70% of Manufacturers Still Collect Data by Hand
AI Reads Old Equipment Displays from a Photo
A Manufacturing Leadership Council survey of the NAM found that 70% of manufacturers still collect data manually. Not because they don't have a digital transformation strategy. Because the equipment on the factory floor — the production counter bolted to a 1990s press, the temperature display on a pre-IoT oven, the seven-segment panel on a legacy conveyor controller — was built before network ports were standard. Swapping the equipment costs six figures and stops production. Adding sensors means wiring, PLC programming, and protocol wrangling. This article covers the third option: photograph the display and let AI read the number directly from the image.
Key Takeaways
- 70% of manufacturers still collect production data by hand from working equipment displays, while 44% of the same manufacturers already need twice as much data as they did two years ago.
- The production count is already glowing on a working LED panel — the only thing keeping it out of your spreadsheet is a 2-foot gap bridged by a person with a clipboard, a task that one machinist described bluntly as "always fails."
- Photograph that same display and let ImageToTable.ai read the number — a 30-second read-and-scribble per machine drops to a 2-second photo, and every extracted row arrives with its source image as an audit trail.
70% Manual. Not Because They Chose It.
The Manufacturing Leadership Council survey didn't ask about preference. It asked about practice: how does production data actually get from the machine into the system? For 70% of respondents, the answer is "a person reads it and types it." That number isn't about Luddism or resistance to technology — 44% of the same respondents reported that data volume in their organizations had at least doubled in the past two years. The appetite for data is growing. The collection method is stuck.
Why stuck? Because the equipment generating the data wasn't designed for connectivity. A young manufacturing engineer on Reddit described the starting point plainly: "Our production line has virtually no data collection. I'm looking to start simple." His line wasn't broken. It just predated the era when machines came with Ethernet ports. That's the default state for tens of thousands of production lines worldwide: functional equipment, zero digital output.
70% of manufacturers collect data manually. 67% cite aging equipment as their top future challenge. 44% have seen data volume double in two years. The gap between "more data needed" and "same manual methods" is widening by the quarter.
For a broader look at how AI reads any type of meter or gauge — water, gas, electric, and industrial — from a photo, see our complete guide to AI meter reading without smart meters.
But There's Already a Display Showing the Number
This is the detail that gets overlooked in digital transformation discussions. Most legacy equipment already has a digital readout. The production counter on a stamping press shows the cycle count on an LED panel. The temperature controller on a curing oven displays the current setpoint. The flow totalizer on a chemical dosing system shows accumulated volume on a seven-segment LCD. The numbers exist. They're visible. They're accurate. They're just not connected to anything.
A manufacturing engineer setting up new manual production lines asked the forum how other plants track output. The answers ranged from "whiteboard and marker" to "operator tallies on a clipboard" to "someone walks the floor every hour with a notepad." A machinist tracking insert wear and part counts was more direct: "Most efforts to have the operators take tallies every time they index an insert after 'x' parts always fails and our data gets skewed." The data collection method isn't just slow — it's unreliable. Operators forget. Tallies drift. The numbers that reach the spreadsheet are a best-effort approximation, not a measurement.
| Equipment | Display Type | What It Shows | Current Collection Method |
|---|---|---|---|
| Stamping press | LED production counter | Cycle count / parts produced | Operator writes on whiteboard each shift |
| Curing oven | Digital temperature controller | Setpoint and actual temp | Hourly walkthrough with clipboard |
| Chemical dosing | Seven-segment flow totalizer | Accumulated volume dosed | End-of-batch manual log entry |
| CNC machine | LCD tool life counter | Parts per insert / tool | Operator tally — "always fails" (Reddit) |
| Conveyor system | Digital speed display | Line speed (units/min) | Rarely recorded unless there's a problem |
| Packaging line | Digital batch counter | Units packed per shift | Supervisor collects at shift change |
In every case, the display is right there on the machine — working, legible, showing exactly the number the plant needs. The problem isn't access to the data. It's the step between "the machine shows a number" and "the number is in the production system." That step is currently a person.
Why IoT Retrofitting Isn't the First Answer
The standard prescription for this problem involves sensors, edge gateways, PLC programming, and protocol conversion — typically OPC-UA or MQTT bridging to a data historian or MES. This works. It's the permanent solution. It's also expensive, slow, and surprisingly failure-prone: 58% of IoT project failures are attributed to device-level problems, and the IIoT market, while growing at 8.1% annually toward $286.3 billion by 2029, reflects a reality where carefully planned deployments take quarters or years to materialize.
For the manufacturing engineer trying to get baseline production data from a line that has never been instrumented, the IoT path asks for capital budget, vendor selection, procurement lead time, installation downtime, and an ongoing maintenance commitment — before a single data point reaches the dashboard. It's the right solution for a line that will run for another decade. It's the wrong first step for proving that data collection is worth doing at all.
