Northbeam's portfolio guidance and the board-mandated 18-month AI roadmap make this a sequencing decision, not an investment decision — the capital is already authorized. Fifty-eight of sixty-four invited leaders and operators completed the full diagnostic, drawn from all six plants and every function but one (People & Safety's two invitees both stalled mid-assessment), and eight sat for a structured interview. This is a representative commercial sample, not a census, and no claim of statistical significance is made anywhere in this report.
Leadership reads Helios's written AI strategy and its two completed pilots as forward motion. The evidence says otherwise: both completed pilots show negative ROI once fully attributed, and 61% of respondents cannot describe how any AI pilot at Helios was measured after it shipped. The most defensible AI capability in the company is a single order-and-dispatch coordinator's personal spreadsheet macro that halves her keying time — undocumented, unmeasured, and never asked about until this assessment.
Two constraints bind together. Cultural trust: a 1.4-point gap between leadership confidence in AI and frontline trust in it, the widest in this cohort, with frozen-middle supervisors fearing replacement at 2.4× the executive estimate. Data fragmentation: floor-execution data lives in four disconnected systems that do not tie to a single part number or work order. Neither can be resolved after the other — a trust program without reconciled data has nothing credible to point to, and reconciled data without frontline trust will not be adopted at the floor. They are funded in the same 90 days or the portfolio does not start.
Fund — Floor-data unification and exception-routing (MES/ERP). The highest-volume, most measurable workflow, and the one place a proven manual technique can become a governed, network-wide standard. Fund — Frontline trust and operating-model reset. The prerequisite every other initiative depends on, and the one place data alone cannot substitute for a named, trusted operating model. Validate — Agent-assisted RFP and compliance-filing consolidation. Real value, unproven at scale; bound it to a dated pilot before committing further spend. Defer — Agent-drafted quarterly board reporting. Low volume relative to build cost until the value-measurement discipline from Priority 1 exists to judge it against. Decline — Standalone, enterprise-wide AI vendor-evaluation copilot. The most requested item on the table and the least defensible: no workflow owner, no target metric, and it would compete with Priorities 1 and 2 for the same scarce change-management capacity this quarter.
1. Name a single accountable owner for the AI portfolio with real decision rights over data access and tooling. — Chief Executive Officer. 2. Ratify a one-page acceptable-use policy and put one sanctioned tool in front of the assessed cohort. — VP, Operations. 3. Baseline three workflows before any build starts: order-exception cycle time, incident RCA time-to-root-cause, and RFP turnaround days. — Chief Financial Officer. 4. Publish a standard part-number cross-reference across all six plants, accepting the loss of local flexibility it costs. — VP, Operations. 5. Put the day-90 portfolio gate on the calendar now, with kill criteria written down before the build starts. — Chief Executive Officer.
- The accountable owner is named by day 15.
- The policy and the sanctioned tool land before any measurement work starts.
- All three baselines are recorded before either funded build begins.
- The day-30 activation checkpoint closes the window.
| Role | Does | Receives | Capable of after |
|---|---|---|---|
| Chief Executive Officer | Signs the ownership decision naming a single accountable owner with decision rights over data access and tooling. | The day-30 activation readout: the three recorded baselines and confirmation no funded build has started ahead of its baseline. | Stating the two coupled binding constraints and the reason the vendor-evaluation copilot was declined, without reference to this report. |
| VP, Operations | Closes out the acceptable-use policy and puts one sanctioned tool live for the assessed cohort. | CEO signature on the ownership decision and access to the four legacy plant item-code systems. | Showing where every AI-touched workflow runs and what data it is permitted to see. |
| VP, People | Runs the first psychological-safety pulse to establish the day-30 trust-gap baseline. | A named liaison at each of the six plants for the trust and operating-model reset. | Quoting the trust-gap baseline as a number, by plant. |
| Chief Financial Officer | Records the first order-exception cycle-time baseline from a measured sample rather than an estimate. | The canonical part-number cross-reference the extraction work will be built against. | Quoting the order-exception baseline as a number and naming which plants it does and does not cover. |
| Director, Plant Operations | Records the first incident RCA time-to-root-cause baseline across all six plants. | The sanctioned tool and a written rule on what a system may draft versus what a person must confirm. | Producing an RCA baseline from recorded incident logs rather than recollection. |
- The part-number cross-reference goes live before the first extracted order is confirmed.
