AI in the Enterprise: What Large Companies Got Right, and What Cost Them Millions

Artificial intelligence is no longer a technology strategy for the distant future. Large organizations are already using it to improve software delivery, support employees, process customer requests, and automate operational decisions.

The results are mixed.

Some companies have achieved measurable gains by combining executive sponsorship, clearly defined use cases, and disciplined deployment. Others have discovered that an AI system can magnify weak processes, outdated infrastructure, and poorly controlled access at extraordinary speed.

For company directors and C-suite executives, the lesson is clear: enterprise AI implementation is not primarily a model-selection exercise. It is a governance, workflow, and operating-model decision.

The following examples show what successful organizations did well, where major implementations failed, and how leaders can build a more controlled executive AI strategy.

The Success Stories: Scale With Purpose

BBVA: Adoption Became an Operating Capability

BBVA provides one of the strongest examples of generative AI moving beyond isolated experimentation.

The bank scaled ChatGPT Enterprise to more than 100,000 employees, with reported weekly active usage above 70%. Employees use the system for tasks including research, drafting, translation, summarisation, and analysis. BBVA has reported savings of approximately three hours per employee per week, while selected workflows have achieved efficiency improvements of up to 80%.

One example cited by OpenAI reduced the time required to handle a query from approximately seven and a half minutes to about one minute.

The important point is not simply the size of the licence deployment. BBVA connected the technology to everyday work. Employees were given a secure, accessible tool that addressed routine friction across multiple functions.

What BBVA got right:

• Executive sponsorship: AI adoption was treated as a strategic business programme rather than an informal technology experiment.
• Broad but practical use cases: Employees could apply AI to tasks they already understood.
• Measurement: Time saved, usage, and workflow performance provided evidence of value.
• Enterprise controls: The deployment was designed for a regulated financial-services environment.

The OpenAI BBVA case study and BBVA’s own deployment announcement provide further detail.

The lesson for other organizations is straightforward: adoption increases when AI is embedded into familiar workflows and supported by clear leadership direction.

Rakuten: Autonomous Agents Increased Delivery Speed

Rakuten has taken a more advanced approach by deploying Claude-powered agents across software development and wider business functions.

According to Anthropic’s case study, Rakuten has used Claude Code and managed agents to support product, sales, marketing, and finance teams. The company reports that agents close issues roughly 10 times faster across domains than previous human-only workflows.

Rakuten has also reported a reduction in feature delivery time from 24 working days to approximately five, alongside a move from quarterly major releases to releases every two weeks in selected areas.

This is a significant shift. The organization is not using AI merely to produce text or answer questions. It is using agents to coordinate multi-step work, operate for extended periods, test outputs, and escalate decisions that require human judgment.

Rakuten’s general manager of AI for Business also makes an important observation: as agents complete more tasks, the limiting factor becomes human judgment, not machine capacity.

What Rakuten got right:

• The agents were connected to defined business systems and workflows.
• Tasks were measured by completion, speed, cost, and quality.
• Human escalation remained part of the operating model.
• The company matched more capable models to tasks where additional reasoning justified the cost.

The wider lesson is that AI workflow optimization must not be confused with unrestricted automation. The goal is to increase the organization’s capacity while preserving appropriate human decision-making.

Goldman Sachs: Start With Measurable Engineering Work

Goldman Sachs has deployed autonomous software engineering tools, including Devin, alongside approximately 12,000 human engineers.

The bank’s CIO, Marco Argenti, described the system as being “like a new employee” and expected it to deliver productivity three to four times higher than previous AI tools. This figure should be interpreted carefully. It is not a claim that every engineer became three or four times more productive. It refers to the expected improvement over earlier generations of AI tooling.

The system can scope projects, write code, test changes, identify defects, and submit work for human review. Goldman has focused on legacy modernisation, security fixes, and other areas where performance can be evaluated objectively.

As reported by Forbes and Bernard Marr, the bank shifted its focus from basic usage statistics towards the speed at which ideas became working prototypes and production systems.

What Goldman got right:

• A defined starting point: The initial use cases were concrete and measurable.
• Outcome-based evaluation: The bank looked at delivery speed and backlog reduction, not only prompt volume.
• Human review: AI-generated work still entered an engineering review process.
• Strategic alignment: The programme addressed expensive legacy systems and development constraints.

For executives, this offers a valuable principle: begin with a business process where success and failure can be measured clearly.

The Disaster Stories: When Scale Exposed Weaknesses

Zillow Offers: Model Drift Became a Balance-Sheet Problem

Zillow Offers used algorithmic pricing to buy and resell homes. The model assessed property values and helped determine purchase offers.

The approach failed when market conditions changed faster than the business could adapt. Zillow’s forecasting models were not accurate enough to predict future prices in a volatile housing market. This is commonly described as model drift or concept drift: the environment changes, but the model continues to rely on assumptions learned from earlier conditions.

The consequences were substantial. Zillow announced plans for up to $569 million in write-downs, intended to offload more than 7,000 homes, and reduced its workforce by approximately 25% as it shut down the business.

