Axiarete reads the software that runs your plants — the MES and the systems around it, the integrations between them, and the custom code and spreadsheets holding it all together — and builds a living digital twin of the technology estate. Then it turns what it finds into measured, verified fixes: faster transactions, shorter blind windows, less unplanned downtime, and a modernization plan built on evidence.
Manufacturers running a customized MES or MOM they cannot yet replace, and a modernization decision they cannot yet scope.
Read the estate from its own code, configurations and logs. Measure it the way you measure the line. Fix what the evidence points to. Hand over the knowledge.
Two hours a week from a project lead, stage access, eight weeks, and an executive decision at the end — on results.
Manufacturing is building again. Annual spending on U.S. manufacturing construction ran about $82 billion in 2021, peaked near $250 billion at an annual rate in August 2024, and still stands near $170 billion — more than twice the 2021 average.1
The labor gap those buildings face is severe, and it is the problem capital is already attacking. Open U.S. manufacturing jobs stood at 580,000 in July 2026, up 35.5% on the year.3 Manufacturers will need 3.8 million workers between 2024 and 2033; up to 1.9 million of those roles could go unfilled.4 Meanwhile 542,000 industrial robots were installed worldwide in 2024, twice the figure of a decade earlier, with about 4.7 million now in operation.5 For the first time, the path to closing the labor gap is visible and funded.
Robots do not shrink the systems problem. They multiply it. Every robot is another consumer of work instructions, schedules, quality specifications and exception handling. Without a system deciding what to make, when, and what to do when the upstream station falters, a robot is stranded capital.
The numbers above are American because the American data is the sharpest. The pattern is not. Wherever capital is flowing into plants, it is flowing into buildings whose software layer has never been measured the way the equipment is.
Sales got Salesforce. HR got Workday. IT got ServiceNow. Below the ERP line, the stack that tracks production, decides what runs where, supervises the machines and records what they did is a different story. The dates tell it.
The first PLC ships to an automaker. Ladder logic is drawn to look like relay wiring so 1970s plant electricians can read it. It still looks like that.8
Windows arrives in the HMI.9
The term “MES” is coined.10
ISA-95, the reference architecture for MES and the layers around it, is published.11
ISA-95 Part 1 is revised. No successor architecture has displaced it.11
Still shaping deployments.
The closest thing manufacturing has to a universal system is not any MES vendor’s product. It is Excel.
None of this reflects backwardness on the part of manufacturers. The structure worked against them. Every plant is an accumulation of its own history: a 1994 press brake beside a 2024 robot cell, Modbus beside OPC UA beside a CSV on a USB stick. Integrating that heterogeneity was human labor billed by the hour, which is why so much of the factory software business quietly became a services business, and why the plant floor so often received a copilot bolted onto a client-server application from 1998.
One question for your next ops review.
If your highest-margin line started producing scrap this minute — no alarm, no stoppage, just bad parts — how long before anyone knew?
Asked candidly, the answers come back in hours. Sometimes a full shift. That is the blind window. It has a visible twin: in ABB’s survey of 3,215 plant-maintenance decision-makers, 69% absorb an unplanned outage at least monthly, at a typical cost near $125,000 an hour.15
Application delays and blind windows never appear as downtime. They appear as scrap, as OEE that plateaus at 60, as operators waiting on screens, and as improvements that die in a release queue. Here is the arithmetic.
Bad parts made between the moment a process drifts and the moment a system tells someone. Hours, sometimes a shift. Multiply by parts per hour and cost per part: that is the bill for a system that documents what happened instead of deciding what happens next.
Near $125,000 an hour for a typical plant;15 up to $2.3 million an hour for an automotive line.19 Siemens’ survey-based estimate puts the Fortune Global 500’s annual losses to unplanned downtime near $1.4 trillion — about 11% of revenue, up from 8% in 2019.19
Typical plants run at 55–60% OEE; world-class is 85.18 Moving from 60 to 85 is 85 ÷ 60 ≈ 1.42: roughly 40% more output from the same machines, the same people and the same walls. Part of that gap is mechanical. Part of it is scheduling, dispatch, changeover and quality decisions waiting on software — and no plant knows which part until it measures.
