Production Verified Real deployment at a named or precisely described company, with a measured result from a primary source.

AI video analytics times every cycle on manual assembly lines, so engineers find and fix the real bottleneck (HELLA, a Tier 1 automotive supplier)

Added 2026-10-26 · Last updated 2026-10-26

Manufacturers in this case: HELLA · all manufacturers using AI

Role

Production (manual assembly lines: cycle time, line balancing and standardized work) · Typical users: industrial or manufacturing engineer, lean or operational excellence lead, production supervisor or shift leader, plant manager, trainer

The problem

Much of assembly is still done by people, and the tools to understand manual work are time and motion studies: hard to scale, based on a few observed cycles, and biased by the observer. Engineers often work on the station they believe is the bottleneck. At a Tier 1 automotive supplier's plant in Guadalajara, Mexico, shift leaders each monitored two to four lines and, however diligent, missed deviations from standardized work. At HELLA, a German Tier 1 supplier of lighting and electronics, one operation required operators to turn their body 90 degrees to grab a bolt from a bin with the left hand while working with the right, and doing it the wrong way caused as much as a 10% productivity slowdown from one operator to another.

How AI helps

A camera above each station streams video; action-recognition neural networks turn the video into data: the sequence and duration of each step at each station, for every cycle. The system compares this with the standard work, shows bottlenecks on dashboards and can alert supervisors, with the video clip, when a step is missed or takes too long. Engineers then watch the clips behind the numbers to understand why, and change the station, the material feed or the training. At HELLA a small appliance on site pre-processes the video and the analysis runs in the cloud. People decide every change; the data is meant for process improvement.

Step by step

Pick one manual line where output or quality matters most and write down today's figures: target and actual cycle time per station, output per shift, defects and scrap.
Before any camera goes up, explain the purpose to operators and their representatives, and agree in writing what the video is used for (process improvement, not individual discipline), who can see it and how long it is kept. Check privacy and labour law where you operate.
Start with a few stations and compare the measured cycle times with a short manual time study to check the data.
Look at the distribution of cycle times per station, not only the average: the long tail and the micro-stops are where the video helps most.
Watch the clips behind the slowest cycles with the operators, and test simple countermeasures (station layout, material presentation, tool position, training).
Measure again after each change and share the results with the line, as HELLA did with periodic reports per shift.
Extend to other lines only once the first one shows a measured gain and the operators accept the system.

Tools

Drishti (action-recognition video analytics for manual assembly, with TRACE video and FLOW analytics): used at HELLA Dhankot and the Tier 1 supplier in Guadalajara
Other tools to evaluate: other video analytics tools for manual assembly lines such as Retrocausal, Invisible AI and PowerArena, and the cycle-time features of your MES
For a first offline study: any LLM with a data-analysis or code tool (ChatGPT, Claude, Copilot, Gemini) on an export of station cycle times, with the prompt below

In the AI tools index: Drishti · all AI tools used by manufacturers

Result

HELLA, sensor assembly line, Dhankot plant (India), Drishti: the cycle-time data revealed slowdowns at stations that were not the focus of improvement, the video showed that the station layout forced a twisting reach that tired operators, and the team reconfigured the station and improved line balancing; the investment paid for itself in less than six months. HELLA's head of operational excellence says the company had already reduced cycle times at its Guanajuato, Mexico plant in 2020 (ASSEMBLY magazine, December 2022). On the HELLA line with the bolt-reach problem, the team recaptured all of the 10% lost efficiency in three to four months (Embedded Computing Design, June 2022, from Drishti). A Tier 1 automotive supplier in Guadalajara, Mexico, not named: by acting on standardized-work deviations, it lowered the defect rate on those lines by 30% in 12 weeks and saved an average of $10,000 per defect; a kaizen using more than 1,000 automatically timed cycles found ways to improve efficiency by 11% and reduce scrap by 15%, and took 50% less time than a traditional event (article by Drishti's founder in SME Media, May 2021).

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

Cameras on workstations raise privacy and trust issues: involve operators and their representatives from the start, use the data to fix the process rather than to discipline individuals, and check privacy and labour law (for example GDPR in Europe, Law 25 in Quebec, works council rights in Germany). In Guadalajara, operators were nervous at first about what they saw as constant surveillance.
Cycle-time gains come from the countermeasures people design, not from the cameras; budget engineering time to act on the data.
The HELLA results come from trade press articles based on Drishti and HELLA statements; the Guadalajara figures come from an article by Drishti's founder about an unnamed customer. None are independently audited.

Try it yourself

You are an experienced industrial engineer for a manual assembly line.

I will paste: (1) cycle time data, one row per cycle and station [timestamp, shift, station, cycle time in seconds, optional step times, optional flags such as missed step or rework]; (2) the standard work for each station [steps, standard time]; (3) the takt time and the line's output per shift.

Do this:
1. Describe the data: number of cycles per station and shift, gaps, and obviously wrong values.
2. For each station, give the formulas or code to compute median, 90th percentile and spread of cycle time, and the share of cycles above takt. Do not compute them in your head.
3. From the results I paste back, identify the true bottleneck and the stations with the most variation, and explain why the average alone can mislead.
4. List likely causes to check on video or on the floor for the slowest cycles (material presentation, reach and posture, tool position, waiting for parts, rework), as hypotheses, not conclusions.
5. Propose 3 countermeasures to test, each with the measure that will show whether it worked.
6. Draft a short, plain explanation for operators of what is measured and why, focused on the process rather than on individuals.

Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter

Related cases

All 12 Production use cases →