A small manufacturer decides to digitise. The proposal that arrives covers production planning, machine monitoring, quality, maintenance, inventory, procurement and dispatch, delivered over eighteen months.
Twelve months later, some of it is running, most of it holds data nobody trusts, and the shop floor has gone back to the register. This is the common outcome, and the cause is not the technology. It is that a programme spanning seven areas has no single point where it visibly works, so nothing accumulates enough credibility to survive the first difficult quarter.
A first year that does three things properly beats one that attempts everything.
Why breadth fails here specifically
Manufacturing digitisation has a property that makes over-scoping unusually damaging: the data is produced by people who are busy doing something else.
An operator running a machine is not a data entry resource. Every field you ask for competes with the actual job, and under pressure the actual job wins. A system requiring fifteen entries per shift will receive them for a fortnight and then receive fiction — which is worse than nothing, because decisions get made on it.
The design constraint follows directly: capture the minimum that produces a usable answer, and capture it as close to automatically as the plant allows.
The order that works
One: know what was actually produced
Before anything sophisticated, establish a reliable record of output — what was made, when, on which line, and how much was rejected.
This sounds trivially simple and is frequently absent. Production is known approximately, from a supervisor's register, reconciled monthly against dispatch. Approximately is enough to run a plant and not enough to improve one.
A dependable output record is the foundation everything else sits on, because every subsequent question — efficiency, cost, capacity, delivery reliability — is a ratio with output in it. The mechanics are in production monitoring systems for small manufacturers.
Two: know what you actually hold
The second area is material. Raw material, work in progress, finished goods.
Manufacturers usually discover here that the recorded stock and the physical stock have diverged substantially, and that the divergence has been absorbed by purchasing extra. That absorption is invisible in the accounts and expensive.
Getting stock accurate is less about software than about the discipline of recording movement at the point it happens — the argument in what actually prevents stock variance.
Three: connect production to orders
With output and material both reliable, the third step is linking them to what was sold: which order a run belongs to, what remains, what can be promised.
This is where digitisation starts paying commercially rather than administratively. A plant that can answer "when will this be ready" accurately has changed how it sells, not just how it records.
What to leave for year two
Machine-level monitoring. Valuable, and it depends on knowing what normal output looks like — which is step one.
Predictive maintenance. Requires history you do not have yet.
Full quality management. Start by recording rejections against runs. The complete system can wait.
Procurement automation. Meaningful only once consumption data is trustworthy.
Anything described as AI. In manufacturing this almost always means pattern detection over historical data, and the honest position is that it requires the history first. What the term should mean is set out in what "AI-powered" should mean in business software.
The pattern is consistent: the advanced capabilities depend on the basic records. Attempting them first produces impressive interfaces over unreliable data.
The capture problem, which decides everything
Every plant we have worked in has the same constraint: whatever the operator has to do must fit inside the work, not alongside it.
What that means practically:
Fewer fields. Every field must justify itself by changing a decision. Fields collected "for analysis later" are how systems become unreliable.
Entry where the work is. A terminal at the line, not a computer in the office that gets updated at shift end from memory.
Tolerant of interruption. A half-entered record during a machine stoppage must survive.
Reconcilable. A supervisor must be able to correct a mistake, and the correction must be visible rather than silent.
Where these are wrong, the system produces confident numbers that are not true — the worst available outcome, and the reason factory operations: what to digitise first treats capture design as the primary decision rather than a detail.
What Truffaire has actually built
We have delivered production monitoring and factory operations systems as part of ten deployments across ten clients, alongside inventory, warehouse, logistics and billing systems.
What that experience supports is a claim about sequence and capture design, not a claim about your specific process. We have not published outcome figures for those deployments and we are not going to estimate them here.
The traceability layer that becomes possible once output and material are reliable is covered separately in production traceability for small manufacturers.
Frequently asked questions
How much should a first year cost?
Less than a full ERP programme, because the scope is three areas rather than seven. The figure that matters is not the build — it is the ongoing cost, which should be agreed before you commit.
Should we buy an off-the-shelf ERP instead?
Sometimes. It depends on whether your process is standard enough to adapt to, which is a genuine question rather than a rhetorical one — examined in off-the-shelf vs custom business software.
What about machines that cannot be connected?
Most small plants have a mix of old and new. Manual entry at the line is a legitimate answer for older machines, provided it is minimal. Waiting until everything is connectable means waiting indefinitely.
Will the shop floor accept it?
If it is faster than what they did before and if they were asked before it was designed. If neither, no — and the reasons are in training staff on a system they didn't ask for.
How long before it is useful?
Reliable output data starts changing conversations within weeks. The commercial benefit of order-linked production takes longer, because it depends on a few months of trustworthy history.
Where to start
Ask what you produced on a specific day three weeks ago, and how much was rejected.
If answering requires finding a register and someone's recollection, that is year one — and it is a smaller project than the eighteen-month programme, with a far better chance of still being in use next year.
If you want a manufacturing digitisation sequence built around your actual constraints, get in touch.