At the machine you paid the most for.
The average digital worker switches applications about 1,200 times a day and loses nearly four hours a week reorienting.22 On the floor, one work order can mean the planner’s screen, then MES, then quality, then maintenance: four applications for one job, thousands of jobs a day.
Your process engineers know how to remove a step, tighten a dispatch rule or cut a changeover. The change waits months for a release window. The plant’s rate of improvement becomes the software’s release schedule.
While this page has been open, one idle automotive line would have burned
$0
Ticking at $638.89 a second — the arithmetic of Siemens’ up-to-$2.3M/hour figure.19 Against numbers like these, the software budget is a rounding error. The opportunity was never the license line. It is the plant’s P&L.
Change any figure; the results update as you type.
At these numbers your estate carries $48,000,000 a year of unplanned downtime, 41.7% of theoretical output headroom, and 33.3 operator-hours a day spent waiting on screens.
We don’t claim all of this is software. We claim you cannot know how much of it is until you measure the software the way you measure the line.
Digital paralysis is the inability to execute, adapt or derive value from technology despite heavy investment. It is not a shortage of systems. It is too many of them, poorly connected, understood by too few people.
Planner, MES, quality, maintenance: four applications to release and run one job. Each interaction is minor. Across thousands of jobs a day the delay, the cognitive load and the execution errors compound silently.
The dispatch list is a workbook. The changeover matrix is another. The hold list is an email thread. The system of record is technically running; the plant is being run around it.
Improvements that would raise yield or cut changeover wait for a release window, so engineers build workarounds instead. The factory’s rate of improvement ends up governed by software release schedules.
Item master in ERP, BOM in PLM, routing in MES, label data in a print server, each maintained separately. Decisions run on whichever version loaded first. The discrepancy surfaces weeks later as a wrong part, a wrong label, a planning miss or a recall.
A decade of customization layered into the MES; point-to-point interfaces nobody documented; HMIs on operating systems that stopped receiving patches years ago; timeouts that silently stopped working. Each decision was defensible. The sum flips from “fine yesterday” to “slow today” with no obvious cause.
The controls engineer who wrote the workaround for Line 3 retires next spring. The MES administrator who knows why hold logic is bypassed on nights is the only one who knows. The documentation describes the system as scoped, not as it runs.
Add acquisitions and it multiplies: five plants, four MES flavors, three ERPs, and a consolidation program stalled because nobody can say with confidence what each system does.
Takt measured in seconds, sequenced delivery in minutes, EDI and ASN in penalties. An MES transaction that waits four seconds at a constraint station is capacity you cannot buy back; a label server that fails is a truck that doesn’t leave.
Every serialized part carries an as-built record that must match the as-designed configuration, across systems integrated by hand and under export-control rules that limit who may look. The estate is the traceability; when it’s opaque, so is compliance.
Validated systems, electronic batch and device history records, and change control that is slow by design. The enhancement queue is measured in quarters. Evidence-based change — baselined, tested against recorded production, documented for validation — is the only kind that moves.
High mix, short runs, constant changeovers, component-level traceability across a supply chain that swaps parts weekly. The dispatch and setup logic is where the seconds and the errors hide.
Lot genealogy, allergen changeovers, recalls that turn on whether the system can say which pallet carried which lot. Label and print failures are shipment blockers; a hold not enforced at transaction time is product still moving.
Recipes, batch records, historians, and control loops that must close within seconds. When the data that closes the loop arrives late or inconsistent, variation spreads across batches before anyone intervenes.
Manufacturers learned long ago that what gets measured gets improved. Equipment has availability, OEE, Cp/Cpk, MTBF and PM compliance — trended, owned, reviewed. Applications have availability and ticket counts, which say whether the software is on, not whether it is helping or hurting the plant. A system can hold 99.99% availability while adding two seconds to every transaction, fragmenting the data planners decide on, and blocking the process improvements that would raise OEE. From IT’s view it is healthy. From the plant’s view it is a constraint running every hour of every shift.
