Summer Reading
I've spent more of this summer on family visits and sewing dress-up clothes for my grand nieces than I have at my desk, and I don't regret a minute of it. But I still make time to read, and a few pieces from the consultancies this month were worth pulling together. Here's a quick roundup, with the throughline that connects them.
ICYMI: Four Reads Worth Your Time
1. BCG: The Agentic Leadership Playbook
BCG's Mark Abraham and Neveen Awad make a distinction that ops leaders should hold onto: the organizations succeeding with AI are not asking where to deploy agents. They are asking how work itself must be redesigned. Technology follows strategy; it does not lead it. Before an organization invests in agents, it must define the business outcomes, the decision rights, and the workflows it actually wants to improve.
This is alignment in its plainest form. An agent deployed against an undefined workflow will execute precisely and improve nothing, because there was never a clear line between the strategy and the action it was meant to serve.
2. McKinsey: Cost versus Value: Managing Agentic AI System Performance
One line from David Tepper stayed with me: most organizations are measuring token cost before they have defined business success. McKinsey's argument is that AI should not be evaluated by how cheaply it runs, but by whether it improves cycle time, quality, customer experience, or revenue.
The deeper problem here is collaborative, not technical. If a team cannot articulate what success looks like before launch, no amount of measurement after the fact will produce a defensible answer. The KPI has to be signed off before the agent goes live, not reconstructed afterward to justify the spend.
3. Bain: Resilience by Design
Bain's reminder is a quieter one, and arguably the most structural: dependency on AI providers is no longer only a commercial risk. Vendor concentration, regulatory shifts, and geopolitical events can all interrupt access to the models an organization has built its operations around. Resilience must be designed in from the start, not bolted on after the first disruption.
Trust, in this context, is not a feeling. It is the confidence that the systems underneath your operation will still be standing next quarter. That requires flexibility and optionality as design choices, not as a response plan written after something breaks.
4. BCG: Why AI Agents Need an Identity, Not Just Instructions
This was my favorite of the four. Björn Ingenleuf and Vladimir Lukic argue that AI agents do not simply execute tasks; they represent the organization that deployed them. An agent built around metrics alone, without the values and judgment that define how the company actually treats people, will produce interactions that are consistent but not distinctive, efficient but not principled.
This is the clearest connection to trust of anything in this roundup. Customers do not know whether they are speaking with a person or a system trained on that person's judgment. What they experience either reflects the organization's character or quietly erodes it, one interaction at a time.
Books to Revisit
A few books came to mind while I was reading through these four pieces, and each one deserves another pass in light of them.
The Advantage by Patrick Lencioni. BCG's point about redesigning work before deploying agents only lands if the organization is healthy enough to make that redesign stick. Lencioni's case that organizational health outperforms strategy and technology is worth revisiting alongside the agentic playbook piece; a misaligned leadership team will not suddenly align because an agent was introduced into the workflow.
Team Topologies by Matthew Skelton and Manuel Pais. If agents are changing how work gets decomposed and handed off, the team structures around that work must change with it. Skelton and Pais built a framework for team boundaries and interaction modes years before agentic AI existed, and it maps cleanly onto the decision-rights question BCG raises: who owns a workflow when part of it is executed by a system rather than a person.
Thinking in Bets by Annie Duke. McKinsey's David Tepper describes agent deployment as a probability question, not a certainty question; an agent only needs to succeed often enough that verification costs less than doing the work outright. Duke's argument that good decisions and good outcomes are not the same thing is the exact discipline leaders need when they are deciding whether an agent's success rate is good enough to trust with a given task.
Unleashed by Frances Frei and Anne Morriss. The BCG piece on agent identity is fundamentally about trust: whether a system built on behalf of an organization behaves in a way people recognize and believe. Frei and Morriss's model of trust, built on authenticity, logic, and empathy, gives leaders a concrete way to evaluate whether an agent's behavior is actually trustworthy or simply efficient.
Right Kind of Wrong by Amy Edmondson. McKinsey notes that agents rarely fail on the last step; they fail several steps earlier, in ways the final output can obscure. Edmondson's distinction between basic, complex, and intelligent failure gives leaders a way to ask the right question when an agent's outcome disappoints: not simply whether it failed, but what kind of failure it was and what it reveals about the system underneath.
Bearing Check
If an agent were handling your organization's most sensitive customer interactions tomorrow, would it make the same judgment calls your best people make today? What would it need to know about how your organization actually operates — not just what it's supposed to do — to get that right?