CurrentNorthstar LabsQ4 FY26
PR
Portfolio artifact — not a customer-facing surface

Product brief

The business case behind the prototype: who has this problem, why this wedge, why a unified platform wins it, what ships first, how it would be measured, and what is still a guess.

The wedge

The trigger is not “use AI.” It is: an approved business objective is at risk because workforce capability will not arrive in time.

That moment already exists in every company large enough to run a headcount plan, it already has an owner, and it is currently resolved by three functions emailing each other. Current is sold into that moment, not into a general desire for automation.

Who has this problem

Buyer, user and the person whose objective is actually at risk are three different people.

  • Economic buyer

    CPO or Chief People Officer, jointly with the CFO

    The purchase is justified by delivery risk and workforce cost, not by an HR feature. Finance has to be in the room because the product routes spend decisions to them on day one.

  • Primary user

    Workforce planning and Talent leadership; recruiting leads

    They already own the translation from "we are short" to "open a req" — the translation this product automates.

  • Decisive user

    The functional leader whose objective is at risk — a VP Engineering, in this scenario

    They feel the gap first and they hold the approval. If they do not trust the modeling, nothing downstream happens.

Initial segment

Organizations already running headcount planning, recruiting and at least one talent development system on the same platform — where the data to compare the four paths already exists and only the orchestration is missing.

Why a unified platform wins this

The argument has to survive “why can’t a point solution do this?”

  • A point recruiting tool can only recommend recruiting

    Ask an ATS how to close a capacity gap and the answer is always a requisition, because that is the only verb it has.

  • An LMS only knows learning; a planning tool only knows headcount

    None of them can compare a redeployment against a contract against a hire in the same units, because none of them holds all three.

  • The comparison needs one graph

    Capability supply lives in delivered work and learning records. Cost lives in the headcount plan and the contingent budget. Authority lives in the permission model. A recommendation that spans them has to resolve against all of them at once.

  • Permissions are the moat, not the model

    Agent actions are only safe if they inherit an existing, already-audited permission model. A standalone agent has to invent one and ask customers to trust it.

What ships first

Broad vision, narrow wedge. The MVP is two paths executed properly, not four executed shallowly.

  1. Phase 1This prototype

    See and plan

    Detect the gap from approved inputs, model the four paths, show the evidence, get a human decision.

  2. Phase 2This prototype

    Execute Build and Buy deeply

    Development plans, learning enrollment, success profiles, sourcing, governed outreach campaigns and scheduling — end to end, with the boundaries enforced.

  3. Phase 3Next

    Widen the trigger surface

    Survey-to-action, performance action plans, contingent and automation paths as first-class workstreams rather than one modeled option.

  4. Phase 4Later

    Outcome learning

    Measure which interventions actually closed gaps and correct the estimates that were wrong. The ramp assumptions in this prototype are exactly what Phase 4 would replace with evidence.

How success is measured

Four levels, and the reason each one is on the list.

Product
Median time from validated signal to resolved gap
The north star. Everything else is a leading indicator of it.
Share of gaps resolved by target date inside cost and policy boundaries
Resolving late or over budget does not count.
Agent
Policy violation rate
Zero. A gate, not a threshold.
Signal precision
Above 80% judged actionable by the owner. The failure mode is alert fatigue, not a missed gap.
Human interventions per plan
Tracked, not minimized. Falling to zero means the boundaries stopped working.
Reversal rate
Below 5% of executed actions undone within 30 days.
Tool failure rate
Every action runs through a real product API and can fail like one.
Customer
Time-to-action on a workforce signal
Days of coordination removed, measured against the manual baseline.
Internal mobility rate
The behavior change the product is arguing for.
Plan-versus-actual accuracy
Whether the forecast is worth trusting.
Business
Talent suite attach rate
The hypothesis is that orchestration pulls modules in.
Multi-product adoption depth
A gap closed across three modules is a customer who now uses three modules.
Consolidation wins against point solutions
The competitive claim, tested.

Business hypotheses

Stated as hypotheses because that is what they are. No revenue is invented here.

  • Orchestration raises Talent suite attach

    A gap closed across three modules is a customer using three modules. Unproven.

  • Workflow depth raises switching cost

    Approval routing and policy configuration are sticky in a way reporting is not.

  • It is a consolidation argument against point solutions

    The comparison a point tool cannot make is the reason to buy the suite.

  • It gives four products one agentic layer

    Recruiting, Headcount Planning, Performance and Learning stop needing separate AI stories.

Assumptions I would go validate next

No customer interviews informed this. These are the questions that would change the product if the answers came back differently.

  • Who actually notices the capacity gap today, and how late?

    The whole product assumes the answer is "a VP, too late." If planning teams already catch this in a quarterly cycle, the wedge is narrower than modeled.

  • Who owns the hire-versus-redeploy-versus-contract decision?

    If that authority is genuinely split across three functions with no forum, the product is a coordination layer. If one person already owns it, it is a decision-support tool — a materially smaller product.

  • How much autonomy will People and Finance actually permit in year one?

    The autonomy model here is a hypothesis. Level 3 may be unsellable to enterprise legal without a pilot period at level 2.

  • What has to be true before an agent sends external candidate outreach?

    Recruiter authorization is modeled as sufficient. Real customers may require legal review, brand approval, or a per-message human send.

  • Which workforce signals produce action rather than noise?

    Three signals a week is a guess. The precision target is a design constraint, not a measured result.

  • Would a customer accept a plan that leaves 6% of the gap uncovered?

    The residual-risk acknowledgement assumes people prefer an honest partial answer. It is equally possible they read it as the product failing.

Synthetic data is not customer validation. This prototype demonstrates a thesis and a system design; it does not evidence demand.