Your Hiring Req Is Now a Capacity-Planning Decision

Atlassian has created a role called Director of Capacity Planning – Human & AI, Strategic Modeling: a job whose entire remit is deciding which work goes to people and which goes to agents. The company describes the shift as moving “from managing ‘headcount’ to architecting ‘capacity’.” That reframe lands on one document before it lands anywhere else, and the document is your hiring requisition.

A req used to be a budget line. Money exists, therefore a person. It is turning into something harder: an argument that a specific unit of work is best served by a human, by an agent, or by some fraction of both. Most req forms still ask for title, level, band, and start date. None of them ask the question your board is now asking.

One clarification before we go further, because the phrase is contested. “Capacity planning” in the project-management sense means allocating people you already employ to projects you already committed to. That is not this. This is capacity planning in the workforce-design sense: deciding what kind of capacity to acquire in the first place.

Atlassian made a whole job out of deciding what to hire for

The role is real and documented in Atlassian’s own words, not just in trade coverage. In a June 24, 2026 post on Atlassian’s blog, Chief People and AI Enablement Officer Avani Prabhakar described the job as building “the frameworks and models that help leaders make confident decisions about how human talent and AI capability work together across the organization.” The thesis in the same post is the quotable part: “We need to move from managing ‘headcount’ to architecting ‘capacity’ – the total mix of human talent and agentic capability.”

Alicia Lenart, Atlassian’s VP of HR Business Partners, put the operational version to HR Executive: “Capacity is two things, right? It’s the human folks that you have, but it’s also the agents that you have.”

Two caveats, both worth saying out loud. First, Atlassian sells agents. Rovo is a product. A framework that treats agentic capability as a line item next to headcount is commercially convenient for the company publishing it. Second, the live job posting is not publicly indexed, so the role and its remit are well sourced, but its output is not. Atlassian has created and described this job. Nobody outside Atlassian has seen what it produces.

What makes the move interesting is the posture. Atlassian explicitly refused the obvious version of this. “We do not believe in mandates about AI adoption at Atlassian because that produces a fear-based culture,” Lenart told HR Executive. And Lenart’s answer to who should own the modeling is the part a 40-person company can actually use: “Who knows the work best? It’s the leaders and the managers in that space.” Not HR. Not finance. The manager holding the req.

The mandate era came first, and one of them got reversed

Atlassian is not first. It is the second act. The first act was two CEOs who shipped the conclusion of capacity planning without doing any capacity planning, and only one of those decisions survived contact with reality.

Shopify, April 7, 2025. Tobi Lütke published his own internal memo on X after it leaked. The operative sentence: “Before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI.” Strip the AI framing and this is a change to the requisition approval workflow: a burden of proof, assigned unilaterally to the requester. Efficient as a hiring brake. Close to useless as a decision procedure, because “prove this technology cannot do the job” is unfalsifiable against a technology that changes monthly.

Duolingo, the same month. Luis von Ahn’s “AI-first” memo said headcount would grow only “if a team cannot automate more of their work,” and that the company would “gradually stop using contractors to do work that AI can handle.” Same instinct, almost the same wording.

Duolingo, one month later. After public backlash, von Ahn clarified on LinkedIn: “I do not see AI as replacing what our employees do (we are in fact continuing to hire at the same speed as before).”

Do not read the walk-back as a gotcha. Read it as the best available evidence for why Atlassian built a modeling function instead of issuing an edict. Von Ahn had a policy. He did not have a model. Nobody had worked out what the agents could actually absorb, so the policy met the work, lost, and had to be retracted in public. A capacity model is the thing that prevents that specific failure. Fifteen months on, startups are still copying the Shopify memo into their own req forms, one Slack thread at a time, and still without the model.

How far has this actually gone? The honest numbers

Further than nothing. Much less far than the discourse claims. Here is the whole evidence base, with its weaknesses attached.

Now the correction, because you have probably seen the wrong version. A widely shared 2026 headline claims entry-level hiring is “down 80%” at companies adopting AI. That figure is a misrendering of the Hosseini & Lichtinger paper, which reports a 7.7 percent decline in junior employment relative to non-adopters after six quarters, rising to roughly 10 percent in one specification. There is no 80 percent finding anywhere in that paper. The real number is off by an order of magnitude from the one circulating, and the real number is still a meaningful, worrying result. You do not need the fake one.

The paper’s own framing matters too. The mechanism is seniority-biased: senior employment kept rising at the same firms. Junior roles shrank through slower hiring rather than separations, which is exactly what it looks like when reqs quietly stop being opened. If you are staffing the bottom of the org chart, we wrote separately about hiring juniors when AI eats the entry rung.

Nobody can prove the agents are working yet

Here is the counterweight, and it comes from the same vendor pushing the capacity framing. Atlassian’s State of Teams 2026 report surveyed 12,035 knowledge workers and 173 Fortune 1000 executives in January and February 2026. Two findings from it:

• 85% of knowledge workers use AI at work. Only 29% have embedded it in how they actually work.
• Only 6% of executives say they have clear examples of organization-wide AI ROI.

Sit with the second one. Six percent, published by a vendor with every incentive to report a higher one. Ninety-four executives in a hundred cannot point to evidence that the agents they already deployed are paying for themselves.

That is the sharpest answer to the reflexive “can’t an agent do it?” challenge. If you cannot demonstrate return on the agents you already run, you cannot responsibly price an open req against an agent you have not built. Not an anti-AI position. A request that both sides of the comparison meet the same evidentiary standard, which right now they do not.