McKinsey's research on factory digitization confirms the upside: up to 20% reduction in quality costs and up to 40% reduction in maintenance costs. But those returns assume the digitization project is complete. The 70% of plants still collecting data manually haven't started, and the gap between "no data" and "full IoT instrumentation" is wide enough that many never cross it.
The practical question isn't "should we install IoT?" The practical question is "can we get production data into a spreadsheet this week, from the equipment we already have, without taking the line down or spending capital budget?" For most plants still in the 70%, the answer to that question matters more than the five-year IIoT roadmap.
How AI Reads a Digital Display from a Photo
The mechanism is the same column-name extraction approach that handles analog gauges, utility meters, and field inspection data. It works on digital displays because the AI doesn't need to understand the display technology — it just needs to understand that what it's looking at is a number.
A seven-segment LED counter showing "8472" is trivial for a vision large model. The digits are standardized, high-contrast, and unambiguous. An old LCD panel with slightly faded segments is harder but usually within tolerance. The AI reads the display the way a person does — by recognizing the digit shapes — but without the drift in attention that makes manual tallies unreliable over an eight-hour shift.
In practice, the workflow is: operator photographs the display → uploads via Collection Link (no login required) → enters column names like "Machine ID, Cycle Count, Timestamp" → AI extracts the values → output is one Excel row per photo. The same batch processing from F1-F3 applies: photograph 30 displays across a shift, upload together, receive 30 populated rows. Read the full field data workflow →
What changes: the operator photographs the production counter instead of writing down the number. The 30-second "read-and-write" action becomes a 2-second photo. The end-of-shift manual tally — "always fails," per the Reddit machinist — is replaced by an Excel file where each row is backed by a timestamped photo.
Live Demo: Photograph a Display, Get the Reading
Upload any photo containing a readable digital display — a counter, a temperature panel, a seven-segment readout — to see the extraction in action.
Files are processed securely and not stored.
When to Photograph, When to Instrument
The decision between photo-based AI reading and full IoT instrumentation isn't philosophical. It's about frequency, criticality, and timeline:
| If you need... | Photo-based AI reading | IoT instrumentation |
|---|---|---|
| Data this week | Yes — start today, no hardware | No — procurement, installation, testing take months |
| Once-per-shift or daily readings | Ideal — operator photographs during existing rounds | Overkill for this frequency |
| Second-by-second real-time data | No — this requires continuous sensor input | Yes — the permanent solution |
| Proof of concept before capital request | Ideal — demonstrate value with zero investment | Wrong place to start |
| Data from 30 different machine models | Works — one AI, any display | 30 separate integration projects |
| Safety-critical alarm triggers | No — batch data, not real-time alerts | Yes — purpose-built for this |
The two approaches aren't competitors. They're sequential. Photo-based AI reading gets data flowing this week and proves the value of instrumented monitoring — building the case for the IoT retrofit that follows. The plants still in the 70% need the first step before they can justify the second. The same AI mechanism handles analog gauges on plant rounds →
Frequently Asked Questions
Does this work on faded or damaged displays?
Moderate fading is handled — the AI recognizes digit shapes even with reduced contrast. Dead segments (e.g., a seven-segment display where one segment is burned out and always off) can cause misreads, just as they confuse human readers. Severely degraded LCD panels with ghosting or bleeding may produce unreliable readings. The practical test: if a person can confidently read the number from the photo, the AI generally can too.
What about displays that show multiple values cycling through screens?
A single photo captures whatever was displayed at the moment of the shot. This works well for static displays and single-screen readouts. For equipment that cycles through multiple parameter screens (e.g., "temp → pressure → speed → temp"), you'll need one photo per screen you want to capture, timed to the display cycle. The AI extracts whichever number appears in each photo — the operator just needs to know which screen they photographed.
Can one person photograph dozens of machines in a single round?
Yes. A walkthrough that previously required stopping at each machine to read the display and write down the number becomes a walkthrough where the operator photographs each display and moves on. The batch is uploaded at the end of the round through a Collection Link. All readings populate one Excel file. The time savings compound with machine count: 30 machines at 30 seconds each (read + write + verify) is 15 minutes of data entry per round. Photographing 30 machines takes about 60 seconds.
What's the accuracy on different display types?
LED and seven-segment displays achieve near-perfect accuracy under normal conditions — the digit shapes are standardized and high-contrast. LCD panels with backlight are similarly reliable. Older reflective LCDs (no backlight, common on 1990s equipment) depend on ambient lighting during the photo — a well-lit shot produces accurate readings. Dot-matrix displays (character-based, not graphical) work well for numeric content but may struggle with mixed text-and-number formats.
Is the data traceable back to the source photo?
Yes. Each extracted row corresponds to one uploaded photo. For production environments where traceability matters, the photo serves as both the data source and the audit record. If a reading ever looks suspicious, the original image is available to verify. This is a capability that manual tallies can't provide — once the operator writes "8472" on the clipboard, no one can check whether the display actually showed 8472 or 8427.