies plan to put AI agents into production in finance. Only 11% have done so.
Intuition says the delay comes from the tools. The most cited reason in the survey lies elsewhere: data, governance and security.
The explanation is in what an agent does. It runs rules on the process that already exists. If the process records where each number came from, the agent works with that origin. If it does not, the agent adds everything up the same way, including the saving that did not yet exist.
A bank already talking about scale shows the other side. In January 2026, Lloyds Banking Group, a British bank (lloydsbankinggroup.com), called 2026 the year agentic AI moves from experimentation to use at scale. The same bank says it deploys its agents with rigorous testing, human oversight and real-time monitoring.
Now picture Kraft Heinz with an agent in consolidation. The scheme lasted about three years, from 2015 to 2018, at the pace of people signing contracts. An agent that consolidates in minutes, with no tracing of origin, would have grown the same pattern to the same size in months.
Where the error shows up first, with and without governance
People planning to bring an agent into finance tend to worry about whether it will get the math right. After Kraft Heinz, what matters more is where the error will show up.
When every number has an owner and a record, the error shows up at its source, on the desk of whoever booked it, before it reaches the result. Without tracing, it shows up years later, in an audit, already added to everything else.
How BudgetXpert would do it differently
Only 11% of companies that plan to use agents in finance have put them into production. Before joining that count, it is worth asking whether today's process could handle an agent consolidating numbers on its own. At Kraft Heinz, the three governance questions went unanswered. On BudgetXpert, each one would have a recorded answer. With BudgetXpert, Kraft Heinz would have seen the origin of each number before adding it up, because every assumption has an owner, only people with permission change the value and the discussion stays next to the number, in the cell.
| Governance question | What was missing at Kraft Heinz | What BudgetXpert does |
|---|---|---|
| Who booked the number? | A record of who booked each saving, visible in consolidation | Every assumption has an owner and only people with permission can change it. Everything the AI creates or changes is recorded, with author and date |
| On what basis? | The assumption behind the discount, which the contract described differently | The record keeps the change made and the conversation that originated it. The discussion stays next to the number, in the cell |
| Who checked it before adding it up? | A check of the assumption in controllership, before the sum | Whoever consolidates sees the owner and the record of each number and knows whom to ask. Each user's AI spending is controlled, with the option to switch the module off |
See at budgetxpert.ai how each number records who entered it, on what basis and who checked it.
Source: cited by Neurons Lab (neurons-lab.com), an agentic AI engineering consultancy, "Agentic AI in Financial Services: A Research Roundup for 2026" (2025 survey). Lloyds Banking Group (lloydsbankinggroup.com), British bank, "2026: The year of Agentic AI, and a new era for finance" (January 21, 2026). Kraft Heinz case: SEC, "SEC Charges The Kraft Heinz Company and Two Former Executives for Engaging in Years-Long Accounting Scheme" (2021). The D&O Diary (dandodiary.com), a publication on securities litigation, "Kraft Heinz Securities Litigation Settles for $450 Million" (2023). Journal of Forensic and Investigative Accounting, an academic forensic accounting journal, "Kraft Heinz Company and the $15.4 Billion" (2023).