What a hiring req actually is now

A req is not a budget line. A req is an assertion that a bounded unit of recurring work exists, is valuable, and currently has no owner.

Headcount was always just the default answer to that assertion, not the only conceivable one. Contract, agency, offshore, and “the existing team absorbs it” have been the alternatives for decades, and the list has always been settled by argument rather than analysis. What changed is the volume of the argument. When the alternative was a contractor, nobody made you defend the req in a board meeting. Now they do, and most req forms carry almost none of what that defense requires. Title, level, band, and start date cannot answer a single question anyone is actually asking.

The fix is not a new title. At 40 people you are not hiring a Director of Capacity Planning. You are adding five fields to a form.

The five questions to answer before you open a req

Answer these before the req goes to whoever approves budget. The point is not to block hires. It is to make the hire defensible in the room where it gets challenged, and to catch the small number of reqs that genuinely should not exist.

1. What is the unit of work? Not the title. The recurring output. “Ships the weekly billing reconciliation and resolves exceptions” is a unit of work. “Operations Analyst” is a job title wearing a unit of work as a costume. If you cannot state it in one sentence, neither a person nor an agent can be scoped against it, and the req will go vague in exactly the way that drags out time-to-fill on unclear requisitions.

2. What throughput does it need, and how spiky is it? Steady, predictable load favors a hire. Bursty load favors capacity you can turn on and off, which is the honest structural argument for agents and contractors alike. A queue that runs at 20% capacity for three weeks and 300% in the fourth is not a full-time job. It is a scaling problem.

3. What is the supervision cost? An agent is not 1.0 FTE of capacity. It is some fraction of the throughput plus the human time to review, correct, escalate, and be accountable for the output. Almost every “agent instead of a hire” plan quietly omits the second term. Write it down as a number of hours per week and watch how many proposals stop penciling out.

4. What happens when it is wrong? Cheap and reversible failure means automate freely. Irreversible, customer-facing, regulated, or safety-relevant failure means hire, and probably hire more senior than you planned. Blast radius is the single most underweighted variable in these debates.

5. What does the human option actually cost? Not your internal band. The market rate, right now, for that role in that region, at the level you need. And the adjacent question: can you actually fill it? A req you cannot close in two quarters is not really competing with an agent you can deploy in two weeks.

Answer all five and you have something better than an opinion. You also have a written record of why this role exists, which is worth more than it sounds when someone asks the same question again in six months.

Where the data for those answers already lives

Full disclosure first, because it cuts against us. We sell an applicant tracking system. Kit gets paid when reqs open, so an ATS vendor telling you some reqs should not be opened is arguing against its own revenue. Weigh that alongside everything else here, including Atlassian’s framing and every capacity-planning vendor bidding on this keyword.

That said: the person-or-agent argument is being conducted in slide decks and Slack threads with no data, at the exact moment when most of the relevant data already sits in the hiring system nobody thinks to open.

The market price of the human option (question 5). Kit’s Compensation Research addon exposes compensation_get_salary_benchmark, compensation_compare_roles, and compensation_get_market_trends, backed by ranges scraped from live job boards across 23 role clusters and 10 currencies rather than survey data with six months of lag. Coverage skews European, with strong Polish and wider EU board data plus some US sources, so check your region before you lean on it. It is a separate $29/month addon, not part of Hiring. We wrote more about why comp benchmarking belongs in the ATS.

What a stage costs in cash. Kit’s payout config lets any stage in your process carry a real amount and currency, which is how paid work samples and assignments get compensated. That makes part of your funnel unusually concrete: for those stages, you have a literal price per candidate. Most systems cannot tell you this because they never held the number.

What a stage costs in time. hiring_get_stage_details returns started_at, completed_at, and duration_seconds for each candidate’s progress through each stage. Aggregate it and “our interview loop is slow” becomes a figure you can put in a model. Be clear on what this is: the raw stage timings are there, but Kit does not ship a time-to-fill dashboard, and you should not expect one.

Where the funnel leaks. hiring_list_applications by stage, hiring_list_pending_decisions, and hiring_list_reviews give you conversion and dwell time per stage, which is the input to both the throughput question and the supervision question. The method is the same one in our guide to finding the stage that leaks.

The req as a durable artifact. Once the five questions are answered, the answers belong on the posting, not in the thread where the decision was argued. Kit’s process templates and metafields let you attach the capacity rationale to the req itself, so it is auditable later. And because all of this is exposed through MCP tools, if you do want an agent doing your capacity modeling, it can query the actual pipeline instead of inventing one.

To be exact about what Kit is not: there is no capacity-planning module, no human-versus-agent calculator, and no workforce model. Kit holds the inputs. The judgment is yours.

What to do on Monday

The evidence says something narrower than the headlines. A minority of large companies have stopped opening some entry-level reqs. Junior hiring at AI-adopting firms is down single digits, not 80 percent, and the field disagrees with itself. Almost nobody can prove their existing agents are paying for themselves. Meanwhile headcount growth still tracks revenue growth more reliably than cost-cutting does.

None of that means hire recklessly. It means the argument deserves a model instead of a mood, and the model fits on a req form. Add the five questions. Answer them from your own pipeline data rather than from vibes. Then open the req, or do not, and be able to say why either way.

Start a free Kit trial if you want the pipeline data behind those answers somewhere you can actually query it.