Enterprise software learned this lesson thirty years ago: nobody buys ERP from a platform vendor and installs it alone. SAP built an ecosystem of implementation partners; Salesforce did the same. The pattern exists because the hard part was never the software - it was wiring the software into how a specific organisation actually works. Manufacturing AI is now learning the identical lesson, at pilot-graveyard scale.
Where AI pilots actually die
Industry surveys have put the share of manufacturing AI pilots that never reach production somewhere between 70% and 90% for years. Having watched this from the shop-floor side, the causes are rarely algorithmic:
The data was never wired for it. Timestamps that disagree between systems, lot numbers that change format at a process boundary, parameters logged only when someone remembers. The model is fine; its inputs are fiction. Fixing this is unglamorous integration work that a remote vendor cannot do and an IT department deprioritizes.
The output lands nowhere. An anomaly alert with no agreed response is decoration. If the platform's recommendation does not connect to a daily management routine - who sees it, who decides, who acts, who closes the loop - it joins the andon lights everyone has learned to ignore.
Nobody owns it on site. The vendor owns the model, IT owns the servers, production owns the line - and the deployment as a thing that must produce value is owned by a steering committee that meets monthly. Orphaned deployments decay quietly.
The pilot proves the demo, not the economics. A proof-of-concept on curated historical data proves the technology works. It proves nothing about live operation on this line with these operators. Pilots designed without a hard operational baseline cannot graduate, because graduation was never defined.
The implementation-partner model
The fix is structural, and it mirrors the SAP/Salesforce pattern: separate the platform from its deployment, and put the deployment in the hands of someone local who owns the outcome.
This is how Jidokai delivers. The platform brings the technology - anomaly detection, batch traceability, causal reasoning, agentic workflows. Deployment is led end to end by a senior Lean Competence practitioner: someone who has run operations, speaks shift-floor and boardroom in the plant's own language, and is contracted against an operational outcome, not a go-live date. Behind the practitioner stand forward-deployed manufacturing and AI engineers who do the wiring - data plumbing, integrations, model configuration - on site and alongside the plant's own people.
The division of labour matters. The practitioner decides where AI should bite first (one line, one dominant loss, a baseline everyone accepts), builds the response routines into the existing daily management system, and owns the handover to internal owners. The engineers make the platform true to the plant. Neither role substitutes for the other - which is exactly why deployments that ship only one of them stall.
What to ask any AI vendor
Four questions expose whether a deployment model exists behind the demo: Who is on our site in week three, and what have they run before? What happens to a recommendation nobody acts on - show the routine, not the feature? What is the pilot's operational baseline and its graduation criterion? And who owns the deployment after handover - name the internal role, not a support tier. Vendors with real implementation models answer in specifics; platform-only vendors answer in roadmap.
FAQ
What does an AI implementation partner do that the platform vendor doesn't?
The partner owns the plant-side work: choosing the first use case against an operational baseline, fixing data plumbing on the pilot line, wiring recommendations into daily management routines, training internal owners, and being accountable for the operational result. Vendors ship capability; partners ship outcomes.
Why do so many manufacturing AI pilots fail?
The dominant causes are operational, not algorithmic: unreliable data foundations, alerts with no agreed response routine, no on-site ownership, and pilots designed without a graduation criterion. Estimates of pilots that never reach production have run at 70 to 90% across industry surveys for years.
How does Jidokai's deployment model work?
One senior lean practitioner leads the deployment end to end and is accountable for the operational outcome; forward-deployed manufacturing and AI engineers handle integration and configuration on site. The platform runs cloud-agnostic (AWS, GCP, Azure) or on-premises for regulated environments.
Should we start with a plant-wide AI strategy or a single line?
A single line with one dominant, measured loss. Plant-wide strategies without a producing reference line become slideware; a line that demonstrably makes money becomes the strategy.