Application Health closes that gap. It measures each application, and the estate as a whole, by whether it enables the plant to operate efficiently, adapt quickly and hit its outcomes. Not every three to five years in a consulting engagement, but continuously, the way equipment health is measured now — on the living twin the MRI keeps current.
Trended, owned, reviewed.
Says whether the software is on. Not whether it helps the plant.
Equipment, automation, AI and people all depend on software. Applications are no longer IT assets. They are production assets.
Manage every system that participates in execution by its contribution to plant outcomes, not by service levels alone.
No plant waits three years to inspect a critical machine. Application health belongs on the same review as OEE and on-time delivery.
Every customization, business rule, interface and workaround encodes institutional knowledge. Capture it in the twin before it retires.
Replace what constrains the plant or blocks the future, not what happens to be old.
Agents decide only as well as the systems, data and context beneath them. On a fragmented estate they accelerate bad decisions at machine speed.
Manufacturing watched the Industry 4.0 wave crest and break. GE spent six years and more than $4 billion on a digital push built around Predix and retreated by 2017.16 A generation of IoT platforms stopped at connectivity and dashboards, leaving manufacturers to build the applications themselves, in plants with no software teams to build them. All of it was attempted a decade before a language model could read a decades-old program.
Two costs have always made factory software hard to buy and harder to deploy: the cost of integration and the cost of interface. AI agents are collapsing both — reading old code and configuration, making sense of naming conventions from 2003, drafting the mappings and adapters for an engineer to verify. When integration cost falls, per-site uniqueness falls with it, and the multi-year deployment with that. In this market, deployment time is the decisive variable.
Retrofitting action onto a 1990s data model is how transitions are lost. The way across is not a bigger record. It is a living model of the estate that agents can stand on and be governed while they act.
No language model belongs in the safety loop. Interlocks stay hard. Safety-rated logic stays deterministic. Those standards exist because lives were lost establishing them. Axiarete’s agents live above the control layer, in the supervisory space where people spend their days coordinating, diagnosing, planning and firefighting.
The instinctive conclusion is to remove the legacy and install something modern. That instinct is what produced the industry’s multi-year deployments and its inventory of stalled programs. Few organizations retain a complete, current picture of the systems they propose to replace. The engineers who configured them have retired; the documentation describes the system as scoped in 2009. The most reliable record left is the systems themselves: their code, their configurations, their logs. That is where Axiarete begins.
The Axiarete MRI reads what the estate is made of and assembles a living digital twin of the technology estate: every application, dependency, customization, data flow, cost and risk, kept current as the estate changes.
The estate, finally legible
The MRI recovers what is embedded in code no one can still read — including much of the institutional knowledge now retiring with the workforce — and keeps it current. Which customizations quietly carry the plant, where debt concentrates, what each system costs to keep, where the risk sits. Decisions that waited on committee cycles are made in days, on evidence.
Rationalize what sprawls, modernize what matters
The twin turns into execution. Rationalize: retire what is redundant, consolidate what overlaps, reclaim the spend and attention. Modernize: rebuild what matters in staged, evidence-backed moves rather than bet-the-plant replacements. Throughout, the invisible is made financial — a defect becomes a business case, a slowdown a quantified cost, a vulnerability a priced risk — and work is ranked by return.
The estate as a living system
Optimization that isn’t operated decays. The twin stays current, drift is caught early, and the portfolio doesn’t slide back into opacity. This is also where the agent question is settled: agents without context guess; agents standing on a living model of the estate can act, and be governed while they act.
Operation feeds what it learns back into understanding.
Parses and decompiles every artifact; builds and maintains the twin; scores each application on the six dimensions and the 300-point Technical Health Assessment; ranks findings by production impact; drafts remediations — rewritten queries, index designs, consolidated logic, configuration corrections; and monitors continuously for regressions and new risk.
