Elephant Index
목록으로
AI Nativeness23분EN

AI Compressor · Which Lead Times Should Manufacturing Software Companies Reduce?

From discovering shop-floor needs to product release and the second customer

연재 「AI Compressor in Manufacturing」 2번째 글

한국어 원문

AI Compressor, Part 2. A beam of blue light passes through two transparent gates, illustrating the move from a shop-floor need to a second customer.

The quality team at an automotive parts plant asks a software company to “use AI to create inspection reports.” Following one real case reveals that writing the report takes only 20 minutes. Matching inspection-machine results with the manufacturing execution system (MES) by serial number and establishing the final disposition after reinspection take longer. If the supplier ships document generation first, the quality specialist still checks the same three screens while reviewing a faster draft.

A company building AI products for manufacturing has two sets of lead times to reduce. Customers need to identify a problem and resume production, or demonstrate compliance after receiving delivery requirements. Suppliers need to understand the site, validate an initial feature, and apply the result at another customer. This article examines the robot and software investments discussed in Part 1 through the supplier's product strategy. Part 1 · AI Compressor and manufacturing investment

From customer needs to a product and the next contract
  1. Observe the workflowFollow one recent defect through
    retrieval, waiting, and approval
  2. Release a small featureOne product and one machine
    from history retrieval to approval
  3. Validate quality and timeCheck omissions and disposition errors
    and on-site support effort
  4. Deploy at the next customerReuse common features
    and validate the differences

Use errors and waiting observed in operation to prioritize the next feature.

Read the first flow from the beginning of the customer's work to its completion. Before developing the requested automated report, identify the retrieval, waiting, and approval steps that actually consume time. Build a feature that completes that work for one product, then measure omissions, disposition errors, and on-site support time together. At the second customer, check equipment and data differences. Reusing the common features while validating those differences is how the supplier can shorten deployment time.

Validate a feature request against one real case

Summarizing interviews and categorizing support requests can accelerate research preparation. A product manager can use the time gained to review one recent defect with an operator and a quality specialist. Record what triggered the work, where evidence was found, whose confirmation was awaited, and how often information was reentered. Checking whether the same problem occurs across shifts and customers builds a basis for the roadmap.

In Harvard Business Review, Andrew Ng described the limits of applying internet companies' large-data playbook directly to manufacturing. Inspection problems and available data vary by process and product. Applied to this case, that suggests first identifying customer differences and choosing a recurring task as the product's initial scope. An MIT Sloan article on industrial AI also explains how getting the right information to the people making shop-floor decisions helps organizations assess and improve outcomes during operation. Andrew Ng, Harvard Business Review, 2021 · Sara Brown, MIT Sloan, 2025

For a first site visit, examine a recent defect-handling record, approved work instructions, and an inspection report that was actually submitted. Where possible, compare two cases on the same line with different products. Result storage, approval of reinspection, and handoff during the night shift may differ. AI can identify recurring phrases in interviews and support records. Return to the source material to distinguish observations from participants' assumptions, then establish whether the cause is missing data, an authorization wait, or inconsistent work criteria.

Turn the findings into buying conditions. In an “automate inspection reports” request, the quality specialist wants faster access to source records, the plant manager wants deployment without stopping the line, and IT reviews where customer data will be stored. Acceptance criteria differ within the same customer. Confirm the sponsor, budget, and schedule, then start with a deadline-driven need such as a quality audit or supplier qualification. That makes it possible to define the pilot's scope and the purchasing decision date together.

Scope the first release around one task quality staff can complete

At this plant, the first feature could retrieve inspection history for one standard product. When a user enters a serial number, the product shows the criteria version used, each inspection result, any reinspection, and the approved final disposition, with links to original records. AI compares OEM requirements with available records to find gaps. System rules check the relationship between the final disposition and the underlying records, and a quality specialist approves the report.

Before release, test normal cases, reinspections, missing results, and changed criteria with the customer. Run the trial alongside the existing process for selected products. Measure time from serial-number search to approved evidence, missing records, agreement with the final disposition, and on-site engineering effort. Even if weekly support work falls by six hours, trial production still requires a continuous line slot and quality staff availability. Confirm that allocation to explain which customer milestone has moved forward.

A Siemens-sponsored MIT Technology Review Insights report examines virtual testing of products and production processes using digital twins. The Economist also discusses simulations making design and process validation faster and less costly in manufacturing automation. Where line access is scarce, a software supplier can first reproduce a problem using historical records or a test environment, then use the result to define the on-site trial. MIT Technology Review Insights, The Emergent Industrial Metaverse · The Economist, The “ChatGPT moment” has arrived for manufacturing

Define the implemented pilot narrowly. Connect one inspection machine for one approved product and gather records for a serial number so that the responsible person can approve submission evidence. As a hypothetical example, handle 20 comparable requests over two weeks. Agree beforehand to measure median retrieval-to-approval time, cases with omissions, cases showing incorrect records, and engineering intervention time. With a small sample, inspect exceptions individually and extend observation before deciding.

