Right Arrow

TABLE OF CONTENTS

Grey Down Arrow

Continuous improvement on video evidence, not anecdotes

Continuous improvement in manufacturing on every run, not one walk: Spot AI scores changeovers and SOP steps, so a change is verified and rechecked by day 90.

By

Rish Gupta

in

|

14 min

|

Continuous improvement on video evidence, not anecdotes

Continuous improvement in manufacturing on video evidence, not anecdotes

Updated November 2026

Continuous improvement in manufacturing holds up when the baseline, the test, and the verdict all come from the same record of every run, not from one observed shift and a few recollections. Spot AI's AI Operations Assistant gives a continuous improvement (CI) lead that record on the cameras a plant already has: it scores every changeover by phase and checks every SOP run step by step, so a change is reviewed over a planned window of runs instead of one walk. Two old rules set the bar: Walter Shewhart's rule of thumb asks for at least 25 in-control samples of four before calling a process stable (Source: NIST), and EPA's lean guidance describes follow-up events at 30 and 90 days after a Kaizen event to check that the gains are sustained (Source: EPA).

Key takeaways

  • Run the loop on the record of every run: a baseline from footage, a written hypothesis, one change, and a review window planned in advance.
  • Stability and improvement are separate questions; Shewhart's 25 in-control samples of four concern stability, and one observed shift is a single sample.
  • Spot AI's AI Operations Assistant scores each changeover by phase and flags steps that were skipped, done wrong, or done out of order, on the cameras a plant already has.
  • Re-measure at 30, 60, and 90 days on every shift, not only the one the CI lead works.
  • Use footage for coaching, not discipline: Spot AI has found that weekly or biweekly check-ins tied to positive incentives answer the "Big Brother" worry.

Why does continuous improvement stall on anecdotes?

Improvement work stalls when the evidence behind a decision is a sample of one: the shift the CI lead walked, the changeover a supervisor remembers, or the operator's account of what usually happens. Spot AI's AI Operations Assistant replaces that sample of one with a record of all the runs, each checked against the standard. On a plant running three shifts five days a week, a Gemba walk through one shift sees one shift in fifteen, under 7% of the week. Supervisors can't cover every line on every shift either, so SOP compliance often comes down to occasional spot checks with little structured follow-up, and crews drift into their own versions of the process.

The move from assumption to evidence is mostly a change in where each answer comes from. The table below pairs the questions a CI lead asks with what the anecdote says, what the full record shows, and where that record comes from:

Question a CI lead asks

What the anecdote says

What the full record shows

Where it comes from

Where does the changeover lose time?

"Teardown takes forever."

The phase with the widest spread across runs, which may be the first-piece checks rather than teardown

Spot AI's AI Operations Assistant scores each changeover by phase

Does every shift run the standard?

"Nights skip the purge check."

Which steps were skipped, done wrong, or done out of order, by shift and crew

Spot AI's step-by-step SOP checks

Did the last Kaizen work?

"It feels faster."

Phase times and step flags for the runs before and after the change

Spot AI's weekly reports by line, shift, and crew

Why did that run go long?

The operator's recollection

The footage of that run, found with a plain-language question

Spot AI's Iris, which answers questions about what the cameras saw

Who needs coaching on what?

The supervisor's impression

Each operator's scorecard, with video of what went well and what to adjust

Spot AI's operator scorecards after each run


The Gemba walk stays: it is where a CI lead talks with the crew and sees constraints a clip can't show, and the footage decides which line to walk, as in AI-assisted Gemba walks.

What does the improvement loop look like with video at each step?

