In February 2026, Harvard Business Review published a piece called "To Thrive in the AI Era, Companies Need Agent Managers." It was written by Suraj Srinivasan, a professor at Harvard Business School, and Vivienne Wei, the chief operating officer of the United Agentforce Platform at Salesforce.
Their argument is simple. AI used to answer questions. Now it does work. Which means somebody has to decide what these things should be doing, check what comes back, and catch it when it is confidently wrong. That somebody is the agent manager.
Here is the line that matters most for you. The authors are direct that the best agent managers are not engineers. They are project managers, operations leads, quality analysts and domain experts. People who already know what good work looks like in a specific business.
This is a judgment job, not a coding job. And judgment is not something you download. It is something you have been building for years without anyone giving it a title.
Below is the seven step playbook. It is deliberately built around work you have already done, because that is your actual advantage here, and almost nobody is telling you to use it.
First, the self-audit: six competencies
The HBR article names six competencies an agent manager needs. Read them slowly, because the split is the whole point:
- Deep knowledge of business processes — you almost certainly have this
- Systems thinking — seeing how one change ripples through the rest of the operation. You have this
- Change resilience — you have lived through reorgs, new software, new bosses. You have this
- Hybrid workflow design — deciding what the machine does and what the human does. You are closer to this than you think
- AI operations literacy — knowing how agents actually behave and fail. This one is learnable in weeks
- Prompt engineering — also learnable, and far less mysterious than the name suggests
Four of six you already own. Two you can build. That ratio is the reason this guide exists.
Step 1 — Find the agent manager work already on your resume
Before you learn anything new, take inventory. Most people underestimate this by a mile, because the work never had this name attached to it.
Prompt 1 — The inventory
I am exploring a role called "agent manager." It is the person who decides what AI agents should be doing in a business, defines what good output looks like, checks the work, measures the impact, and diagnoses failures when they happen. Harvard Business Review describes the strongest candidates as project managers, operations leads, quality analysts and domain experts rather than engineers. Here is my background: [paste your resume, or 10 to 15 honest lines about what you have actually done at work] Go through it and find evidence of five things: 1. A process I owned end to end and could describe step by step 2. A time I defined the standard for what "good" looked like for other people's work 3. A time I caught an error or a quality problem before it reached a customer 4. A time I trained, onboarded or corrected someone doing work I knew well 5. A time I measured or reported on whether something was working For each one, quote the part of my background it came from, and tell me in one sentence why it maps to the agent manager role. Then tell me the single strongest piece of evidence I have, and the biggest genuine gap. Do not flatter me. If the evidence is thin somewhere, say so.
What comes back is your raw material for every step after this one. Save it somewhere you can find again.
Step 2 — Pick one process and write the spec
Do not try to become an agent manager in general. Become the agent manager of one specific thing.
Pick a process you genuinely know: the refund workflow, new client onboarding, the monthly report, invoice chasing, appointment scheduling, the intake form, first-response triage. Something with a beginning, an end, and a version you would be embarrassed to send out.
Then write the spec. This is the single most valuable document you can produce, and it is the thing companies quietly discover they do not have.
Prompt 2 — The spec for what "good" looks like
Help me write a working specification for a business process I know well, so it could be handed to an AI agent. The process: [name it] How it works today, in my own words, messy is fine: [describe the steps as you actually do them] Who it affects: [customers, staff, vendors, whoever] Produce: 1. The process broken into numbered steps, with the trigger that starts it and the condition that ends it 2. For each step, what a GOOD output looks like and what a BAD output looks like, in concrete terms 3. The inputs the agent would need access to in order to do the step at all 4. The three steps where an error would be most expensive, and why 5. The exact points where a human must review before anything moves forward 6. Every judgment call in the process that is currently in my head and written down nowhere Ask me up to five clarifying questions first if the process is unclear. Then write the spec.
Question six is the one to sit with. The judgment calls living only in your head are precisely what a company cannot automate without you, and precisely what you have never been paid extra for.
Step 3 — Build AI operations literacy the only way that works
This is competency five, and there is no shortcut. You need to have watched an agent fail in front of you, more than once, so that you stop being impressed and start being useful.
Take the process from step 2 and run it, by hand, in whatever AI tool you already have. Give it the spec. Ask it to do a real instance of the work. Do this ten times across two weeks with real, varied examples, including the awkward ones.
Pay attention specifically to:
- Where it is confidently wrong, which is far more dangerous than where it says it does not know
- What it invents when the input is incomplete
- Which instruction it quietly drops when you give it too many at once
- How the quality changes when the example is messy rather than clean
- What it does at the edges: the refund that is out of policy, the customer who is both angry and right
If you are not sure which tool or which level you are at, start here: The Three Levels of AI. You want to be operating at level two before you call yourself literate at level three.
Step 4 — Keep a failure log
This is the step that turns a fortnight of experimenting into evidence, and it is the one nearly everyone skips.
Every time the output is wrong, log five columns: what you asked, what came back, what was wrong with it, what you think caused it, what you changed. Keep it in a spreadsheet or a plain document. Ten to fifteen entries is plenty.
Do this and you will have a document almost nobody applying for these roles has. "Diagnose why it broke" is one of the three things companies will pay for here, and a failure log is that skill made visible.
