Skip to content
LUNTA

The delivery system

Nothing advances without evidence.

Enterprise AI fails in the seam between a working demo and a changed operation — the MIT NANDA initiative’s 2025 report The GenAI Divide found 95% of enterprise GenAI pilots delivering no measurable P&L impact. Our delivery system exists to close that seam: five phases, each ending at a gate, each gate requiring named artifacts before anything moves forward.

We publish the system in full. It is how we work on every engagement, and it is how you should judge us — phase by phase, gate by gate.

  1. 01 · New moon

    2–4 weeks · fixed fee

    Diagnose

    See the system as it is.

    Before anything is built, we establish ground truth: where value actually sits in your operation, what your data can honestly support, and which candidates deserve investment — including the ones we will advise against.

    • Diagnostic report: value map and ranked opportunity portfolio
    • Data-reality assessment for each candidate
    • Written evaluation thresholds the pilot must meet to proceed
    • Baseline metrics that Full Value will be measured against
    If this gate does not pass
    Diagnose does not fail — it concludes. If nothing in the portfolio is worth piloting, that is the finding, written up with its evidence rather than a candidate manufactured to keep the work going. You keep the value map, the data-reality assessment, and the baselines either way.

  2. 02 · Waxing crescent

    4–8 weeks

    Pilot

    First light, under production conditions.

    A working system on your real data and real workflows — not a demo on curated examples. The pilot is judged against the thresholds agreed in writing during Diagnose. A pilot that misses them ends here, with reasons.

    • Working system operating on production data
    • Evaluation results against the pre-agreed thresholds
    • Cost, latency, and failure-mode profile
    • A written go / no-go recommendation — either way
    If this gate does not pass
    The pilot closes here, in a written no-go: what was tested, the result against every threshold in the signed schedule, the failure taxonomy, and what would have to change for the answer to change. Softening a threshold once the results are known is not an available option — the schedule was signed before there was anything to evaluate.

  3. 03 · First quarter

    6–12 weeks

    Scale

    From one workflow to the operation.

    The proven system is engineered to production standard: integrated with your platforms, secured, monitored, and held to service levels. This is the phase where AI programmes most often stall quietly — so it is where our engineering discipline is heaviest.

    • Production deployment with monitoring and alerting
    • Integrations live across the systems of record
    • Service levels defined and met
    • Security review closed
    If this gate does not pass
    A system that cannot hold its service levels or close its security review does not reach more users. This is the phase most often kept quietly alive, so both options are named: extend against a written remediation plan with its own dates and gate, or withdraw the system to the pilot cohort. “Rolling out while we finish hardening” is not one of them.

  4. 04 · Waxing gibbous

    4–8 weeks

    Embed

    Your organisation runs it.

    A system nobody uses is a cost, not a capability. We build the operating model around the technology: trained owners, governance that works in practice, and adoption measured against targets — not assumed.

    • Adoption measured against the agreed target
    • Named internal owners, trained and operating
    • Governance and risk controls in routine use
    • Support and escalation model live
    If this gate does not pass
    If adoption misses the agreed target, the answer is not more training. We report the measured number, name what the operating model is missing, and — where the workflow was never going to absorb the system — recommend withdrawing it rather than defending a licence nobody opens.

  5. 05 · Full moon

    At programme close

    Full Value

    Measured outcomes against the day-one baseline.

    The programme closes on a number, not a narrative: value realised against the baseline set in Diagnose, with the measurement method agreed by both sides. This is the only moment that earns gold.

    • Value-realisation statement against the Diagnose baseline
    • Measurement method reviewed and signed by both sides
    • Run-cost and ownership economics documented
    • Handover complete — or a scoped next cycle, on evidence
    If this gate does not pass
    If measured value falls short of the case, the statement says so — using the measurement method both sides signed before the measurement was taken. A shortfall written down and explained is a programme your organisation can learn from; one renegotiated into a narrative is one it will repeat.

A gate, in writing

This is the format every Pilot gate uses — an illustrative sample, not a client’s numbers. Real thresholds are set with you during Diagnose, against your baseline, and signed by both sides before the pilot begins.

Illustrative pilot exit-gate thresholds
First-pass resolution on production tickets≥ 62%41% · measured in Diagnose
Answer accuracy, adjudicated sample (n = 500)≥ 95%—
P95 latency at operating volume≤ 4 s—
Cost per resolved case at volume≤ $0.40$3.10 · human-only
Any threshold missed → a written no-go recommendation, with reasons and what would have to change for the answer to change.

See the other three gate artifacts

Why progress is never walked back

We advance a programme only when its exit gate passes — so a phase, once reached, never has to be surrendered. Progress at LUNTA does not get walked back, because it was never claimed before it was proven.

What this makes impossible

Pilot purgatory.
A pilot cannot idle indefinitely — it passes its gate or it closes, in writing.
Scaling on momentum.
Nothing reaches production because a steering committee is embarrassed to stop it.
Demo-driven decisions.
Every verdict is rendered against production data and pre-agreed thresholds — never a stage-managed walkthrough.
Unmeasured success.
The programme closes on a number against the day-one baseline, with the method signed by both sides.

Phase 01 is a fixed-scope, fixed-fee diagnosis.

Two to four weeks, ending in the Diagnose gate artifacts — including the thresholds any pilot would have to clear. Yours whether or not phase 02 happens with us.

Start at phase 01See the artifacts each gate produces