ForesightOps
    Outcome-Owned Architecture

    The agentic build-out is failing in the open.

    Money is being staked on pilots the way it is staked at a table, on a demo and a promise, with no model of the return and no honest accounting of what a wrong bet costs.

    Request the briefingDownload the white paper
    A working session on your own numbers, not a pitch.
    40%+
    Of agentic AI projects scrapped by the end of 2027.
    Gartner, June 2025
    95%
    Of enterprise AI pilots return nothing measurable to the business.
    MIT, cited Feb 2026
    88%
    Of AI agent proofs of concept never reach production.
    IDC / Deloitte / Forrester, 2026
    Why programs fail

    Most agentic AI programs fail for one reason: the work was scoped and owned badly, before any technology was chosen.

    By the time the technology is chosen the failure is already built in. No one set what the agents would own, where a person stays accountable, or what risk-adjusted return the work had to produce. When the work is scoped and owned badly, that is an operating model problem, and it is the gap no software vendor can close for you.

    The structural inversion

    Structure the enterprise around the outcomes it owes.

    The org chart is shaped like its workflows because coordination was expensive, information moved in batches, and execution required a person. Those constraints have collapsed. OOA re-derives the organization from the stakeholder outcomes it owes, making the decision the unit of structure and the decision boundary the seam where human judgment stays.

    Today · organized by function, optimized for control
    One outcome crosses every silo.

    Onboard a hire, and the work touches HR, IT, Security, Finance, and Facilities. The work itself is fast. The lead time is consumed by what happens between the silos.

    HRITSECURITYFINANCEFACILITIESSTARTPRODUCTIVE○ THE OUTCOME WAITS AT EVERY SEAM
    Where the lead time goes

    85 to 95 percent of lead time is spent waiting between silos. The queues, tickets, and handoffs at the seams consume it. The work in the silos is the thin slice.

    Tomorrow · organized by outcome, optimized for flow
    Each outcome owns its decisions.

    Agents execute the knowable decisions with no wait. Humans enter only at the decision boundaries that carry real consequence. The colored edge on each decision is its decision boundary.

    Hire is productive
    Identity & eligibility
    Access bundle
    Access exception
    Readiness
    Vendor paid correctly
    Match & code
    Exception
    Material dispute
    Books are true
    Reconcile
    Estimate
    Attestation
    Where the lead time goes now

    The queue collapses everywhere except the few real decision boundaries. Agent work runs with no wait. Only the seams still queue, and only briefly.

    Agent-owned · known, low tail-costSeam · the agent prepares, a human ratifiesHuman-owned · unknowable or irreducibly accountable
    What changed

    The functional org chart answered four constraints. Agentic capability can remove them.

    Place the capability into the existing structure and the constraint stays, because the queues, the decision rights, and the seams are still shaped by the old assumption. The capability makes removal possible. The redesign is what removes it.

    Information
    The old limit
    moved in weekly batches
    ↓ can be removed
    can be instant
    Every function reads the same live numbers the moment they change.
    Coordination
    The old limit
    across functions was costly
    ↓ can be removed
    can be nearly free
    A standing meeting becomes a single update to a shared system.
    Execution
    The old limit
    required a person
    ↓ can be removed
    can run in software
    The decision is carried out the moment it is made, inside set limits.
    Decisions
    The old limit
    waited for a schedule
    ↓ can be removed
    can be continuous
    Made against what is true now, not last week's snapshot.
    The law beneath the lossAcross knowledge work, value-add time is a small share of total lead time. The rest is queue, and the queue forms at the decision seam. Two lenses then compound: flow efficiency locates the loss, and the OOA model assigns each decision a risk-adjusted value from the outcome it serves, then places it across the decision boundary.
    The decision boundary

    The line where ownership transfers is the operating model.

    Every decision runs as a loop. The question OOA answers is which part of that loop software owns inside declared policy, and where a named human stays accountable. That line is the judgment boundary, and drawing it correctly is the whole of the work.

    The decision loop, partitioned by the judgment boundary
    The machine owns the loop inside declared policy. Humans own the three loops that govern and interpret it. Ownership transfers only at the seam. Expand for the full figure.
    GOVERNED BY HUMAN JUDGMENT The decision loop, partitioned by the judgment boundary A cybernetic control system. Outcomes revise policy, and ownership transfers only at the seam. THE JUDGMENT BOUNDARY the live edge, the single line where ownership transfers agents may always hand up, never hand down STANDING ACCOUNTABILITY the Policy Loop owner is accountable for the loop's outcomes THE MACHINE-OWNED LOOP Operates within declared policy. The five-tuple ⟨τ, I, π, χ, φ⟩. Trigger τ · the event that opens the loop Information ι · the state assembled to decide Policy π · the rule that turns the information set into a commitment HUMAN-OWNED, UNDER CHANGE CONTROL Autonomous commitment χ · the agent executes, inside policy Outcome φ · results observed, the loop's feedback element HAND UP to the named receiving loop, on typed conditions POLICY UPDATE the Policy Loop revises π, re-enters the machine loop THE HUMAN-OWNED LOOPS Humans own loops, they are not in the loop. Exactly three classes. Policy Loop Owns the rules that govern the agents: thresholds, watchlists, budget lines, boundaries. Owns π. PERIODIC CADENCE · THE ACCOUNTABLE OWNER Trade-off Loop Owns the calls the numbers do not settle. The agent surfaces options, the human owns the commitment. EVENT-DRIVEN CADENCE Meaning Loop Owns pattern interpretation across instances: what it means and what we do about the relationship. PERIODIC REVIEW The Policy Loop owner also owns the machine loop's outcomes. "The agent decided" is never an account of anything. LEGEND Execution flow (machine-owned) Feedback: φ revises π Hand up + policy update The judgment boundary (the seam) Standing accountability CYBERNETIC CONTROL A closed loop: outcomes produce feedback that revises policy, which governs future decisions. POLICY ON THE PATH Every commitment passes through policy. π maps the information set to the commitment, never bypassed. THE JUDGMENT BOUNDARY The central safety and accountability artifact. It partitions judgment, not execution. HAND UP, NEVER DOWN Agents escalate to a named human loop when outside policy. They never delegate downward. HUMANS OWN LOOPS Policy, Trade-off, Meaning. Humans do not step into machine loops, they own the outcomes across the seam. SEPARATE PHENOMENA Execution, feedback, ownership, and accountability are different relations, read in different visual languages. From Outcome-Owned Architecture (OOA). The decision loop as a five-tuple ⟨τ, I, π, χ, φ⟩. © 2026 Jamie Bernard and William Haas Evans, Fugue Strategy Advisors.
    /Download the PDF ↓
    The method

