Article
AI automation when the CRM is still messy
Most teams asking for AI automation already know their CRM is a bit untidy. They just hope the automation will cope. Here is what breaks first when it does not, what to fix before you build workflows, and how we scope that without starting a whole new project.
The request usually sounds like this: "We want leads routed automatically, follow-ups drafted, and a weekly summary we do not have to build by hand."
All reasonable. Then, a few minutes later: "The CRM is a bit of a mess, but that is fine, right?"
Sometimes it is. Often it is not.
People have been working around a messy CRM for years. They know which fields to ignore and which ones only get updated at the end of the month. Automation has none of that context. It reads what is in the field and acts on it, every time, at speed.
We have written about fixing the data before you add the agent and about where agents read bad CRM fields. This post is the automation side of the same problem: what goes wrong when the workflow is live and the fields underneath are still drifting.
What breaks first
The failures are rarely dramatic. They are small, confident, and repeated.
Leads go to the wrong person
Routing rules depend on fields like "Service type", "Region", or "Source". If those are typed by hand, or filled in differently by different people, the rule matches some records and misses others. Leads land with the wrong person, or with nobody, and it takes a while before anyone notices the pattern.
Emails that read well and say the wrong thing
An automated follow-up pulls a name, a stage, and the last thing discussed. If the stage is stale or the note is out of date, the email is polite, fluent, and wrong. It thanks someone for a quote they already declined, or chases an invoice that was paid in cash. The customer sees it before your team does.
Deals move without anyone deciding
Workflows that advance a stage when a field changes are only as good as whatever changes that field. A form integration or an old sync step updates a value, the workflow fires, and a deal jumps forward on a trigger nobody actually made. Now the pipeline says something your team does not believe.
Reports that look precise
An automated weekly summary has tidy totals and clean charts. If "Value" is still the first-quote guess and duplicates are counted twice, the report is precise and wrong at the same time. Because it arrives on schedule and looks finished, people start trusting it more than they trusted the spreadsheet it replaced.
None of these are AI problems as such. They are data problems that automation makes faster and harder to see.
What to fix before the workflows
You do not need a perfect CRM before you automate anything. You need the fields that feed this workflow to be trustworthy.
The order is the same one we use in our Data Intelligence work, and where agents read bad CRM fields walks through it in detail. In short:
- Start from the workflow, not the whole CRM. List the fields the automation would read or write. Usually that is a handful, not hundreds.
- Give each one an owner and a winner. One person responsible, one system that wins when two disagree.
- Turn free text into fixed options where a rule depends on it. Routing on a dropdown is predictable. Routing on whatever someone typed is not.
- Stop other tools quietly overwriting people. Find the syncs and old automations that write to the same fields, and decide which ones stay.
- Keep a person on anything customer-facing or money-related. The automation drafts and suggests. Someone still approves.
If the CRM itself is the bottleneck, wrong structure, wrong objects, or a system the team has outgrown, that is a CRM design and development job first. The automation goes on after.
How Discovery scopes this
This is not a separate audit or a new package. It is part of how we already run Discovery for AI automation engineering.
In Discovery we:
- Pick the workflow you actually want. Lead routing, follow-ups, reporting, intake, whatever is costing your team the most time.
- Look at the fields that would feed it. Where they come from, who updates them, what writes over them, and whether the team trusts them.
- Sort what is ready from what is not. Some workflows can be built straight away on fields that are already clean. Others need a small amount of tidy-up first.
- Decide the order honestly. If the data underneath a workflow is not ready, we say so and suggest fixing that first, or automating something else in the meantime. Sometimes the right call is "automate this later".
You come out with a clear view of which automations are safe to build now, which fields need work before the rest, and what that work involves. No invented urgency, and no automation built on fields your own team does not believe.
If you are a Melbourne business weighing this up, our page on AI automation for Melbourne businesses covers the kinds of workflows we usually start with.
Next step
If you want automation but you are not sure your CRM is ready for it, that is a good place to start the conversation.
We look at the fields the workflow would depend on, tell you what is ready, and only then talk about building.
Amir
Alphesda Interactive