Services as software, and how it lands. AI-augmented engineers, forward-deployed: architects who scope against the twin and engineers who ship inside your pipelines from week one. They validate every change against recorded production transactions, decide what ships, deploy to one plant first, watch for 72 hours, keep a tested rollback warm, and hand over the knowledge base, test suites and procedures so the capability stays with you. Each engagement strengthens the platform; every estate read makes the next one faster.
Neither works without the other. We don’t pretend otherwise.
Start where it hurts. The same twin powers all six, so every engagement compounds the next.
Forensic, code-level tuning of the MES and the systems around it: custom code, model artifacts, database, platform configuration, integrations and hold enforcement. Continuous detection, RCA and remediation once the baseline exists.
Outcome40–60% lower shop-floor transaction latency; production configuration risks defused; zero functional regressions across two production tuning engagements.
How MES optimization works →Reconstruct the knowledge base for hundreds of applications across plants and business units directly from code; identify consolidation, retirement and modernization candidates with SME-validated evidence; build the business case.
OutcomeIn one engagement, an $11M savings plan on a $40M baseline; 21% of the portfolio identified as reduction opportunity and validated by SMEs; months of manual analysis compressed to weeks.
Feature-level replaceability assessment against target platforms; auto-generated migration plans, test cases and requirements; risk discovery for version upgrades and post-acquisition plant consolidation.
Outcome90% accurate replaceability assessment validated by in-house experts.
The 300-point Technical Health Assessment across architecture, quality, performance, security and governance; code-level drill-down; business-impact modeling per issue; agentic remediation for what matters.
Outcome50–80% reduction in architect and developer effort to understand and remediate issues; a standing foundation for enterprise technology risk governance.
Map plan-to-produce, order-to-ship and quality-event-to-CAPA across the systems that execute them, with health, performance and cycle time per step; find the toggle tax, the manual reconciliation and the friction; optimize.
OutcomeProcess bottlenecks removed where software, not equipment, is the constraint.
Unified risk discovery across source code, software supply chain and runtime for the applications that touch production; qualification by exploitability and production impact; agentic remediation.
Generic application and database advice doesn’t fix an MES. Axiarete’s analysis targets MES constructs directly — grid and dispatch events, custom logic functions, business rule handlers, data objects, site flags, integration and print queues — with the failure modes known for each.
Every change ships with its evidence: the baseline, the test results against recorded production transactions, the single-site deployment record, the 72-hour monitoring log and the rollback proof.
A standing capability, not a report. It keeps running after the engagement ends.
Both engagements were forensic: we decompiled the estate, measured every transaction, and rewrote what the evidence pointed to. Every rewrite was validated against recorded production traffic before it shipped. The platform in both cases — Siemens Camstar / Opcenter Execution on Oracle — runs discrete, electronics, medical-device and semiconductor plants worldwide; the pathologies are not industry-specific. Every figure in this section can be walked through, artifact by artifact, under NDA.
The situation. Shop-floor grids and track-in / track-out had slowed unpredictably. Execution plans flipped overnight: fine yesterday, slow today.
| Finding | Before | After | Count |
|---|---|---|---|
| Setup-matrix lookups sorted entire tables to return one row | ROW_NUMBER() over a 25-level CASE sort after 28 joins; every row materialized | Indexed specificity score + FETCH FIRST 1 ROW ONLY | 47 matrix queries re-engineered |
| Grid columns executed one query per displayed row | 16-join query per row; a 100-row grid drove 400 history scans | Lookup folded into the grid query as a set-based join | 400 → 1 history scans per render |
| Predicate shapes that disabled every index | (col LIKE ? OR col IS NULL) stacked 20 deep; LIKE on Boolean and Integer columns | Exact-match / wildcard UNION ALL branches with typed equality | 365 index-killing predicates rewritten |
| Index estate rebuilt on evidence | 125 redundant, 16 empty, 20-column-wide indexes; 20 on the hottest table | B-trees on 205 verified join paths; SQL Plan Baselines pinned | 125 redundant indexes retired |
The situation. Move-out, track-in / track-out and lot start had slowed as a decade of customization layered into the model database. Thousands of metadata calls queued ahead of every commit.