- The first extracted order is confirmed by a human before any order is written unattended.
- The Priority 3 pilot starts only once both funded builds are running.
| Role | Does | Receives | Capable of after |
|---|---|---|---|
| Chief Executive Officer | Holds the portfolio to its funded scope: no further AI request enters this quarter without a named owner, a target metric, and a baseline. | One portfolio page a month — movement against each baseline and anything now off track. | Answering, for any request to widen the portfolio, which binding constraint it relieves — and declining it when the answer is neither. |
| VP, Operations | Runs the first extracted orders through human confirmation into the ERP and publishes the exception-routing error rate weekly. | A weekly exception-routing report: volumes, confirmation rate, and where the extraction is wrong. | Approving or rejecting an extracted order against a published confirmation standard instead of re-keying it. |
| VP, People | Stands up the hub-and-spoke liaison structure at all six plants and runs the second psychological-safety pulse. | The day-60 gate evidence: measured movement on at least one funded workflow against its day-30 baseline. | Deciding on published evidence, not advocacy, whether the trust-gap intervention is working. |
| Director, Plant Operations | Scopes and dates the RFP/compliance-filing validation pilot — two plants, ten weeks — and agrees the accuracy margin it must beat. | Defect reports from the funded workflows against the canonical part-number cross-reference. | Stating the margin by which the validation pilot must beat manual review before it can scale. |
| Chief Financial Officer | Publishes the order-exception cycle-time trend weekly against the day-30 baseline. | Run-rate net projections separated from first-year net for both funded workflows. | Telling the board which operating-plan line each AI-attributed figure lands in. |
- The day-90 portfolio gate is held before anything new is funded.
- The validation pilot is funded or killed at that gate rather than allowed to drift.
- The quarterly review becomes a standing commitment before month six.
- A follow-on Sextant re-assessment is scheduled by month nine and run at month twelve.
| Role | Does | Receives | Capable of after |
|---|---|---|---|
| Chief Executive Officer | Holds the day-90 gate on the record: each funded workflow either shows a published before-and-after or is re-scoped, and the validation pilot is funded or killed in public. | The quarterly capital-allocation pack with every AI-attributed figure reconciled to a published metric. | Defending both funded builds to the board on their year-two run rate while stating plainly that year one is thin. |
| Chief Financial Officer | Reconciles every AI-attributed figure in the quarterly capital-allocation pack to a published operating metric, and refuses the ones that cannot be. | Run-rate net per funded workflow, separated from first-year net. | Telling the board which line of the operating plan each AI-attributed figure lands in — and which claimed gains were struck for want of a published metric. |
| VP, Operations | Runs floor-data unification in production against published metrics and reports its run-rate net into the quarterly pack. | The workflow's AI-attributed figure as a line in the capital-allocation pack, not a slide in a program update. | Training a new plant supervisor on the governed workflow without reference to Areté Intelligence. |
| VP, People | Publishes the twelve-month trust-gap trend across all six plants and hands the hub-and-spoke liaison structure to a second cohort of supervisors. | A published proposal-acceptance rate and a trust-gap trend by plant. | Running the operating model for a month without direct involvement and knowing exactly what changes when she does. |
| Director, Plant Operations | Scales the validated RFP/compliance workflow to the remaining four plants only if the day-90 gate passed, and reports its run-rate net. | Quarterly funding decisions with their reasons. | Stating whether the RFP workflow's constraint today is data access or model capability. |
"We've run two AI pilots and they're paying for themselves."
58 of 58 completed respondents could describe at least one AI pilot at Helios by name; 0 of 58 could cite a recorded before-and-after metric for either. Both completed pilots show negative ROI when fully attributed. This is a whole-cohort count at N=58, the only denominator in this report that clears the quantitative floor alongside the frontline/operator layer.
“"Everyone says the RFP pilot is a win. Nobody can show me the before number." — Chief Financial Officer” — Chief Financial Officer
"Our standard operating procedures are the same at every plant."
