Portfolio Management in an Agentic World

Lou has overseen the portfolio for years, and nothing about the work looks any different. The reviews still happen on schedule. The templates still fill in. The numbers still reconcile. But he sees now that the existing processes have become the bottleneck now that the build system outpaces funding.

The delivery teams now build in weeks what once took quarters. The strategy that the teams and portfolio align to revises itself whenever an assumption stops holding. Between those two, Lou’s allocation sits where it has always sat, set once and defended until the cycle turns. His is the only part of the system still moving at its original speed.

In a review that quarter, Alex, a delivery leader Lou has worked alongside for years, repeated something an old mentor had said to him. He did not aim it at anyone in particular.

“Your teams can change their minds every week. Your money changes its mind once a year. So who is deciding what your company becomes?”

Nobody answered. Lou noticed, with some discomfort, that the question was addressed to his job.

Demand Versus Capacity Allocation

Lou’s first move to address the problem was to improve the articulation of the requests, and it was not a foolish one. If the money could not move faster, the objectives could at least be clearer. He sent them back to every product group with instructions to state outcomes rather than outputs and to attach a measure to each.

They came back better, and they came back quickly. Agents drafted most of them, and the drafting was good: well formed, outcome-shaped, measurable. Nothing moved. The capacity underneath those objectives had been committed before they were written and would not be revisited until the cycle turned.

That articulation became cheap is precisely what makes its limits visible. Roger Martin named the failure before any of this was automated. Writing about the gap organizations leave between an objective and its key results, he observes that “the setting of key results will have little or nothing to do with their achievement.” Desire is not a mechanism, and a well-drafted desire is still not one. An objective that no funding decision responds to is a statement of intent rather than a commitment.

Writing a better objective got cheap. Deciding and acting on one did not.

Three Clocks, No Gearing

Lou stopped studying the allocation and started studying the timing.

He wrote down when each outcome actually changed. That moment of change became the “signal.” Strategy changed when an assumption underneath it stopped holding, which was irregular and driven by evidence. Objectives changed on a quarterly rhythm. Capacity changed once a cycle, when the plan was set. Three clocks, running at three speeds, with nothing connecting them.

A system moves at the pace of its slowest clock, and Lou’s deciding and acting was the slowest. Every improvement made downstream arrived at his boundary and waited there.

Funding rhythms vary widely, and some organizations reset in months while others reset in years. The length of the cycle matters less than one question: whether capacity renews faster than the assumptions it was committed against expire. Telemetry now retires assumptions in weeks. An allocation defended for a full cycle is defending a picture of the business that stopped being accurate somewhere in the middle of it.

A Backlog Worth Approving

An objective names a capability gap. It does not tell anyone what to start.

What converts one into the other is a portfolio brief: a problem stated small enough to approve, carrying the outcome it targets, the key result it moves, and its standing against everything else already waiting. Ninety days is the outer edge of that window rather than the target, but smaller targets are even better. Len Greski sets the same ceiling in his work on software economics, funding value streams and products in outcome-gated envelopes lasting no longer than 90 days. A problem that cannot be framed inside a quarter has not yet been made small enough.

A portfolio brief must establish three things: the problem, the outcome, and the key result. The problem, sliced to the first piece a real user can respond to rather than the capability in full. The outcome, named against the people whose behavior is meant to change, and what each of them will be able to do that they cannot do today. That is the capability. And the single key result it moves, ranked against everything else in the backlog, because priority that is not comparative is not priority.

Ranking is where an agent earns its place. Scoring a backlog consistently against a stated rubric is work humans do badly, slowly, and inconsistently, and it is work a model does well. Delegating it removes the oldest distortion in portfolio management, which is that priority tends to reflect who argued hardest. What cannot be delegated is the rubric. A rubric encodes what value means to the organization right now, which makes it a decision rather than a calculation, and it expires. Agents rank against a definition of value. A real person must own the definition rather than delegate it, and revisit it before it goes stale.

Lou ran his own backlog against the key-result test first, and most of it failed to meet the priority thresholds. The items were real. The sponsors were senior. Not one of them connected to a key result anyone could name aloud. A portfolio backlog whose items cannot each name the key result they move is a list of requests wearing a strategy’s clothes. It scored low, and it was not funded.

