This summer I completed Implementing Agentic AI: Building Your Organizational Playbook, a program from MIT Sloan. Over three modules, each participant built a playbook for their own organization: what agentic AI means, how to implement it, and how to scale it responsibly. It ends with a recommendation written for a board.
I expected to learn about technology. I came away thinking mostly about accountability.
Most conversations about agentic AI focus on what agents can deliver. Very few focus on the discipline needed to keep them accountable. This is not a set-it-and-forget-it technology. You can launch an agent in an afternoon. Building and running the governance that keeps it accountable is a multi-year commitment, and it needs owners well beyond IT.
Here are the lessons I took away, and why I think they matter for anyone responsible for client relationships.
Pick the Workflow That Will Teach You the Most
The program made a point that stayed with me. Your first agent should not be the safest use case or the most ambitious one. It should be the one that teaches your organization the most about making agentic decisions well.
For my playbook, I chose the quarterly business review. Anyone in client success knows the drill. A client manager pulls data from several systems, works it through spreadsheets, decides what it means, and builds a presentation. Done well, it can take up to 60 hours per report.
The QBR splits into two parts. The calculation is mechanical. The meaning is where judgement lives. Which trends are significant enough to raise? Which insights matter to this particular client, given their priorities and history? How do you present a weak result so it builds trust and points to a path forward?
An agent can do the calculation quickly. The judgement calls are what force an organization to make decisions it has never written down. How accurate does the data need to be? How confident must an insight be before a client sees it? Who signs off on sensitive content? Who owns the output once it is live? A safer, internal-only workflow would let you avoid all of those questions. That is exactly why it teaches you less.
The Model Is Not Your Safety System
Jacob Andreas, one of the program’s faculty, describes an agent as a language model plus scaffolding. His point is that risk should be architected into the system around the model, not left to the model itself.
Working through my design, the riskiest point was not where I expected. It was data extraction. An agent cannot reliably tell whether it has complete data. If a system export times out halfway and the agent carries on, every insight downstream is built on a gap. The fix is a system-level control, such as comparing each extraction against historical patterns and stopping for a human when something looks off.
The broader lesson is simple. Decide in advance what the system does that the model cannot be trusted to do, and build it in.
Autonomy Is a Dial, Not a Switch
Paul McDonagh-Smith framed this well: autonomy is a dial you calibrate, not a switch you flip. The dial runs from constrained to fully autonomous, and the setting should be chosen for each task rather than for the agent as a whole.
In a multi-agent QBR workflow, data extraction can run with delegated autonomy, retrying on its own and escalating only after set thresholds. Analysis and insights start supervised and may earn more independence over successive cycles. The final client-facing report stays supervised permanently. No report reaches a client without a person signing off.
Some lines should never move, no matter how well the system performs. Knowing which ones is a leadership decision.
Every Agent Needs a Manager
This was the idea I found most useful. Kanchana Patlolla put it plainly: every agent needs a manager.
That role is not the engineer. It is someone close to the work who can tell when output is technically correct but commercially wrong. In my design, that was an analyst who knew the source data and had sat in client QBRs. The role brings genuinely new responsibilities: maintaining the rules the agent follows, monitoring whether it performs as designed, and acting as the link between the business and the technical team.
If no one is named as the agent’s manager, no one is accountable for what it produces. With client-facing work, that gap shows up in front of the client.
Measure What the Time Becomes, Not Just the Time Saved
One of the program’s adoption realities is that existing metrics will not capture the value agents create. Most KPIs were built for a different unit of work.
Hours saved is the obvious measure for an automated QBR. It is also incomplete. It tells you time was recovered. It does not tell you what happened to that time. The metric I proposed instead was reinvested relationship hours: how much freed-up time turns into client-facing activity, and whether that shows up in renewals and expansion.
For client success leaders, this is the real business case. Speed only matters if it turns into stronger relationships.
Compare Against the Real Process, Not Perfection
Tim Kraska made a point boards need to hear. Agents replace processes that people run today, and people make mistakes too. The right question is not whether the agent makes errors. It is whether the agent, with its safeguards, produces better outcomes than the process it replaces.
That comparison is honest in both directions. In my analysis, an agent would likely be weaker at first on insight quality, because it lacks the client context an experienced manager carries. But today’s insight quality also varies widely from one person to the next. An agent offers consistency and an audit trail from day one, and it improves as it is calibrated with input from the people who know the client.
Governance Can Be a Competitive Advantage
For firms that serve regulated clients, governance is not only protection. It is a selling point. Banks and other financial institutions put vendors through demanding third-party risk reviews, and AI is now part of that scrutiny. A firm that can show documented governance, named owners, and independent validation will stand out against competitors who are using AI with none of that in place.
This ties directly to how I think about client success. Trust is the foundation that determines how much a relationship can carry. How an organization governs its AI is quickly becoming part of that foundation.
What This Means for Client Success Leaders
Agentic AI will change how client teams spend their time. The organizations that benefit most will not be the ones that deploy the most agents. They will be the ones that decide, early and deliberately, where human judgement stays, who owns each agent, and how freed-up time is put back into the relationship.
A few questions worth asking now:
- Which client workflow would teach your organization the most, not just save the most time?
- Who would manage that agent day to day, and what would they stop doing to make room for it?
- Where should the dial never move, regardless of performance?
- How will you know whether recovered time is turning into retention and growth?
For a closer look at where AI fits in each layer of the client revenue model, read What AI Can and Cannot Build in Your Client Revenue Model.
Retention is earned. Expansion is engineered. Agentic AI does not change that. It raises the bar on how deliberately both need to be built.