Speak the order, don’t tap it

Watch a sales rep take an order in a busy shop and you notice something the software rarely accounts for: they are not sitting down. One hand holds the phone, the other is on a carton. The shopkeeper is talking, serving a customer, and changing their mind about quantities. The rep is trying to tap through a form.

Everything about that moment argues against typing, and yet typing is what we have asked field teams to do for a decade.

What the interface costs in a real visit

Form-based capture is efficient in the abstract and expensive in the aisle. Each line is a search, a selection, a quantity field and a confirm — four interactions, at a screen the rep has to look at, in conditions that make looking down socially awkward and physically inconvenient.

The consequences are predictable and well documented in every field operation I have watched:

Orders get written on paper and typed up in the evening, which is where accuracy goes to die and where the timestamp stops meaning anything.
Activities go unrecorded. The shelf audit happened; nobody filled in the form, because the form was six taps and the rep was already walking to the next outlet.
Line counts shrink. Reps order the things they can find quickly, not the things the outlet needs — which quietly biases the entire assortment toward whatever sorts to the top of a list.

None of that is a discipline problem. It is an interface designed for someone sitting at a desk being used by someone standing in a doorway.

Speaking the visit

Voice capture inverts the default. The rep talks — "twelve cases of the 500ml, six of the one litre, and note that the display is down" — and the app parses that into a structured draft: two order lines and a merchandising observation, each shown for confirmation before anything commits.

Three design decisions matter more than the recognition quality.

It produces a draft, not a document. A spoken sentence never posts an order unseen. What appears is a normal set of lines the rep glances at and confirms — the same records they would have typed, arrived at differently. This is the difference between a feature people trust and one they turn off after the first wrong quantity.

It writes to the same records as everything else. There is no "voice order" entity. Dictated lines price against the same price lists, apply the same trade schemes, check the same credit limit and land in the same durable outbox as typed capture. Voice is an input mode, not a parallel system — which is why it inherits offline behaviour, permissions and audit for free rather than needing its own version of each.

Correction is conversational too. The natural follow-up to a misheard line is not opening a form; it is saying "make that eight". An interface that accepts speech for creation and demands tapping for correction has solved half a problem.

Why this matters more for compliance than for speed

Speed is the obvious benefit and the less interesting one. The larger effect is on what gets recorded at all.

Every field process has a set of steps that are genuinely performed and inconsistently logged — the competitor price noticed, the damaged stock spotted, the promise made about next week's delivery. They go unlogged because logging costs more attention than the observation was worth in the moment.

Lower the cost of recording to the cost of saying it out loud, and the recorded picture starts to resemble the real one. That is a compliance and data-quality gain, and it shows up in coverage analytics, in scheme claims that reconcile, and in the simple ability to answer "what actually happened at that outlet?"

The same pattern in field service

Everything above applies with more force to a technician, because a technician's hands are unambiguously busy and often dirty.

Checksheets. A twenty-point inspection is twenty taps at a machine, or one pass of spoken readings — "pressure normal, belt tension low, coolant at minimum" — parsed into the same structured checksheet the office expects.
Parts issue. Naming the parts consumed as they are fitted, rather than reconstructing the list at the van at the end of the job, which is where van stock reconciliation usually goes wrong.
Labour and job notes. Dictated at the point of work, while the detail is still precise.

The common thread: in both roles, the person is doing physical work in an environment that punishes screen attention, and the record has historically been assembled afterwards from memory. Voice moves capture back to the moment the thing happened.

What we are still careful about

Two things deserve stating plainly rather than being discovered by a customer.

Noisy environments are hard. A shop floor with a compressor running is a different acoustic problem from a quiet office, and any honest voice feature degrades there. The mitigation is design, not optimism: short utterances, visible interpretation, and a typing path that is always one tap away rather than a fallback the user has to hunt for.

Voice does not remove the need for good structure underneath. Parsing "twelve cases of the 500ml" into a line requires knowing the catalogue, the pack sizes, and what this outlet usually buys. The interface is the visible part; the reason it can work is that the item, the outlet and the price list are all one data model rather than three systems being reconciled.

What I would tell another design team

Watch the work before designing the form. Almost every field-app decision that looks strange from a desk makes sense once you have stood in the doorway of a shop at four in the afternoon.

And when you add a new input mode, hold it to the same standard as the old one: same records, same permissions, same offline behaviour, same audit. A capture channel that bypasses any of those is not a faster path to the data — it is a second source of truth wearing a friendly interface.