What to Ask an Engagement Manager Who Staffs People and Agents

What Does an Engagement Manager Do When Half the Delivery Team Is Agents?

It is Thursday and the client wants the market sizing on Monday. Two analysts are staffed, plus a research agent that produced ninety pages overnight, four of which cite a report nobody can find. The engagement manager decides that night what a person rechecks, what ships, and what the client hears before they find it themselves.

That decision is the job. The title still says engagement manager at most firms, though the work is closer to what Microsoft's 2025 Work Trend Index calls an AI workforce manager, a role 28% of surveyed managers said they were already considering hiring for hybrid teams of people and agents ². The pressure behind it is not speculative either: 47% of workers report spending more time managing and directing AI than doing the work themselves, and 65% of managers expect agents to absorb at least half of their own job inside three years ¹.

Underneath the title change is a structural one. The economics of a consulting firm were built on a wide base of analysts doing research, modeling and data cleanup, and that base is exactly what generative tools reach first. Deloitte has been replacing the analyst-consultant-manager ladder with more specific titles and internal levels, and Fortune reported McKinsey's internal agent fleet growing more than 500% in eighteen months to roughly twenty thousand agents ³. The person managing that mix owns three things at once: the client relationship, the allocation of work between people and systems, and the evidence that the blended output is true.

Which Backgrounds Produce a Real Hybrid Team Orchestrator?

The obvious feeder is a consulting manager two or three years past their own analyst rotation, close enough to the work to know which steps hide the errors. The unexpected ones travel better than they look: newsroom editors who ran stringers and fact-checkers, localization program managers, clinical research coordinators, and agency producers. All four spent years shipping work they did not personally write, on a deadline, with a named checking step.

What those backgrounds share is the habit of treating verification as a scheduled task with an owner rather than a virtue. An editor does not say the copy was checked; they say who checked it, against what, and by when. That is the exact sentence structure to listen for when a candidate describes an agent-heavy workstream.

The practice behind the skill matters more than the résumé. The managers who are good at this got good by doing the work themselves first, usually on something unglamorous: rebuilding a comp benchmark with an assistant and finding the assistant had averaged two different definitions of the same metric, or drafting a client memo and discovering that the two cleanest sentences were the two with no source. They can tell you the specific failure mode of the model they use most, the one they now check first every time. Candidates who only supervised agents describe capability. Candidates who worked with them describe scars.

One adjacent pool is worth naming: people who have built or curated the material agents work from. A manager who has run a source library for an AI system already thinks in terms of provenance rather than output volume, which is most of this job.

Ask the Two Questions That Separate Real Orchestration From Performed Oversight

Question one: walk me through one engagement you staffed with both, task by task, and tell me why each task landed where it did. Question two: tell me about a time the agent output was wrong, what it cost, and what you changed afterward. Everything you need is in how those two answers are shaped.

The tells are specific. A real orchestrator gives you a task-level allocation with reasons attached, and the reasons are about consequence rather than difficulty: the client-facing number gets a human owner because being wrong there is expensive, the first-pass literature scan does not. They can name a task they deliberately kept human because the analysts would otherwise never learn it, which is the sentence firms most need and least often hear.

A performed answer stays at the level of the engagement. Tools get named, hours saved get quoted, and the failure story arrives as a general caution about hallucination with no cost, no date and no change. Watch also for the manager who verifies everything, which sounds rigorous and means they never designed the allocation at all; if a person rereads every generated line, the agent bought nothing and the analysts are proofreaders now.

Press once on escalation. Ask what happens at 9pm when an agent's output looks wrong and the partner is on a plane. The answer should describe a rule that existed beforehand, including who is allowed to stop a deliverable, which is the same discipline behind an explicit escalation path. Improvised heroics on the night are not a process, and clients only see the version that shipped.

Where Do You Find Hybrid Delivery Leads, and What Do They Cost?

Look adjacent rather than at people whose title already says this, because almost nobody's does. The reliable pools are delivery leads at implementation firms such as Slalom, Thoughtworks, EPAM and Publicis Sapient, senior managers leaving the Big Four as the ladder gets rebuilt ³, and technical program managers who have shipped an internal assistant to a business unit and lived through its first bad month.

Communities help more than job boards here. The MLOps Community Slack, Data Council, local AI engineering meetups and the alumni networks of the firms above surface people who talk about evaluation and failure rather than adoption. Conference talks are a good filter: a person who has publicly described what their agent workflow got wrong is telling you more than any résumé line.

On pay, be honest about the evidence. No published salary series exists for this title yet, because the title is barely two years old and firms are still filing the work under engagement manager. The nearest anchors are the base roles: as of mid-2026, levels.fyi reports a median total compensation of $159,500 for management consultants in the United States and $173,500 for consulting managers ⁴. Treat those as the floor of the conversation rather than the answer. Candidates who have actually run hybrid delivery hold the stronger hand and know it, and in-house transformation teams competing for the same people often pay above consulting bands to get them.

Remote norms sit in the middle. The agent-facing half of the work is entirely remote, and the client-facing half is not: expect a pattern of two or three days on the client site during discovery and delivery weeks, remote in between, with fully distributed arrangements more common for internal transformation teams than for client-billed engagements.

Hire the Engagement Manager Who Can Staff Both, and Keep Them

What closes these candidates is authority, not title inflation. They want to decide the staffing mix themselves, to have a real budget for verification time, and to know that a deliverable can be held. Give them the actual constraints in the interview, including the ones that embarrass you, and the good ones will engage rather than flinch.

What kills the offer is a firm that wants the speed without the checking. If the pitch is that agents let one manager carry twice the engagements, an experienced candidate hears an unfunded quality mandate and walks. The second killer is ambiguity about the analysts. A manager who cares about their people will ask what happens to the junior bench, and the honest answers, including a smaller bench doing harder work sooner, land better than a reassuring one nobody believes.

The third is evidence they will be judged on things nobody can observe. Say what you will measure. Client outcomes and defect escapes are fair; utilization measured against a team that is half machine is not, and the candidates you want will say so in the interview.

For the interview loop itself, keep the panel small and make one stage real work: a messy brief, an assistant, and ninety minutes. You will learn more from watching someone allocate and check than from any answer about their philosophy, which is also true when you are hiring the specialists they will manage, including a research scientist whose method has to hold up.

See how it works >