Skip to content
Skip to content

Ship Automation Without the Drama

The difference between an automation that runs quietly for years and one that gets switched off after a bad month is almost entirely in the rollout. Five phases, in this order, with a shadow-run before anything goes unattended.

Work through it in order. The shadow-run phase is the one most projects skip and the one that catches the most expensive problems.

Five-Phase Rollout Checklist

Map, build, shadow-run, cut over, monitor. Do not skip phase three.

0 of 31 complete0%

Your ticks are saved in this browser, so you can work through the list over several sessions.

01Phase 1 · Map

0/6

Document reality before encoding it.

02Phase 2 · Build

0/7

Design the controls in, do not retrofit them.

03Phase 3 · Shadow-run

0/6

The phase everyone skips and everyone should not.

04Phase 4 · Cut over

0/6

A slice, not the whole thing.

05Phase 5 · Monitor and iterate

0/6

Invisible things stop getting reviewed. Schedule it.

A general implementation framework. Specific steps vary by platform and by the risk profile of the process being automated.

The Three Rules

Everything in this checklist follows from three principles. If you remember nothing else, these are the ones that prevent the failures that actually hurt.

Shadow-run before you trust it

Run the automation alongside the manual process for a week or two, comparing outputs without acting on the automated ones. It costs a fortnight and it is the single most effective way to find the cases nobody thought of, before those cases reach a customer.

If you cannot see it, you cannot run it

Monitoring is not optional infrastructure to add later. An automation without alerting will eventually fail silently, and silent failure discovered six weeks on is far more expensive than the automation ever saved.

The team has to know what changed

Automations get worked around by people who do not understand them. Explain what it does, what it does not do, and who to tell when it looks wrong, before go-live, not after the first confused escalation.

The Five Phases

Sequence matters more than pace. Most single-process automations move through all five in four to six weeks.

1

Map

Document the real process end to end, including exceptions, and agree what success looks like in measurable terms before any building starts.

2

Build

Construct the workflow with proper service accounts, least-privilege access, error handling and logging designed in rather than retrofitted.

3

Shadow-run

Run automated and manual in parallel, compare every output, and fix the differences. Nothing acts on the automated result yet.

4

Cut over

Switch on for a defined slice of real work, with the manual path documented and ready to resume.

5

Monitor and iterate

Review exception queues, measure against the success criteria agreed in phase one, and widen scope only once the numbers hold.

The Controls Projects Forget

Four things that are trivial to include during the build and genuinely painful to retrofit once an automation is running in production.

A log of every decision made

When someone asks in three months why a particular invoice was coded to a particular account, you need to be able to answer. Logging every automated decision with its inputs and its reasoning turns an opaque process into an auditable one, and it costs almost nothing to include from the start.

  • Log the input, the decision and the timestamp for every run
  • Keep logs long enough to satisfy your own record-keeping obligations
  • Make logs searchable by the identifier a human would actually use
  • Ensure a non-technical person can read the log without help

A defined stop condition

Decide in advance what would make you switch the automation off, and who has the authority to do it without a meeting. Automations that keep running through an obvious problem do so because nobody was sure they were allowed to stop them.

  • Define the error rate or anomaly that triggers an automatic pause
  • Name who can switch it off, and make sure they know they can
  • Document how to revert to the manual process, step by step
  • Test the stop and revert procedure before go-live, not during an incident

An owner for the exception queue

Every automation produces items it could not handle. Without a named person reviewing that queue on a schedule, it silently accumulates until it becomes its own backlog, at which point the automation has moved work rather than removed it.

  • Name the person who reviews exceptions and how often they do it
  • Set a threshold on queue size that triggers escalation
  • Feed recurring exceptions back into the workflow as new rules
  • Track queue size over time. A growing queue is an early warning

Documentation that survives the builder leaving

Automations frequently become undocumented single points of failure. A short written record of what it does, what it connects to, what credentials it uses and how to change it protects you from the day the person who built it is unavailable.

  • Record what the workflow does and which systems it touches
  • Record which service accounts and credentials it depends on
  • Record how to make a common change, with a worked example
  • Store it where operations can find it, not in a developer’s folder

Next Steps

Automation Buyer’s Guide

Still choosing a partner? Twenty-eight questions worth asking first.

Open the guide

Process Automation Priority Scorecard

Make sure you picked the right process before you build the wrong one well.

Score the process

Automation Readiness Assessment

Check the business fundamentals that determine whether a rollout goes smoothly.

Assess readiness

Frequently Asked Questions

How long should the shadow-run phase last?

Long enough to cover one full cycle of the process plus its predictable variations. For a daily process, one to two weeks is usually sufficient. For anything with a monthly rhythm (month-end close, payroll, reporting) you need at least one complete month, because the exceptions that matter only appear at period end. Businesses that shorten this phase to save time almost always spend that time later dealing with a production problem, usually at a worse moment and in front of a customer.

Should the first cutover be the whole process or part of it?

Part of it, essentially always. Route one category, one customer segment, one branch or one document type through the automation first while everything else continues manually. You get real production signal with a contained blast radius, and reverting is trivial if something is wrong. Full cutover on day one converts every minor configuration gap into an operational incident, and it is the most common reason a technically sound automation gets a poor internal reputation it never recovers from.

What should I measure to know whether it is working?

Agree the measures in phase one, before anyone has an emotional stake in the answer. The useful set is usually four numbers: the proportion of items handled without human intervention, the error rate compared with the manual baseline, the time from input to completion, and the size of the exception queue over time. The manual baseline matters most and is the one most often missing, without knowing your existing error rate, you cannot tell whether the automation improved things or simply made existing errors more visible.

Who should own an automation once it is live?

The business function that owns the process, not IT and not the vendor. The person accountable for the outcome should be accountable for the automation that produces it, with technical support available to them. Automations owned solely by a technical team drift out of alignment with how the business actually operates, because the people who notice the drift have no direct route to fix it. The practical arrangement is a business owner for the outcome and the exception queue, with a named technical contact for changes.

How often should an automation be reviewed after go-live?

Weekly for the first month, monthly for the first six months, then quarterly. The early reviews catch configuration gaps and edge cases. The later ones catch drift, the process changing without the automation being updated, which is the most common cause of a workflow that quietly stops matching reality. Put the quarterly review in a calendar with a named owner, because an automation that runs reliably becomes invisible, and invisible things do not get reviewed unless someone is scheduled to do it.

What if the automation performs worse than the manual process?

That is exactly what the shadow-run is designed to reveal, and finding it there is a success rather than a failure. If it emerges after cutover, revert to the manual path immediately using the documented procedure, then diagnose. Most commonly the cause is an input pattern that was not represented during testing, which is a fixable configuration problem. Occasionally it reveals that the process genuinely depends on judgement that was never articulated, in which case the honest answer is to narrow the automation to the rules-based portion and leave the judgement with a person.

Want This Reviewed Before You Go Live?

Send us your workflow design and rollout plan. We will tell you what we would change, whether or not you are a customer.