Ask a delivery lead to describe their week and the answer usually sounds the same, chasing status, writing it up, sending it out, and chasing it again before the next client call. The work that actually protects a project, reading where it's drifting and having the hard conversation early, gets whatever time is left over. AI should be the moment that changes, and in a lot of firms it isn't, because the tools changed while the job stayed exactly as it was written.

In Part 2 of this series, we looked at why delivery data has to live in one place before AI can do much with it. Part 1, The Accidental Manager, made the case that delegating to agents turns everyone into a manager, including people who have never managed anyone. This piece takes the most practical idea from Part 1 and works it through for one role, which is writing two job descriptions where there used to be one.

One Role, Two Job Descriptions

The easy path is to give every project manager an AI license and leave the role untouched. It feels like progress, but it puts the whole question of what the agent should do, and who checks it, on the person least likely to have time to think about it.

A better starting point is what the role exists to deliver. For a project manager in a services firm, that looks like projects finished on time and on budget, with a client who wants to work with you again. Break the role into its tasks and sort each one into automate (the agent does it and a person checks it), augment (the agent drafts and the person judges and owns the result), or human-only. What comes out is a job description for the agent and a rewritten one for the person.

Writing the Agent's Job Description

The agent's description covers three things, the work it does, the boundaries it works within, and the points where it hands off to a person. For a project manager, the work is usually the recurring administration, status report drafts, meeting notes and action items, timesheet reminders, and risk and issue log updates. The boundaries are just as important, since an agent drafting a change request is useful and an agent committing to one with a client isn't.

Those boundaries matter because AI's strengths are uneven. In a Harvard Business School field experiment with 758 Boston Consulting Group consultants, those using AI on tasks within its capabilities completed about 12% more tasks, finished roughly a quarter faster, and produced noticeably higher-quality work. On a task outside those capabilities, they were 19 percentage points less likely to reach the right answer than consultants working without it. An agent job description is a way of writing that line down, so nobody has to rediscover it on a live client project.

Rewriting the Human Job Description

The person's role doesn't shrink, it changes shape. It keeps the human-only work, client relationships, escalations and difficult conversations, scope and commercial decisions, and accountability for the outcome. It adds the management of the agent's work, setting a clear outcome, giving enough context, judging what comes back, and stepping in when something is off.

That last part is easy to underestimate. Research from Data & Society on how people oversee AI agents makes the point that approval prompts and pause buttons only help when the person can recognize a mistake and intervene before it spreads. A review step that nobody has the time or authority to act on is closer to a formality than oversight, so the rewritten role needs the four things Part 1 described, a defined outcome, authority to accept or stop agent work, preparation to judge the result, and protected time for oversight inside the role rather than on top of it.

Where the Data Lives Shapes What the Agent Can Do

An agent job description is only as good as what the agent can see. If project status lives in one tool, time in another, and client communication in a set of inboxes, the agent works from exports, and the person spends the time they saved checking it.

Cloud Coach's AI runs natively on Agentforce, on live records inside the firm's own Salesforce org, which is why much of the automate column can work as intended. AI Meeting Manager turns client meetings into tracked decisions and action items on the project record, and Project Copilot answers questions about project health and flags risk from the same live data the project manager works in. The agent's output lands where the human reviews it, rather than in a separate tool that needs reconciling.

Cloud Coach customers see about a 20% boost in team utilization on average. A redesigned role is what decides whether recovered capacity like that goes into client work or quietly fills up with more administration.

How We See It

The job description might be the most underused AI tool a services firm has. Writing two of them turns the accidental manager into a deliberate one, with clear work for the agent, clear authority for the person, and time set aside to do the managing well. Once the time comes back, the next question is what to do with it, which is where this series goes next.

Frequently Asked Questions Related to Professional Services

• How will AI agents change the project manager role in professional services?

Much of the recurring administration, such as status drafts, meeting notes, and log updates, can move to agents, while the project manager keeps client relationships, escalations, scope decisions, and accountability, and takes on directing and checking the agents' work.

• What should an AI agent's job description include?

The specific tasks the agent handles, the boundaries it can't cross, such as client commitments or commercial terms, and the points where it hands work to a person for review and acceptance.

• How do you prepare project managers to oversee AI agents?

Give them a clearly defined outcome, the authority to accept, reject, or stop agent work, the context and skills to judge results, and protected time for oversight within the role.

Sources Cited

Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani, "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality," Harvard Business School Technology & Operations Mgt. Unit Working Paper No. 24-013, September 2023 (pre-registered experiment with 758 Boston Consulting Group consultants). https://ssrn.com/abstract=4573321

Samir Passi and Ranjit Singh, "The Oversight Fallacy: Why AI Agents Require More than Humans-in-the-Loop," Data & Society, July 2026. https://datasociety.net/research-library/the-oversight-fallacy-why-ai-agents-require-more-than-humans-in-the-loop/