| Finding | Before | After | Result |
|---|---|---|---|
| One move-out fired thousands of metadata functions before commit | 110 custom logic functions wired to one event: 2,696 calls | Early-exit consolidation; printing and trace codes moved to after-commit | 2,696 → 1,500 functions per move-out; latency down 40–50% |
| A sixth of the schema had no index | 335 of 2,064 tables bare; 22 phantom indexes with zero columns | Composite B-trees on the history mainline | 335 → 0 unindexed tables; history reads up to 80% faster |
| Every dispatch recomputed a static hierarchy, recursively | 254 queries ran CONNECT BY over an unchanging tree | Hierarchy materialized, refreshed on deploy | 254 recursive lookups materialized; 1–4 s back per sequence |
| Dispatch grids fetched all 190 columns to display 15 | 48 SELECT * queries pulled the full row | Column-pruned to the 15 shown | 190 → 15 columns; payloads down ~90%; 1–5 s per dispatch list |
TABLE() context-switch calls re-engineeredReconstructed the knowledge base from 12 million lines of code; ran the Technical Health Assessment on every application; distilled three rationalization scenarios.
Archaeological knowledge construction for hundreds of legacy manufacturing applications; feature-level replaceability assessment against the target MES and ERP; auto-generated migration plan, test cases and requirements.
No workshops before evidence. The MRI starts from your code on day one, and the executive review at the end of week eight decides whether to continue, change scope or stop — on results. We schedule around your shutdowns and change windows, not the other way round.
Run the MRI on code, configuration, logs and incident records; produce the inventory of every customization. Add measurement: response time and database usage per transaction. Map the 8–10 transactions the plant depends on — release, dispatch, start, complete, inspect, hold, label, ship — and agree how incidents will be ranked by production impact.
Performance and KPI baseline by key transaction. Full inventory of customizations and site-specific settings documented. First low-risk fixes discovered, prioritized and delivered — for example, the setting mismatch that logs operators out mid-transaction.
Full roadmap prioritized by business impact and execution difficulty. Continue delivering — for example, timeout corrections so a slow step fails fast and releases its locks instead of blocking other users. SME reviews to validate and refine.
Validate every optimization. A twelve-month plan with priorities and action plans. Executive review: continue, change scope, or stop.
We can start with your top 2–3.
| The rip-and-replace program | Typical MES consulting | The Axiarete standard |
|---|---|---|
| 1–2 years per site, six-to-seven-figure budgets, and a stalled-implementation risk the industry knows well21 | Surveys, interviews and workshops; little attention to code or configuration | Bottom-up analysis of the current state — code, model artifacts, configurations — so the estate is understood before anything is replaced |
| Nothing improves until go-live | Findings gated on workshop availability | First findings in week two; full baseline by week two; quick wins in production by week four |
| Replaces the system, not the knowledge; the same workarounds get rebuilt | Generic application and database advice | Analysis targets MES constructs directly, with the failure modes known for each |
| One big change window | Ends with a recommendations document | AxiareteForge engineers execute — code, configuration, database objects, model — and verify each change |
| A tested way back is rarely part of the plan | Change windows with no tested way back | No production change before baselines; tested against recorded traffic; one plant first; 72 hours; rollback tested before use |
| Institutional knowledge lost in the cutover | Knowledge leaves with the consultants | Knowledge base, test suites, scripts and procedures handed over and kept current; the same evidence supports the upgrade or migration when it’s time |
And versus portfolio tools. CMDB, APM, TBM and EA tools are recordkeepers: they hold what you tell them, decay faster than teams can update them, and answer “what exists,” not “what to fix, retire or modernize, and why.” Axiarete reads the systems themselves and keeps the answer current.