Two Questions at the Gate

Every approval turns on two questions about capacity. The first asks whether the organization can build the thing. The second asks whether the customer it is built for can absorb it. Most portfolios ask the first and assume the second.

Is there build capacity?

Then comes the discipline that makes a backlog mean anything. A problem with no capacity behind it does not get approved. Not deferred with a nod, not approved in principle, not held pending. Not approved.

That rule changes what funding is. Funding is the available capacity to do the work. So when a problem worth solving has no capacity behind it, two honest moves remain. Reallocate capacity from somewhere or fund additional capacity against a longer view of the strategy. Somebody owns each of those decisions. Approving the work regardless leaves the decision unmade.

The third option carries a cost that appears on no ledger. Work approved without capacity behind it does not sit quietly. It keeps drawing effort from people who cannot act on it, in status asked for, plans revised, sponsors updated, and expectations maintained. Writing in Harvard Business Review, Rose Hollister and Michael Watkins list unfunded mandates among the six root causes of initiative overload. A backlog of commitments nobody can execute carries its own change load, and that load lands on the organization long before anything reaches a customer. Declining to approve removes work that was never funded in the first place.

Risk decides which problems reach this gate at all, sorted by reversibility and the cost of being wrong. A small, reversible bet inside one product group belongs to that group, decided there against capacity its own leader already holds. What travels to the portfolio is work that binds more than one group, or commits capacity somebody else is counting on. A gate everything passes through is a queue with a formal name.

Can the customer absorb it?

The second question is the one almost nobody asks, and without an answer to it the decision to commit is not executable.

Lou stopped asking what each product group had shipped and asked something else. What had one customer group received from the portfolio over the last cycle, counting every product that served them? Nobody had the answer, because every group reported its own.

This is the part that changed. Assembling that view used to take a quarter of somebody’s attention and rarely survived contact with the next reorganization of the reporting. A portfolio agent now holds every product’s release record, adoption telemetry, and support volume against a single customer and returns the aggregate in an afternoon. The synthesis is no longer expensive, which means a portfolio team can be held accountable for a view it could previously claim was impractical to produce.

The list was longer than anyone expected. Each group had reported adoption on its own releases and each number had looked fine, but nobody had added them up. Together they described a customer absorbing more change in one cycle than anyone had chosen. Every team had shipped a defensible amount. Together they had shipped too much.

Adam Whaley’s work is where the evidence for this lives. He defines feature usage as “the share of what you’ve shipped that people actually use, measured from instrumentation in the product, not from opinion in a meeting,” and the Standish research he cites puts 45% of shipped features in the never-used column, with another 19% used only rarely.

None of this is new, and none of it is about technology. Hybrid seed corn reached the American market in 1924 with a yield advantage approaching 20%, and a farmer could try it on a single acre. Bryce Ryan and Neal Gross studied 257 operators across two Iowa communities. The average interval between first hearing of hybrid seed and first planting it was five and a half years, and the first commitment was a median of 12% of a farm’s corn acreage. Absorption was slow and partial even where the arithmetic was obvious and the trial was nearly free.

Capacity to build and capacity to absorb are different resources, and only one of them can be bought.

An agentic system can tell a portfolio team that no capacity exists on either side. Only a person can decline to approve the work anyway, absorb the disappointment that follows, and answer for the choice afterward. The ranking can be delegated. The refusal cannot.

Every approval asks whether the work can be built. Almost none ask whether it can be absorbed.

Confirmation Is Not Payback

One piece of gearing remains. An approved problem runs to a period, and at the end of that period something has to happen other than the next cycle arriving.

The window buys confirmation, not payback, and the distinction is what allows the model to survive a conversation with finance. Deloitte’s research on return from AI investment places satisfactory ROI on a typical use case at two to four years, with 6% of organizations reporting payback inside 12 months. Against that curve, a portfolio brief promising a return inside its own charter period is either measuring the wrong thing or teaching the organization to report optimistically. A brief establishing whether the capability is being built, and whether the assumption underneath the bet still holds, is measuring the right thing on the right clock.

