by Michael Domanic, Section Head of AI

Nobody at yesterday's Agentic Enterprise Strategy Summit said deploying agents was easy.

LinkedIn's Prashanthi Padmanabhan called it a grind. Bloomberg’s Shawn Edwards said it’s taken “an incredible amount of time, effort, and money.” Cisco’s DJ Sampath said agentic strategy means “unlearning quickly.”

We spent four hours with leaders from LinkedIn, Cisco, Bloomberg, Cardinal Health, PwC and Palo Alto Networks, and I moderated a few of the sessions myself. The conversations kept circling the same eight ideas, so here they are, with what I took from each.

1. Start with the operational baseline, and let the dollars follow

Alyssa Cox runs the Transformation Management Office within Cardinal Health's Global Business Services, and her rule is that you can't start a business case with dollars (as in, we’re aiming to save $X). For their pricing-dispute research agent, her team first asked what would change operationally (how long a dispute takes, how big the backlog is, how quickly cash comes in) and only then put a number on it.

That requires knowing your baseline, which is harder than it sounds because most teams aren't tracking it. Alyssa's point was that without one, you have no case to bring to an executive and no way to tell later whether the agent delivered.

Prashanthi's team at LinkedIn did the same thing step by step for their Hiring Assistant product. They defined what value looked like at each stage, then measured the “after” for each part of the agentic workflow - not just the final result. Early signals like those matter, because retention and revenue take a long time to show up.

2. Fix the data one agent needs, not all of your data

Alyssa was refreshingly blunt about Cardinal's data. The numbers they report to the Street are in good shape, but day-to-day operating data lives across multiple ERPs and Salesforce, and each system slices customers differently depending on who it was built for. No AI can tell you which answer is right when the systems disagree about what a customer is.

Her response is not to launch a project to clean it all up. Her team instead cleans the specific subset one agent needs, such as the data behind a chat interface for pricing research, and makes that an individual data product.

3. Stop asking whether an agent is safe, and test a specific rule

Cisco's DJ Sampath described a tools paradox. An agent is only as useful as the tools it can reach, and each tool adds risk, so many enterprises respond by refusing access until they understand the risk. That's why so much agent work stays stuck in pilots. "Is it safe?" is a question nobody can answer, so it never gets answered and pilots stall out.

His practical alternative is to pick a specific rule, such as "an agent can't approve payments above $100," and see whether it holds. That tells you what observability and guardrails you actually have, and it shows the effort involved in extending them to the next workflow.

4. Writing down what "good" means is now the real product work

One of the more interesting things Prashanthi said is that product managers used to write PRDs, and now they write product policies. A policy has to include things like which qualifications are required for a role, which are preferred, and where the agent is allowed to bend. Her point was that you can't evaluate an agent without a guideline for what good looks like, and that this wasn't a PM skill two years ago.

The problem is that you have to know the quality bar before you delegate it, and humans are notoriously bad at describing what “good” looks like in their jobs. But it’s a smart exercise: Before you build an agent, write down how you would define the quality of the output and what signals “good” vs “bad.”

5. Test the whole system continuously, because quality degrades when other agents change

Shawn Edwards from Bloomberg made a point that is easy to miss if you only think about agents one at a time. When one agent changes, it can cause degradation in a connected agent and result in quality slips across the agentic chain. For this reason, Bloomberg runs end-to-end evaluations constantly, even for an agent that hasn't changed recently.

The evaluation approach has, of course, evolved with AI. At first subject matter experts typed questions and answers into spreadsheets, then LLMs took over as judges, and now agents run most of the evaluation while the experts guide it and look at the low-confidence cases.

6. Don't use "human in the loop" for everything

“Human in the loop” is one of the most common buzz-phrases in AI today. But Jenn Kosar from PwC said human in the loop isn’t realistic or advisable for many agentic processes - it slows agents down and kneecaps their benefits.

She recommends determining the right level oversight for each action an agent takes, not for the use case as a whole. One agent can have a low-risk step that needs only an after-the-fact log – and a high-risk step that needs approval beforehand.

7. Budget for the running costs, not just the build

Alyssa treats infrastructure as a fixed cost and focuses her business cases on variable costs: tokens, volume and ongoing support. It’s tempting to think you can set an agent up and walk away, but her team has to rework connections whenever Google changes its infrastructure or moves to a new model, and nobody automatically funds that for them.

The harder part is that cost used to mean headcount and a SaaS fee, and now it means tracking token usage per workflow. That means you have to build in observability from the start, not deploy first and ask, "Oops, what did that cost?" later.

8. Expect chaos, and give every team a roadmap and an expiration date

Greg's closing point was that agent chaos is coming and is a sign of success. You'll have duplicate agents, agents with the wrong model burning tokens, and agents nobody uses anymore. Locking everything down to avoid that chaos just delays the upside. Accept that you’ll have a certain level of chaos as you deploy this totally new way of working – and plan for it.

See you next week,

‍
Michael
Your fellow Head of AI

‍