The loop stays the familiar one (baseline, hypothesis, change, and verification), and video changes the evidence at each step rather than the steps themselves. Spot AI's AI Operations Assistant supplies that evidence from the cameras a plant already has. Here is how each step differs:

Loop step

With one observed shift

With every run on video

Baseline

A stopwatch study of the runs someone watched

Phase times and step flags for every run in the baseline weeks

Hypothesis

The loudest explanation in the meeting

The slowest and fastest runs of the worst phase, watched side by side

Change

Announced at a shift meeting

Written into the standard and checked step by step from the first run

Verification

One follow-up walk

A planned window of like-for-like runs compared with the baseline, by phase, shift, and crew


Step 1: set the baseline from footage

Pick one process with a clear cost. A changeover on a constrained line is the usual first choice, because every minute over standard is production the plant has already paid for and didn't get. Define its phases and the events that start and stop it, such as the last good part of the old run and the first good part of the new one.

Some plants hold back because the standard isn't written down. Many Spot AI customers find that implementing the tool is itself a forcing function to standardize and document the process: the first baseline is how the standard gets written.

Size the baseline before it starts. For a 95% confidence interval on the average, with a known standard deviation, NIST's sample-size formula says n is at least (1.96 × standard deviation ÷ allowed error) squared, so the runs needed grow with the square of the spread (Source: NIST). As an illustration, if comparable changeovers (the same product-to-product transition on the same line) vary with a standard deviation of 6 minutes and the CI lead wants the average within 2 minutes at 95% confidence, that is about 35 changeovers; a standard deviation estimated from the data calls for a somewhat larger sample. That sizes the baseline average; judging a change is a separate comparison, in step 4. At two changeovers a shift on three shifts, footage collects them in about six working days; a stopwatch study that catches one changeover per walk needs 35 walks. For mapping a whole flow with measured times, see value stream mapping with real cycle and wait times.

Step 2: write the hypothesis from the worst phase

Spot AI scores each changeover by phase, so the CI lead can start from the phase with the widest spread rather than the longest total. Watch the five fastest and five slowest runs of that phase side by side, and ask Iris in plain language what the cameras saw during the slow ones. Often the difference is concrete: a cart that isn't staged, a tool fetched mid-phase, a check done twice. Waiting and motion, two of the eight wastes of lean manufacturing, often show up here.

Write the hypothesis so the next runs can prove it wrong. For example: "If the die cart is staged at the press before the line stops, the teardown phase median drops from 14 minutes to under 10."

Step 3: make one change and put it in the standard

Change one thing at a time, so the comparison points to one candidate cause. Add the new step to the standard work, brief every crew, and let Spot AI's step-by-step SOP checks flag any run where the new step was skipped, done wrong, or done out of order. That separates two questions that anecdotes merge: did the crews run the change, and did the change work?

Step 4: verify over a planned review window

Set the review window before the change goes in: define the smallest improvement worth detecting and size the comparison from the baseline's variation, not the baseline count; for a frequent changeover, the next hundred runs is a practical window to plan against. Keep the same product transitions, equipment, and operating conditions as the baseline, and spot-check detected phase boundaries against the clips. Then compare the target phase's median and spread with the baseline, by shift and crew, next to the new step's flags. Whether the process has settled is a separate question: Shewhart's rule (25 in-control samples of four) concerns stability, not proof of improvement. The combinations point to different next moves:

What the review window shows

What it suggests

Next move

New step followed; phase time down on every shift

Evidence consistent with improvement

Write it into the standard and pick the next phase

New step followed; phase time unchanged

No improvement detected in this comparison

Go back to the slowest runs and write a new hypothesis

New step skipped on some shifts; phase time down where it runs

Evidence of improvement where the step is run

Coach the shifts that skip it before judging the change

Target phase down; other phases up

Work may have shifted between phases

Check where it went, and judge the whole changeover


The same phase-level view is how plants find hidden capacity on the lines they already have.

Separate adoption from effect before calling a Kaizen a failure. Spot AI's step flags show whether crews ran the new step, and the phase scores show whether it worked. Read them in that order.

How do you make the gains stick at 30, 60, and 90 days?