Use the initial operating results to scope the second release. If staff repeatedly reopen source records, improve source links and time and unit displays. If reinspection approval creates the longest wait, improve handoff and notifications. If departments perform similar work with different forms, consider self-service report configuration. Agree on the target time reduction and acceptable error criteria for each update so that its results can be shown to the next customer.

Measure from completed code to approved deployment

The supplier's developers can use AI to find field names in equipment documents and draft integration code or tests. They still need to establish whether the result handles machine-specific time zones, units, missing values, and retransmission correctly. Extra review and on-site fixes can transfer time saved in coding into commissioning. Record requirements-to-site-acceptance lead time, post-deployment corrections, and engineering hours per customer alongside coding time.

In the Quarterly Journal of Economics, Brynjolfsson, Li, and Raymond reported that AI assistance increased customer-support agents' issues resolved per hour by 15% on average. In METR's 2025 randomized experiment, experienced developers working on familiar open-source projects took 19% longer to complete tasks when they could use the AI tools available at the time. The studies involved different work, participants, and tools. For manufacturing software teams, they provide a reason to measure completed outcomes separately in support and development. Brynjolfsson, Li, and Raymond, Generative AI at Work · Becker et al., METR developer productivity experiment

The Siemens–Microsoft collaboration illustrates how industrial knowledge can be combined with AI development capabilities. In Microsoft's published thyssenkrupp Automation Engineering case, engineers use Industrial Copilot while the company supplies task-specific data and component-level engineering. The account says thyssenkrupp controls its proprietary data and does not share product-specific data with Siemens. This supplier-published case illustrates the importance of letting customers determine who may access which records when bringing industrial AI into established work. Microsoft Source, Siemens and thyssenkrupp case study

Define the respective roles of AI and rules in the product. AI can identify candidate products from handwritten notes or attach evidence links to changes in customer requirements. The product checks serial numbers against criteria versions and controls approval through user permissions. If mandatory source records are missing or sources conflict, it flags what needs review. Retain the records and criteria used so that quality staff can reconstruct the facts.

Engineering teams turn field exceptions into validation cases for the next release. Retest retransmission of the same serial number, differences between local and server time, and selection of an outdated form after a model update. Deploy to a few lines first and check the time needed to return to the existing approval process. If AI drafts integration code faster but field corrections and support hours rise, recalculate deployment lead time.

Operating results determine the next update
  1. Review operating recordsWhat prevented
    completion of the task?
  2. Address the causeMissing records → validate integration
    Approval waits → improve handoff
  3. Remeasure the same workTime to approval
    Corrections and support cases

What the manufacturer measures
Evidence preparation, reinspection, and audit deadlines

What the software supplier measures
Integration and validation time, support cost, and contract conversion

The second flow shows how to choose the next feature after release. If missing history prevents completion, improve equipment integration and record validation. If approval waits dominate, improve handoff and notifications. On remeasurement, the manufacturer checks evidence preparation and reinspection time, while the supplier checks integration, validation, and support effort. Together, these results show whether customer work is getting faster and repeat deployment is becoming easier.

The second customer reveals the supplier's market

After solving the first customer's inspection-history problem, establish why another customer would pay for the same capability. The first may be preparing for a new automaker's supplier audit, while the second is responding to a quality dispute with an existing buyer. Both need rapid access to source records for individual parts and evidence of the criteria and final disposition. Serial numbers, inspection timestamps, criteria versions, and approval history can become common product features, with equipment data connections and submission formats configured for each site.

Start market development with customers that repeat this work and have a clear reason to buy. Sales confirms events such as audit deadlines or contract requirements that create purchasing urgency, along with quality, production, and IT acceptance conditions. Start a paid pilot with one line and one product, then use verified records and time improvements to support the next line's deployment. Providing equipment makers or site integrators with the same validation procedure and integration scope can expand the customer base a small engineering team can support.

At the second customer, price common features and additional work explicitly. Common capabilities include inspection retrieval, source links, and approval history. Agree on new equipment connections, dedicated forms, work scope, and acceptance items before contracting. Suppose the first deployment takes 80 engineering hours and the second takes 45. Record machine counts, data condition, and validation scope to establish whether the difference comes from a reusable connector or simpler equipment at the second site.

Use the workflow to define expansion into another industry too. Equipment maintenance offers the interval from work order to approved recovery. High-mix manufacturing offers the interval from order confirmation to approved operating settings. The responsible people and reasons to buy differ across markets, so the case for the same product needs fresh evidence. First checking pilot conversion, support time per customer, and additional-line deployment time within the inspection-history segment helps determine how much engineering time can be allocated to exploring another market.

  • Manufacturing customers track evidence-preparation time, reinspection time, audit deadlines, and the products using the system after an order is won.
  • Suppliers track contract-to-first-useful-result lead time, integration and validation hours per customer, operating support, and pilot conversion to a full contract.
  • Use field support records to prioritize the next release. Frequent omissions call for better integration and record validation. Time spent changing evidence formats may justify configuration features.

A plant's request becomes a market opportunity when the next customer completes similar work sooner and the supplier also spends fewer engineering hours on installation and validation. The proposed outcome becomes concrete. Starting from a feature such as AI document generation, the supplier can specify which records will be validated during the first trial and when submission evidence will be ready.

목록으로