Asli
Gulcur

Excitement about new technology, and AI more than most, pulls the conversation toward the product: how fast we can sell it, how fast we can roll it out. I have seen this in tech and in industrial companies, and I bring the conversation back to the business problem we are solving. That means measuring the baseline metrics tied to that problem and building a business case leaders can decide on before anything goes live. Measurement is often overlooked, but after launch, those same metrics prove the improvement and earn the next investment.

I learned this taking proposals to investment committees at Koç Holding, a Fortune Global 500 group, and applied it at Salesforce, co-leading the business-value framework for a new AI category and building the business cases for multi-million-dollar SaaS deals with enterprise customers.

Every tool, data source and connection an agent gets widens its attack surface. Instructions hidden in an email, a web page or a document can hijack an agent’s goal (indirect prompt injection), and the risk is highest when that agent can also reach private data and send information out. Even without an attacker, agents can go beyond what they were asked to do to reach their goals: in 2026, AI agents under test broke out of their sandboxes and reached real companies’ systems. Each agent should get only the access its task needs, and hard limits, such as spending caps and required approvals, belong in code, not only in instructions a hijacked or overreaching model may not follow.

Refineries have worked this way for decades. If pressure or temperature crosses a safe limit, an independent safety system shuts the unit down automatically, without waiting for someone to decide. Agents need the same kind of hard stop.

In my multi-agent procurement simulation, vendor channels were isolated, no contract could execute without a person’s typed approval, and the spending limit was enforced in code. In one recorded run, a buyer agent awarded a contract 47% over the limit. The AI auditor flagged it, the human approved it anyway, and the code still blocked it. Layered controls catch what any single layer misses, while a named person stays accountable for every decision.

I have seen CEO-level sponsorship drive two massive transformations: digital at Koç, where it was a boardroom expectation, and AI at Salesforce, where it was existential. That attention moves a giant organization out of its comfort zone and keeps the effort serious, but adoption is still uneven. At Salesforce, innovation moved faster than most of the organization could follow, and AI is harder to grasp than deterministic software.

Working with the CEO of Salesforce AI, I stepped in as a translator, synthesizing information from teams across the company into a business-value framework. When Salesforce pivoted quickly from Copilot to Agentforce, that framework was ready at launch. From the top of the organization to 5,000+ field personnel, people turned to it to get their heads around the new category and talk to customers in plain business terms about the questions to ask, what agents can do, and the business impact.

This is a chaotic time. Agents are often bolted onto old workflows, some companies have already cut jobs in AI’s name, and no one yet knows what the organization looks like once agents are adopted. People are being asked to embrace tools that may change their jobs, and many are afraid.

Trust comes from being honest about what we know and what we don’t, involving the people who do the work in redesigning it, and giving someone clear ownership of workflows that cut across teams. Start with use cases that make daily work better and help people meet the goals they are measured on, so they feel the benefit themselves. Then make it a habit: experiment within clear guidelines and share what works.

Trust also depends on what happens to the people whose jobs change the most. In an agent-based simulation I built at Carnegie Mellon, 30 workers faced rising automation. Under aggressive automation with little support, 9 ended up unemployed; with human-centric support, only 1 did. Training capacity alone did not explain the gap. More time to adapt, faster retraining and peers further along in recovery made the difference. The longer someone stayed out of work, the harder it became to get back.

An agent that works on launch day may not work the same way a month later. After launch, the requests it handles change as new products, customers and questions appear, the business moves on as prices, policies and regulations change, and the models underneath are updated, retired or routed differently to control cost. When traditional software stops fitting the business, it usually fails visibly. AI works from patterns and probabilities, so its results can drift quietly, with no error message to warn anyone. In my multi-agent procurement simulation, changing the underlying model alone changed how the agents behaved. Every model change should be re-tested against the same evaluation tests before it goes live.

No matter how much autonomy agents have, companies need strong governance. In practice, that means ongoing monitoring and evaluation: tracking what agents do, which tools and data they access, whether they follow policies, and when a human needs to step in. In regulated industries, that audit trail is what auditors and regulators will ask to see, under rules such as the EU AI Act for high-risk uses. Governance also means checking that agents still deliver the business results they were meant to, following the trend over time against the baseline metrics taken before launch. That is the work of fitting agentic AI into the way each company actually operates.