Re-measure the same process against the baseline at 30, 60, and 90 days, reading every shift rather than only the one the CI lead works; Spot AI's weekly reports summarize performance by line, shift, and crew, so each check is a reading rather than a new study. EPA's lean guidance describes follow-up events at 30 and 90 days after a Kaizen event to assess performance and find the adjustments needed to sustain the improvements (Source: EPA); a 60-day check adds a midpoint before small drift becomes the new normal. The three checks ask different questions:

  1. Day 30, adoption. Is the new step run on every shift and by every crew? If one shift skips it, coach that shift before judging the change.
  2. Day 60, holding. Has the phase time held on all shifts, including crews with new operators? A spread that widens again is an early sign of drift.
  3. Day 90, standard. Is the change now how new operators learn the process, and which phase is the longest now? The answer to the second question is the next turn of the loop.

Approved customer reports cover different measurement periods. A Fortune 500 packaging company reports changeover time down from 28 to 21 minutes over six months, worth $15M in added throughput per plant per year at zero new capex (customer-reported). A $12B manufacturer reports changeover time down 15% in the first three weeks with Spot AI's AI Operations Assistant (customer-reported). Neither report describes a 30, 60, and 90-day routine; the checks are how a plant confirms that its own gain holds.

What does visual feedback look like on the floor?

Visual feedback works when each role sees the evidence it can act on, close to the run it describes. Spot AI's AI Operations Assistant flags deviations in real time and splits the record by role:

  1. Operators get a scorecard after each run, with video of what went well and what to adjust.
  2. Supervisors get a shift review: a stage-by-stage summary, a task timeline, and key-moment clips for the whole shift.
  3. CI leads get phase scores and step flags across runs, the material for the improvement loop.
  4. Operations leaders get trend dashboards that compare plants side by side, plus weekly reports by line, shift, and crew.

The best runs double as training: a veteran's fastest clean changeover becomes the reference new operators learn from, the subject of turning a veteran's know-how into the standard. Silver Bay Seafoods, which gave managers across its plants one Spot AI view, reports a 15% gain in operational efficiency in its public customer story.

How do you use footage for feedback rather than discipline?

Use it in regular coaching, not in incident hunts. Spot AI has found that weekly or biweekly check-ins that tie positive incentives to hitting goals are a positive way to use the AI Operations Assistant, and they answer the objection crews raise first: "It sounds like Big Brother." HR and unions often raise the same worry, so the answer has to show in how the program runs:

  1. Coach the step, not the person. Feedback names the step and shows the clip; it doesn't rank people.
  2. Start from what went well. Spot AI's operator scorecards include video of what went well, so the conversation opens on the best part of the run.
  3. Agree who sees what before go-live. Spot AI's role-based access control sets user permissions at the camera and feature level.
  4. Keep the video in the plant. Spot AI's Intelligent Video Recorder (IVR) keeps full-resolution video on site, and only metadata travels to the cloud.
  5. Bring HR and union representatives in early. Process footage can show identifiable operators, so show them a real scorecard before crews see one and agree who may access the footage, how long it is kept, and whether individual scorecards may be used in employment decisions.
  6. Let operators help write the standard. A standard the crew helped define is one they can check their own runs against.

"We're not spying on anyone, but we can't be out there watching 24 hours a day. Spot AI helps us spot the processes that need to change."

Kevin McDaniel, Director of Safety and Asset Protection, Unique Industries

For the wider rollout conversation with crews, HR, and unions, see answering the workforce question before rollout.

A lean leader's checklist for the first 90 days

The checklist turns the loop into a plan for the plant manager. Spot AI's forward-deployed engineers build AI Operations Assistant deployments from a paid pilot to a production rollout, so most of these items are decisions for the plant, not setup work:

  1. Name the process, the number it should move (for example, median changeover minutes on one line against its standard), and the person who owns that number.
  2. Confirm which cameras see each phase. Spot AI runs on the cameras a plant already has, so list any phase no camera covers before the baseline starts.
  3. Write or confirm the standard and its phases, with start and stop events.
  4. Set the baseline period and size it with the sample-size arithmetic above. Keep setup, minor stops, and waiting under distinct loss codes in the MES, and compare the same product-to-product transitions before and after a change, with a separate standard where a transition's requirements differ.
  5. Agree access rights and the check-in cadence with the plant manager, HR, and, where there is one, the union.
  6. Take the IT/OT and security review early: Spot AI is SOC 2 Type II, supports SSO through the customer's identity provider, and keeps full-resolution video on site.
  7. Decide where results go. Spot AI's open APIs, webhooks triggered by AI agent detections, and a read-only Model Context Protocol (MCP) endpoint let the plant combine video results with MES, maintenance, or quality data.