    A reproducible instrument the company owns and re-runs.

    The method is four moves, run in order. Where others deliver a readiness assessment or a build, OOA delivers a number the company owns and re-runs.

    01
    Map the decision seams
    Take one cross-functional process and find every point where a decision is made, and who makes it.
    02
    Classify each decision
    Sort every decision into one of four classes by who should own it once software can hold it inside policy.
    03
    Screen, then rank
    Run PRISM to remove what should not proceed, then rank what remains by risk-adjusted yield.
    04
    Price and sequence
    Output a risk-adjusted number and the order to move the boundaries, against the date the value has to land.
    Decision inventory
    Four decision classes
    Every decision belongs to one class, which sets who owns it: software inside policy, or a named human at the boundary. The inventory is the raw material the rest of the method runs on.
    The economic model
    ρ = V / (C + R)
    ρ = V / (C + R)
    Risk-adjusted yield: value over cost plus risk. It ranks the moves so capital goes to the boundary that pays the most against the horizon that matters.
    The screen
    PRISM
    Nine risk dimensions with hard gates, run before the ranking. A gate can remove the second-highest-yield lane, which is exactly the point: it catches what does not pay before capital commits.
    Why trust us

    The method is built to reach answers that cost us money.

    Most automation advice runs one direction, toward more spend, because the advisor is paid when you buy. This method establishes what the work is before any platform enters the analysis, so it can reach the opposite conclusion: that a loop should not be automated, that a workload does not justify its compute, that the right answer is fewer licenses or none.

    Decision before platform
    The work is mapped before any tool is named.
    Platform capability enters only as a derived output, downstream of the decision inventory, never as an input that decides which loops get valued. Let the tool shape the question and every answer tilts toward the tool.
    The mandatory inverse test
    Every engagement produces a loop we tell you to keep human or kill.
    At least one decision loop is carried through the full method with an expected outcome of human ownership or elimination. If the analysis cannot find a single thing not worth automating, the analysis is wrong. That is a finding, not a failure.
    The number you own
    A model built from your work, that you keep and re-run.
    The risk-adjusted model is built from your own volumes and rates. You own it, re-run it when token prices move, and take it to your board yourself. It renews on your terms, not the seller's.
    The method comes from having run the work, as the executive accountable
    Johnson & Johnson Global IT. Eli Lilly. Royal Caribbean, an operating model for agentic systems across more than forty ships, with an AI lab and innovation capability built. Twelve major operating-model transformations, run from inside the operation, not advised from a deck.
    What it is worth

    The redesign pays back $24.2M in risk-adjusted NPV, built from the company's own work.

    Most automation cases start with a percentage from an industry report and multiply. This one starts with the tasks the company's people actually do, the time each takes, and the cost of getting one wrong. Every figure traces to something the company can check against its own records.

    $24.2M
    NPV risk-loaded, discounted at 12 percent. The number the company funds against.
    2.35 yr
    To pay back
    90%
    Less human effort
    9.2×
    Benefit to cost
    Clean NPV $36.2M before the risk charge · risk-loaded $24.2M is the number you fund against · P10 NPV $19.2M across a Monte Carlo run on the six biggest drivers
    The engagement

    One decision boundary, proven in thirty days.

    The smallest engagement that proves the value, or kills it. No transformation bought on faith.

    What we run

    One cross-functional process. Map its decision seams, classify each decision by the four classes, screen with PRISM, and rank by ρ against the date the value has to land.

    The output is a risk-adjusted NPV, a payback, and an EBITDA impact for the top one to three moves, before any agent is built.

    Week 1 populate the model · Weeks 2 to 3 map one stream · Weeks 3 to 4 move the first boundary

    What counts as proof

    The number goes to the board or the executive committee, and it changes a capital decision. A deployment gets reordered, killed, or greenlit because of it.

    Then a pull for more: the same diagnostic on a second process or unit. A changed decision, and a pull for more. Interest and a good meeting do not count.

    Prove the return first. Then build what captures it.

    Request a briefingDownload the white paper