Plant application estates encode recipes, routings, supplier relationships and, in regulated industries, validated process logic. We handle them under a signed NDA and a written data-handling agreement before any artifact moves, and we work from your stage environment with read-only database access wherever the work allows. Nothing Axiarete deploys touches safety-rated logic. Interlocks stay hard; the safety loop stays deterministic; our agents live in the supervisory space above it.
Manufacturing Vertical
Former CIO, SkyWater; former IT and operations executive at Intel, Micron, Maxim / Analog Devices, GlobalFoundries and Infineon. Author of Axiarete’s Perspectives series on the digital factory.
Strategy
Former CIO, Varian. Advisor to manufacturing CIOs. Formerly PwC and Deloitte. 2021 Bay Area CIO of the Year.
Solution Architecture
Leads Axiarete’s application rationalization and MES optimization engagements. Formerly Bain & Company and HP.
Customer Engagement
Leads Axiarete’s application rationalization and technical-debt engagements. Formerly technology strategy at PwC and Deloitte.
Co-founder
Expert in enterprise systems, AI/ML and knowledge graphs — the foundations of the living twin. Former Oracle, Visa, Marqeta and Lattice.
Robots will close the labor gap. Lean will keep grinding out process gains. But the software that runs our factories is still built on thirty-year-old ideas, and it stands between reshoring announcements and reshoring results. Twenty-one sourced references.
Read the perspective →Why application intelligence has become the next competitive advantage. Introduces Application Health and its six dimensions, and the model for connecting software health to throughput, yield, cycle time and OEE. Written for semiconductor manufacturers; the argument holds for any plant.
Read the paper →How complexity is slowing decisions, eroding value and stalling AI. Four field cases from the authors’ careers in manufacturing IT, what they have in common, and a 90-day path out.
Read the paper →Send us the custom code and the MES model export — or the database workload reports and 90 days of logs — and within two weeks you will have a measured baseline of your key transactions, an inventory of every customization, and the first ranked list of what to fix. No workshops first. No deck about your industry. Evidence from your own estate. We would rather demonstrate on your systems than present slides.
Two hours a week from a project lead. Stage access only. An executive decision at week eight, on results.
We’ve received your request and will reply shortly to agree which inputs to start from.
Our deepest experience, and both production tuning engagements, are on Siemens Camstar / Opcenter Execution on Oracle, with .NET-based customization. The MRI, the twin, Application Health, and the portfolio, modernization and technical-debt solutions apply across the estate regardless of MES vendor.
Every change ships with its evidence package: baseline, test results against recorded production transactions, single-site deployment record, 72-hour monitoring log and rollback proof.
Both, in the right order. The twin shows which applications constrain the plant and which are redundant. Retire what’s redundant, modernize what blocks the future, and optimize the core you’re keeping — including the MES you’ll be running for years either way. The same evidence carries into the upgrade or migration when that is the right move.
With the MRI across all of them, because the consolidation decision depends on knowing what each system does. Then one plant, one transaction set, one baseline, and the change protocol above. Every estate read makes the next faster.
It is a platform operated with you by AxiareteForge, forward-deployed engineers. The platform reads, models, scores, ranks, drafts and monitors; the engineers validate, ship, watch and hand over. The platform section above states exactly which is which.
Each customer operates in a dedicated AWS environment with AES-256 encryption, customer-managed keys and zero-trust access, with data residency defined by agreement. Customer code, models and logs are not used to train Axiarete or third-party models. See Security, safety and data handling above.
First findings in week two. Quick wins — typically configuration and timeout corrections, index additions and the highest-volume query rewrites — in production by week four. In our second tuning engagement, 20–30% of the total latency gain landed in the first two to four weeks.
A project lead for two hours a week, subject-matter experts for roughly a day in total across eight weeks, and a sponsor for two hours. Everything else we take from the artifacts.
Reference numbers match Axiarete’s perspective Systems Are the Last Frontier of the American Manufacturing Revival; reference 22 is new to this page. Time-series figures (references 1 and 3) are as of the September 2026 releases. Engagement figures are from Axiarete engagement records; engagements are described by the work performed, not by the customer.