Prompt 3 — Turn the log into a diagnosis
Here is a log of failures from an AI agent I have been testing against a business process I know well. The process: [name it] The log: [paste your entries] Analyse it and tell me: 1. The failure PATTERNS, not the individual failures. Group them. 2. For each pattern, the likely root cause: bad instructions, missing context, missing data access, an unclear standard, or a task that should not be automated at all 3. Which of these I could fix by rewriting the spec, and which are genuine limits of the tool 4. The guardrail I should put in place for each one 5. Which single step of this process I should recommend stays fully human, and the argument for it in two sentences Write point 5 the way I would say it to a manager who wants to automate everything.
Step 5 — Put a number on it
"Proving the impact in numbers" is the second thing companies pay for, and it is where most people from non-technical backgrounds lose the room. Not because they cannot count, but because they never framed their work in those terms.
You need a before and an after. Time per instance, error rate, cost per instance, turnaround time, volume handled. Even rough numbers beat none, as long as you are honest that they are estimates.
Prompt 4 — The impact case
I tested an AI agent against a business process and I need to quantify it the way an operations or finance person would. The process: [name it] Before: [how long each instance takes today, how often it runs, how often something goes wrong, roughly what an hour of that time costs] After my testing: [what changed, what it handled well, what still needed a human] Give me: 1. A before and after comparison with the assumptions stated openly next to each number 2. The annual impact if this ran across the full volume, with the arithmetic shown 3. The three numbers a decision maker would actually care about, ranked 4. The strongest objection someone would raise against my numbers, and the honest answer 5. A five sentence version I could say out loud in a meeting Rules: do not invent figures I did not give you. Where a number is missing, write [NEED THIS] and tell me how to get it. Flag anything I have estimated so I never present an estimate as a measurement.
Point five matters more than the spreadsheet. The person who can say the number out loud, plainly, without hedging, is the one who gets handed the project.
Step 6 — Design the handoff
Hybrid workflow design, competency four, is where your judgment is worth the most, because it is the part engineers are worst at.
Using everything from steps 2 through 5, write one page that answers three questions for your chosen process:
- What runs on its own. The steps where the agent is reliable and the cost of an error is low.
- What a human checks before it goes out. The steps where an error is expensive, public, or hard to reverse.
- What never gets automated, and why. The moments that need a person: the exception, the upset customer, the judgment call that depends on context nobody wrote down.
That third category is the one to defend hardest. Anyone can argue for automating everything. Being the person who can say precisely where the line sits, and defend it with the failure log to back it up, is rare and immediately valuable.
Step 7 — Raise your hand
You now have a spec, a failure log, a diagnosis, an impact case and a handoff design, all on a process you already understand. That is a portfolio, and it took a fortnight.
If you are employed: do not ask for a new title. Ask for a meeting about one process, and bring the one page from step 6. Say what it costs today, what you tested, what broke, and where you think the line should sit. Most organisations right now have people asking to "do something with AI" and almost nobody arriving with a diagnosis. Titles follow that, they do not precede it.
If you are looking: the title is real and it is appearing on job boards, though it is still inconsistent. Search for "agent manager," "AI operations," "agentic operations," "AI enablement" and "automation lead," and read the descriptions rather than the titles, because the same job is being posted under all of them. As one example of what the money can look like, a Customer Success Operations and AI Agent Manager posting at the health-tech company Redox listed a base range of $110,000 to $125,000. One posting is not a market, so treat it as a data point, not a benchmark.
Then write it up properly. The method for turning this into resume language a human and an AI screener both accept is here: How to Make Your Resume Unrejectable. And if you want the work itself to have a named organisation attached to it, the platforms in How to Get Real Work Experience Without a Job will give you a real project to run this playbook against.
Where to actually start, honestly
Search this topic and you will be told to take one of three certifications: Anthropic's Claude certification, OpenAI's agent-building course, or IBM's agentic workflows course on Coursera. They are all real and all perfectly good.
They are also all builder courses. They teach APIs, SDKs and frameworks. Recommending them here would contradict the entire argument of this guide, and if you are coming from operations, service, admin or sales, there is a decent chance you would bounce off them in week one and conclude the role is not for you. It is.
So start in this order instead:
- Read the HBR article itself. Here it is. It is the source everyone else is paraphrasing, and being the person in the room who read the original is a small unfair advantage.
- Do steps 2 through 4 of this playbook. Two weeks of hands-on testing with a failure log will teach you more about how agents behave than any course, because it is your process and you can tell good from bad instantly.
- Then, if you want structure, take a course. By that point you will know which questions you are actually trying to answer, which is the difference between a certificate and a skill. And a builder course will make far more sense once you have watched an agent fail on your own work.
Do not do this in reverse. The certificate first is the expensive, discouraging path, and it is the one being recommended everywhere.
What to do this week
Not all seven steps. Two.
- Run Prompt 1 against your real background. Fifteen minutes. Read what comes back honestly.
- Name the one process you will use. Write it on something. That decision is 80% of whether the rest of this happens.
Then run Prompt 2 this weekend, and start the log on Monday.
One last thing, because I think it is the real point.
Nobody is going to walk over and tell you that the thing you have been doing for fifteen years is now the qualification. The knowing-what-good-looks-like, the catching-it-before-the-client-sees-it, the this-will-cause-a-problem-in-March. It never had a name, so it never showed up on your resume, so you stopped counting it.
Start counting it. You do not need to become someone else for this one. Just one small step past who you were yesterday.