"Who's still waiting on a quote from last week?" That's a question an owner might ask a good office manager without a second thought. Asking the same question of a CRM has usually meant opening a report, picking a date range and filtering by status. Voice and text commands change that. You type or say the question in ordinary words, and the software answers, or proposes the change you asked for and waits for your yes.
It's also one of the easiest features in modern software to oversell. Here's how it works in practice, what to expect from it, and where it genuinely earns its keep.
How voice and text commands work in a CRM
There are three steps under the hood, and knowing them makes the rest of this article easier to follow.
First, if you spoke, speech recognition turns your audio into text. Second, an AI model reads that text and works out what you want: a search, an update, a report or something else, plus the details (which customer, which date, which job). Third, the CRM runs that request against your data, using the same rules and permissions as if you'd clicked through the screens yourself.
Text commands skip the first step. You type "jobs booked for Thursday in Geelong" into a command bar instead of building a filter. In my view typed commands are the underrated half of this feature. They're quiet, they're easy to correct, and at a desk they're often quicker than any menu.
The important design point is that the AI doesn't get a back door into your database. It translates your words into the same actions the normal screens perform, so everything it does can be checked, logged and limited.
Searching: the easy win
Finding things is where commands pay off fastest, because the risk is low. If the answer is wrong, nothing has changed; you just ask again or open the record. Examples that work well:
- "Show me open quotes over $5,000 that nobody has followed up in a week."
- "What's the gate code for the Hendersons' site?"
- "Which customers in Ballarat haven't had a service in over a year?"
- "When did we last speak to Priya at the dental clinic, and what about?"
These are hypothetical, but they're the shape of question an owner or coordinator asks every day. Each one would otherwise mean a filter, a sort and a scroll. Searching by meaning helps too. "The bloke with the leaking hot water system in Richmond" can find the right job even if nobody typed "leaking" into the notes.
Updating records without opening five screens
Updates are where commands save the most clicks, and where the design has to be most careful. Typical examples:
- "Mark the Smith job as complete and add a note that the client wants a follow-up in March."
- "Move Tuesday's 2pm inspection to Wednesday at 9 and let the customer know."
- "Add a new lead: Sam Nguyen, kitchen renovation quote, referred by the Taylors."
- "Set the Collins account to on hold until the overdue invoice is paid."
Notice that some of these do more than one thing. "Let the customer know" means sending a message, which is a different kind of action from changing a date in your own system. A well-built CRM treats those separately, which brings us to confirmation.
Reporting by asking questions
Reports are the third big use. Instead of exporting to a spreadsheet, you ask:
"How many enquiries came in last month, how many did we quote, and how many did we win? Split it by source."
The answer should come back as numbers you can click into, not just a sentence. That's my one firm rule for AI reporting: every figure it gives you should link to the records behind it, so you can see that "14 jobs won" really means fourteen jobs and not thirteen plus a duplicate. A chatty summary with no way to check it is a liability dressed up as convenience.
Commands work best for one-off questions. Numbers you watch every week still belong on a dashboard, where you can see the trend at a glance. Our page on CRM reporting dashboards built around your numbers explains how we set those up so both approaches draw on the same data.
How it confirms before changing anything
A command that reads data can just run. A command that changes data should show you what it's about to do and wait. In practice that looks like this:
- You say "Move the Patel install to next Friday."
- The CRM replies with the specifics it understood: "Move the Patel solar install (job 2418) from Tuesday 10 November to Friday 13 November, 8am. Send the customer an SMS? Confirm or cancel."
- You confirm, change a detail, or cancel.
The read-back matters because it shows the exact record and the exact values, not a paraphrase of your words. If there are two Patels, it should ask which one rather than pick. Riskier actions deserve a firmer gate: anything that emails or texts customers, changes money (prices, discounts, invoice amounts), deletes records or touches many records at once should spell out the full effect and need an explicit confirmation. Bulk changes are the obvious danger. "Mark all of last week's jobs as invoiced" should tell you it's about to change 23 records and list them.
Every change made by command should also land in the record's history with who asked and when, the same as a manual edit. If something goes wrong, you can see it and reverse it.
Permissions: it can only do what you can do
A command runs as the person who gave it. If your apprentice can't see pricing in the normal screens, asking the assistant "what margin did we make on the Jones job?" shouldn't reveal it either. This sounds obvious, but it's worth asking any vendor directly, because a system that gives the AI a single all-powerful account and relies on it to "behave" is a system waiting for an awkward conversation.
In the CRMs we build, staff see only what their role needs and sign in with multi-factor authentication, and voice and text commands sit inside those same rules. Voice adds one more wrinkle: anyone near the phone can hear the answer. Reading a customer's address aloud in a ute is fine. Reading out a clinic patient's details on a crowded train is not, and the OAIC's guidance on privacy and commercially available AI products makes the point that privacy obligations apply both to personal information put into an AI system and to what it produces. We cover the wider picture in our article on AI, privacy and customer data for Australian businesses.
Accuracy, and when to check
Speech recognition is very good now, but it isn't perfect, and its mistakes follow a pattern. Names, street names, product codes and numbers are where it slips. "Fifteen" and "fifty" sound alike in a windy driveway. Unusual surnames are hard because the system has rarely heard them.
Google's documentation for its speech service is candid about this. It says each result carries a confidence estimate that can be used to decide whether to ask the user for confirmation, and that rare words can get a low likelihood even when they're recognised correctly (see Google Cloud's Speech-to-Text request and confidence documentation). Microsoft's speech service offers phrase lists so a system can be primed with domain terms and proper nouns, according to the Microsoft Learn speech to text overview. In a CRM, the practical version is that your own customer names, suburbs and product names should be part of what the system listens for.
So when should a person check? A short rule of thumb:
- Searches: trust, then glance. If the result looks odd, open the record.
- Notes and simple updates: read the confirmation before saying yes.
- Anything with a number in it (dates, times, prices, quantities): read the number back every time.
- Anything a customer will see: check the wording, not just the action.
Where voice beats typing, and where it doesn't
Voice shines when your hands or eyes are busy. A plumber finishing a job with wet gloves can say "job done, replaced the tempering valve, customer wants a quote for a new unit" and have the note, the status and a follow-up task sorted before leaving the driveway. Sitting in the van between jobs, parked, is another good moment: three or four quick commands while the details are still fresh beat a pile of paperwork at 8pm.
Driving is a different story. Road rules on using phones behind the wheel vary from state to state, and the safe habit is simple: pull over, then talk to the CRM. We'd rather a note was written ten minutes late.
Voice is weaker in a few places, and I'd argue it's overrated at a desk. On a noisy site, near an angle grinder or a compressor, recognition suffers. In an open-plan office, everyone hears you. For long text, like a detailed quote description, typing and editing on screen is still faster and easier to get right. And for exploring data, where you want to compare, scroll and spot patterns, a screen beats a voice reading out a list. Typed commands sit in the middle: as quick as talking, but quiet and easy to correct.
If you're curious how this fits into a CRM designed around AI from the start, our explainer on what an AI-first CRM actually is is a good next read, and our custom CRM builds show where commands sit alongside the rest of the system. Or, if you'd like to try asking your own business questions out loud and see how the answers come back, get in touch with our team for a 30-minute call.
— Craig Winder, Founder