Book the 30, 60, and 90-day reviews on the day the change goes in, and read all shifts at each one. A review already on the calendar is the one that happens, and a gain checked on one shift says nothing about the other two.

Limits worth planning for

Four limits to plan around:

  • Steps inside a machine or behind a guard need machine or sensor data.
  • Footage shows what happened, not why someone did it; the conversation at the Gemba still confirms the cause.
  • A wrong standard gets followed as reliably as a right one; review the standard itself at day 90.
  • Low-volume processes take longer to verify: a changeover that runs once a week takes months to fill a useful review window.

Spot AI's part in the loop is narrow: it keeps a record of each run, so the baseline, hypothesis, and verdict rest on the same evidence. See how Spot AI approaches continuous improvement on the AI Operations Assistant page, from changeover scoring by phase to step-by-step SOP checks and operator scorecards.

Key terms

  • Baseline: the measured performance of a process before a change, taken from enough runs to show its normal spread.
  • Changeover phase: one part of a changeover, such as teardown, setup, or first-piece checks; Spot AI's AI Operations Assistant scores each phase separately.
  • Standard work: the agreed, written way to run a process, with its steps in order; the reference each run is checked against.

Frequently asked questions

What is continuous improvement in manufacturing?

Continuous improvement in manufacturing is the ongoing loop of measuring a process, finding what slows it, changing one thing, and confirming the change held; Spot AI's AI Operations Assistant supplies evidence for each step from the cameras a plant already has. Kaizen emphasizes ongoing, employee-led improvement; Lean focuses on removing waste and improving flow; and DMAIC structures a project as Define, Measure, Analyze, Improve, and Control.

How is video used in continuous improvement?

Spot AI's AI Operations Assistant uses existing cameras to score each changeover by phase and check every SOP run step by step, flagging steps that were skipped, done wrong, or done out of order. A CI lead uses that record to set a baseline, choose the phase to fix, and verify the change on the runs that follow.

How do you sustain continuous improvement gains?

With Spot AI, a plant can re-measure the same process at 30, 60, and 90 days from weekly reports that summarize performance by line, shift, and crew, rather than scheduling new studies. Check adoption of the new step first, then whether the time held on every shift, then whether new operators learn the new standard.

Is using cameras for process improvement the same as monitoring workers?

Spot AI's AI Operations Assistant records process footage that can show identifiable operators, so the honest answer is set by the plant's own rules: agree with HR and worker representatives before rollout who may access the footage, how long it is kept, and whether individual scorecards may be used in employment decisions. Spot AI supports those rules with role-based access control at the camera and feature level and full-resolution video kept on site, and Spot AI has seen weekly or biweekly check-ins tied to positive incentives work well.

How many runs should a CI lead review after a change?

Spot AI's AI Operations Assistant checks every run, so the review window can follow the process rather than the observer's time. Before the change, define the smallest improvement worth detecting and size a like-for-like comparison from the baseline's variation at an agreed confidence and statistical power; the baseline arithmetic in step 1 estimates one mean and does not size this comparison, and stability is a separate check.

About the author


Rish Gupta, CEO and Co-founder

Rish Gupta is CEO and Co-founder of Spot AI, leading the charge in business strategy and the future of video intelligence. With extensive experience in AI-powered security and digital transformation, Rish helps organizations unlock the full potential of their video data.

Tour the dashboard now

Get Started