Among the frontline/operator layer, the only layer at or above the N=10 quantitative floor (N=31), 24 of 31 described a plant-specific workaround for order exceptions. The supervisor layer (N=7) falls in the 5-to-9 band and is reported qualitatively with a small-N caveat, not crossbroken: supervisors described the same pattern and attributed it to legacy customer relationships predating the corporate ERP rollout.
“"Plant 3 does a rework hold one way and Plant 6 does it another, and both of them think they're following the SOP." — Director, Plant Operations” — Director, Plant Operations
"Our floor teams are eager for more automation — they've told us so."
33 of 58 completed respondents said a coworker's manual number is more trustworthy than a system-generated one. Frozen-middle supervisors report fearing job replacement at roughly 2.4× the rate executives estimate they do, drawn from the diagnostic's psychological-safety block.
“"I'll use the tool if someone I trust tells me it's right. Right now that's nobody." — Shift Supervisor” — Shift Supervisor
"The ERP has the numbers. Reconciliation is a reporting exercise, not a data problem."
31 of 58 completed respondents described assembling a routine number by hand from two or more systems. Four of six plants still run pre-rollout item codes that do not tie cleanly to the corporate part-number master.
“"I can get you a scrap number in an hour. I can't get you the same scrap number twice." — VP, Operations” — VP, Operations
"IT and Operations are jointly driving this."
The IT & Data cell has 4 completed respondents, below the reporting floor of 5, so this finding is reported qualitatively and is not crossbroken. Across the eight interviews, three different functions named three different accountable owners, and no charter or budget line for AI work was produced on request.
“"I own the machines. Nobody's handed me AI, and there's no line item for it." — Director, Plant Operations” — Director, Plant Operations
Helios sits at the low end of Experimenter, and the distance to Builder is structural rather than incremental. Governance is the strongest of the low pillars — industrial-safety discipline extends naturally to AI oversight — but it sits on top of the two pillars that gate every opportunity in the portfolio: a 1.0 Culture score (the widest leadership-frontline trust gap in this cohort) and a 1.1 Data Foundation score (floor-execution data trapped in four disconnected systems that do not reconcile).
A written AI strategy exists and a sponsor-mandated 18-month roadmap is in force, so Strategy reads higher than the pillars beneath it can currently support. That mismatch is the finding: Helios has more ambition than substrate. 33 of 58 completed respondents said they trust a coworker's manual number over a system-generated one, and supervisors report fearing job replacement from AI at roughly 2.4 times the rate executives estimate they do.
Builder placement is not reached by adding more pilots. It is reached when the floor-data unification and the frontline trust program move together — data alone does not make the culture gap smaller, and a trust program alone cannot be measured without a reconciled system of record.
- No accountable owner and no decision rights for AI work
- No canonical part-number cross-reference across the six plants
- No baseline on any workflow, so nothing can be promoted on evidence
Helios sits at the Tool user rung: individuals use informal techniques inventively for personal productivity, exemplified by one coordinator's undocumented reconciliation macro. The next rung, Workflow author, is reached when plants write reusable, shared, grounded workflows instead of re-inventing personal ones — which requires the governed tool, the part-number cross-reference, and the three baselines this report sequences. Nothing above Workflow author is reachable while both coupled constraints remain open.
Nothing in this report compounds on its own. Two things would.
The first is the data. The part-number cross-reference is funded once, as a shared enabler, and no opportunity's return is credited with having built it. Once it exists, the work not funded this cycle stops being expensive: the RFP/compliance pilot has a reconciled spec source to validate against, and the board-reporting workflow has a measurement discipline to be judged by.
The second is the trust program. A frontline-leadership psychological-safety gap either narrows on published evidence or it does not, and today there is no published evidence at all. Closing that gap is what makes the exception-routing workflow adoptable past the plants where the original coordinator's technique was already trusted informally.
Year one is thin, and we have said so on every page that carries a number. The run rate is not thin. A second cycle inside the same operating-plan year would begin from a reconciled part-number cross-reference, three live baselines, and two plants that have shipped a governed workflow and can scope the next one with far less of us in the room. Representative Sextant output — fictional company and modeled assumptions.