Skip to content
All insights

Workflow automation

The internal relay is the reason answers take two days.

A customer asks where their order is. The person who reads the email cannot see the answer, so they forward it. Two days later someone replies. The question took thirty seconds to answer and two days to reach the person who could.

Published April 14, 20266 min read

What this covers

  • Most operational requests are lookups across two or three systems, not problems requiring judgment.
  • Routing accurately is often worth more than answering automatically, and it is the safer place to start.
  • The reply has to be drafted from the live record, not from a template with the record pasted in.
  • Complaints, claims and anything commercial should route to a person by rule, whatever the agent could have answered.

Open a shared operations mailbox on a Tuesday and the pattern is the same in every B2B business. Where is my order. Do you have 400 of these in stock. Can you resend the delivery note for last month. We need to change the delivery address on PO 44192. Why does this invoice not match what we agreed.

Almost none of these require thought. They require access. The answer exists in the order record, the stock table, the document store or the invoice, and the only reason the customer waits is that the person who read the email cannot see all four.

The relay, drawn out

It is worth writing down what actually happens, because the cost is invisible when each step looks reasonable.

  1. 01The request arrives

    In a shared mailbox, addressed to nobody in particular, at 16:40.

  2. 02Someone triages it

    Reads it, decides which team it belongs to, forwards it. This person adds no information and consumes the first hour.

  3. 03It queues

    In a second inbox, behind that team's own work. Overnight, typically.

  4. 04Someone looks it up

    Opens two systems, finds the answer, which takes under a minute.

  5. 05Someone replies

    Writes the answer back, often without the context of the original thread.

  6. 06The customer chases

    Because by the time the answer arrives the question has moved on, or the reply answered one of the two things they asked.

Step four is the work. Everything else is transport. And step six generates a second request that goes through the whole cycle again.

Start with routing, not with answering

There is a tempting order of operations that usually goes wrong: automate the answers first, because that is the visible win. The safer and often larger win is to automate the triage.

Classifying an incoming request correctly, identifying the account and the order it refers to, and putting it in front of the right team with the record already attached removes the first hour and the overnight queue without the agent committing to anything customer-facing. It is the lowest-risk step in the entire category, and in many operations it is where most of the delay lives.

One published figure exists on this site for exactly this: a European logistics company measured 40% faster ticket routing, where the agent identifies the intent of each incoming request and routes it into the right operational workflow. The customer name is confidential, and the number belongs to their operation rather than being a benchmark for anyone else's.

Answering from the record, not from a template

When the agent does draft a reply, the difference between a good and a bad implementation is where the content came from.

A bad implementation matches the question to a template and fills in a field. It produces confident replies that are wrong whenever the situation is not the one the template assumed: the order that was split across two shipments, the line that was substituted, the delivery that went to the site rather than the registered address.

A good implementation retrieves the actual state and describes it. Three of your four lines shipped on Thursday and are with the carrier under this reference. The fourth is on back order with a confirmed date of the eleventh. That is a different answer from a template, and it is the answer that prevents the follow-up email.

Every reply that does not fully answer the question generates a second request. Half-answers are not half the work, they are more than double.

What routes to a person, always

Some requests should never be answered automatically, regardless of how clearly the agent could answer them. These are rules, not confidence thresholds.

Request typeHandlingWhy
Order status, ETA, trackingAnswered from live dataA lookup with a factual answer
Stock and availabilityAnswered from live dataSame, with the caveat the system states
Document requestsRetrieved and sentThe document exists; finding it was the work
Change requestsPrepared, then approvedIt changes a committed transaction
Complaints and claimsRouted, with context assembledCommercial and relationship judgment
Anything about price or creditRoutedCommercial decision, never automated
An angry customerRouted, flaggedTone is a signal the agent should act on by escalating

The last row matters more than it looks. An agent that detects frustration and escalates rather than replying is better service than one that answers correctly and coldly. Escalation is a feature, not a failure.

What to measure

  1. 01First response time, measured from the customer's timestamp rather than from when the ticket was opened internally.
  2. 02Share of requests resolved on the first reply, which is the number that predicts follow-up volume.
  3. 03Share of requests that never reached a person, split between routed-only and answered.
  4. 04Reopen rate, which is the honest check on whether the automated answers were actually complete.

A rising reopen rate alongside a falling response time means the agent is answering fast and incompletely, which is worse than the original problem. Watching both together is the discipline.

The Customer Operations Agent sets out what is built before a deployment and what is configured against your systems, tone and escalation rules.

Questions

Questions we get about this.

Does it reply to customers directly?

Only where you decide it should. Classification and routing can run on their own, drafting can sit behind a person's approval, and automatic sending can be enabled for specific request types. Each is a separate decision, and most deployments start with routing.

How does it know the answer is current?

It reads the live record rather than filling a template. An order status answer describes what the order record, the shipment and the carrier reference actually say at that moment, including partial shipments and substitutions.

What happens with a complaint?

It routes to a person, with the account, the order history and the relevant documents already assembled. Complaints, claims and anything touching price or credit are routed by rule rather than by confidence, because they are commercial and relationship decisions.

Can it handle requests in several languages?

Reading and replying across languages is normal rather than a separate capability, which matters for shared mailboxes that receive mail from several markets. Your escalation rules and tone guidance apply the same way in each.

Where do the answers come from?

From the systems the request is about: the order record, stock, shipment and document store, read at the moment the reply is prepared. Nothing is answered from a cached copy, because a stale answer to a status question is worse than no answer.

See what this would look like in your operation.

A Workflow Review traces one of your real workflows end to end and names the first step an agent could take over. No platform to evaluate first.

Start with one workflow

Fixed scope for the first workflow, agreed before we start. Expansion is your decision once it proves value.

How it works

Security, stated plainly

Where the agents run, what they can and cannot do, and exactly what we do and do not claim.

Security