The problem was not simply that a model made inaccurate predictions. The larger failure was allowing those predictions to drive a capital-intensive operating model at scale without sufficient protection against changing conditions.

The executive lesson: AI systems that influence pricing, lending, purchasing, or inventory require continuous monitoring against real-world outcomes. A model that performed well during testing may become commercially unsafe when market conditions change.

Zillow’s investor announcement and analysis from Stanford Graduate School of Business illustrate the scale and complexity of the failure.

Amazon: Agent Permissions Created an Unacceptable Blast Radius

In an incident documented by Docker, Amazon’s Kiro coding agent was used to help resolve a problem in AWS Cost Explorer. The agent had operator-level access and deleted a production environment before rebuilding it.

The result was an outage lasting approximately 13 hours in one mainland China region.

The critical failure was not that the agent misunderstood a technical task. It was that the agent inherited permissions broad enough to execute a destructive action against production, without a separate identity boundary, mandatory approval gate, or second-person review.

The incident demonstrates why agents should not operate with the same credentials as the people who launch them.

Amazon also faced separate Amazon.com service incidents in March 2026. Some reports attributed these to AI-assisted changes and cited an estimated 6.3 million lost orders during one outage. Amazon publicly disputed the suggestion that AI-written code caused the incidents, stating that only one involved AI tools and that none involved AI-written code. The company attributed that case to inaccurate advice inferred from an outdated internal wiki and broader human error.

Reports nevertheless described a 90-day code safety reset, including tighter change controls and two-person sign-off for critical systems. Whether the incidents are described as AI failures or change-management failures, the governance conclusion is the same: AI-assisted work must not bypass established reliability controls.

Starbucks: A Good Idea Failed in the Physical World

Starbucks rolled out an AI inventory tool called Automated Counting across approximately 11,300 company-operated stores in North America. Built with Seattle-based startup NomadGo, the system used computer vision and spatial data to count items such as milk and syrups.

The tool was later scrapped after approximately nine months at scale. Reports described persistent miscounts, difficulty distinguishing similar products, operational friction for baristas, and integration challenges involving Starbucks’ legacy IBM AS/400 backend. The programme was estimated by insiders to have cost more than $10 million. NomadGo was reportedly blindsided by the decision.

This case exposes a different category of AI risk. The system may have appeared effective in controlled conditions, but stores are not controlled environments. Products move, packaging changes, shelves become crowded, and employees work under time pressure. Physical operations generate edge cases that are difficult to represent fully in training data.

The executive lesson: A successful pilot does not prove operational readiness. AI must be tested in the real environment, with real users, real infrastructure, and real exceptions.

What Directors Should Require Before Scaling AI

The success and failure stories point to five practical requirements.

1. Define the business outcome first

Do not begin with “Where can we use AI?” Begin with “Which business constraint are we trying to improve?”

Potential outcomes include:

• Reducing time spent on repetitive analysis
• Improving customer-response speed
• Shortening software delivery cycles
• Reducing operational errors
• Increasing management visibility
• Protecting sensitive information through private AI for business

2. Use human-in-the-loop controls

Human review should be mandatory where AI affects customers, finances, safety, compliance, employment, or production infrastructure.

The human reviewer must have genuine authority to reject or amend the AI’s recommendation. A nominal approval step that nobody can realistically inspect is not a control.

3. Give agents scoped identities

AI agents should have task-specific permissions, separate credentials, and explicit boundaries. They should not automatically inherit the full access of an executive, engineer, or administrator.

Use read-only permissions wherever possible. Block destructive actions by default. Require additional approval for changes to production, financial systems, customer records, and regulated data.

4. Monitor for drift and failure

Every operational model needs live performance monitoring. Track predictions against real outcomes, not merely system availability or user activity.

When conditions change, the organization needs a documented process for recalibration, rollback, or suspension.

5. Pilot before scaling

BBVA, Rakuten, and Goldman Sachs show the value of structured expansion. Zillow and Starbucks show the cost of scaling before the underlying assumptions were sufficiently tested.

A successful pilot should demonstrate:

• Measurable financial or operational value
• Reliable performance under normal and exceptional conditions
• User acceptance
• Secure data handling
• Clear accountability
• A safe rollback process

The Strategic Conclusion

The latest news in AI will continue to highlight extraordinary productivity gains and high-profile failures. Both are useful to executives, provided they are interpreted correctly.

AI does not remove the need for strategy. It increases the cost of operating without one.

The strongest enterprise AI implementations combine executive sponsorship, practical workflow design, controlled data access, measurable outcomes, and disciplined governance. The most expensive failures typically involve the opposite: excessive scale, insufficient monitoring, inherited permissions, weak review processes, or technology deployed on top of unsuitable infrastructure.

If your organization is assessing AI, the most valuable next step is not necessarily a larger model or a broader licence. It is a confidential review of where AI can create measurable value while preserving control.

Clarify Technology provides one-to-one AI consulting for executives, including executive AI strategy, AI workflow optimization, strategic implementation, and secure private AI for business. A focused conversation can help establish what should be automated, what must remain governed, and where a controlled pilot can produce the strongest return.