That evidence is also what should open the next tranche. Len Greski states the mechanism plainly, describing portfolio approval for the next scope of work as something that “releases per verified outcomes, not an annual spending plan.” A boundary that opens only when the calendar turns converts fast evidence into a queue.

Evidence triggers used to be aspirational, because nobody was watching continuously and the review calendar was the only thing that reliably arrived. That constraint is gone. An agent can hold the assumptions underneath every funded bet and surface the ones that have stopped holding, which turns the evidence trigger from an intention into something that actually fires. It surfaces the condition. It does not open the boundary, because opening a boundary moves money and commits people, and that has an owner.

What Lou Learned

Two quarters on, the portfolio review opened with the backlog rather than the spend. Three of 11 approved problems had been stopped at their first boundary, when the evidence said no and stopping was still cheap. Two more had never been approved at all, because nothing had been freed to build them and nobody was willing to pretend otherwise.

Alex asked what Lou would tell a peer walking into the same position. Three things:

1. Briefs, not objectives. An objective names the distance between the capability the strategy requires and the capability that exists. A brief turns that distance into a problem small enough to approve, carrying the outcome it targets, the single key result it moves, and its standing against everything else waiting. Agents can rank that backlog consistently, which is more than most portfolios manage, but the rubric they rank against is a decision with an expiry date on it.
2. Capacity, not permission. Ask both questions, whether the organization can build the thing and whether the customer can absorb it. Funding is the available capacity to do the work, and a customer with less absorption capacity does not adopt later so much as adopt less.
3. Evidence, not the calendar. Ninety days is rarely long enough to earn a return. It is long enough to learn whether the work is heading somewhere worth funding. The first round is judged on what it proved, and the next round opens when the proof arrives rather than when the quarter ends. The real test is whether funding decisions get revisited faster than the facts behind them go stale.

The question Alex repeated still gets asked, and now it has an answer. A company becomes whatever its capacity keeps confirming. The same people still decided. What changed was that the deciding could finally hear something back before the next cycle began.

This post comes from our Management Consulting practice, which specializes in designing and implementing operating models that align governance, processes, and technology to drive measurable business outcomes.

Sources

Bryce Ryan and Neal C. Gross, Acceptance and Diffusion of Hybrid Corn Seed in Two Iowa Communities, Iowa Agricultural Experiment Station Research Bulletin 372, Iowa State College, 1950. Originally published as “The Diffusion of Hybrid Seed Corn in Two Iowa Communities,” Rural Sociology, 1943. https://didawiki.cli.di.unipi.it/lib/exe/fetch.php/wma/agricultural_research_bulletin-v029-b372.pdf

Deloitte, AI ROI: the paradox of rising investment and elusive returns, October 2025. https://www.deloitte.com/global/en/issues/generative-ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html

Eliyahu M. Goldratt and Jeff Cox, The Goal: A Process of Ongoing Improvement, North River Press, 1984. Character homage, names only, in tribute.

Leonard Greski, “The New Software Economics: Earn the Right to Invest Again, in 90-day Cycles,” LiminalArc, May 2026. Originally published in Architecture & Governance Magazine, April 2026. Direct-quote source, and the source of the ninety-day funding window referenced. https://www.liminalarc.co/2026/05/the-new-software-economics-earn-the-right-to-invest-again-in-90-day-cycles/

Roger L. Martin, “Stop Letting OKRs Masquerade as Strategy,” Playing to Win / Practitioner Insights, November 2021. Direct-quote source. https://rogermartin.medium.com/stop-letting-okrs-masquerade-as-strategy-a57fc2cea915

Rose Hollister and Michael D. Watkins, “Too Many Projects,” Harvard Business Review, September–October 2018. https://hbr.org/2018/09/too-many-projects

Adam Whaley, “Are You Building the Right Thing: The Metrics That Measure How Fast You Learn,” LiminalArc, July 2026. Direct-quote source, and the source of the feature usage measure and the Standish figures cited. https://www.liminalarc.co/2026/07/are-you-building-the-right-thing-the-metrics-that-measure-how-fast-you-learn/