AI or automation: picking the right tool for the job
An eighteen-month job finally wraps up, the client is happy, and somebody on the marketing side now has to go and ask them how it went. That means digging out the feedback questions, which live in a Word document on somebody's own laptop in a version nobody is entirely sure about, and then working out who to call, because the person who signed the contract eighteen months ago may well have moved on. If the answers do come back, they get typed into another document on that same laptop, where nobody else will ever go looking for them, and the write-up and the testimonial request become jobs for a quieter week that never quite arrives.
The business I am describing had worked this out for itself years ago. Somebody put a feedback process together and it never got off the ground, because it was an extra job sitting on top of everybody's real one and it took time nobody had. In all the years since, they did not complete a single feedback form.
So does that problem need AI?
The problem was never the asking
Every step in that sequence is easy on its own — find a document, look up a contact, send an email, save a file — and any one of them takes two minutes. Strung together at the end of a long project, when the team has already moved onto the next job and the person who has to do it has eleven other things on, the whole thing quietly goes into the too-hard basket. Not because any part of it is difficult, but because it is a dozen small frictions in a row and no single one of them is actually anybody's job.
When an owner hears "we should automate this", they tend to picture something that thinks — a system that works out what to do and goes and does it, deciding things on their behalf. Fair enough, and they are right to be careful about handing judgement to a machine.
But look again at what that task is. Somebody reads a name in one place and types it somewhere else, attaches a form, and files a result. Nobody is deciding anything; they are being a pipe, and a pipe is not an AI problem.
What it looks like now
The marketing person opens the project in a SharePoint list and picks "run feedback". A message arrives in Teams asking one question: email or phone?
If they choose email, a draft appears in their Outlook drafts — addressed, written, with a link to the feedback form already carrying the project name in it. They read it and press send. If they choose phone, the Teams message comes back with the contact's name and number and a link to the same pre-filled form, so they can ring and fill it in as they talk. When the form is submitted the answers land in a list automatically, go straight to the project manager and the boss, and marketing gets told it happened.
There is no AI anywhere in that. It is a planned route — when this happens, do that, then that, then stop — and the only decision in the whole thing is made by a human, on the one question that genuinely needs a person: is this a client you ring, or one you email? Notice too that the route never sends anything. It writes the email and stops, which means it cannot embarrass anyone in front of a client, and the step that could do real damage stays with someone who can see the context. Every system I build works that way, producing a draft rather than a decision.
I built ten or twelve of these for that business, and most are duller than the one above — adding an ID number, moving something from one list to another. One of the more useful does nothing clever at all: rather than social media posts arriving for approval in dribs and drabs all week, they are collected into a single review in Teams, sent on the day and time the approver prefers. That was not a day I worked, which is rather the point — the work no longer had to wait for me.
Why the boring version is the one that works
This is the part I would argue with people about. The reason to prefer a plain automation over something cleverer is not that it is cheaper, and it is not really that it is safer either. It is that you can tell whether it worked.
A fixed route runs the same way every time, so it either ran or it did not and you can go and look. Something that decides afresh on each pass gives you a different answer each time, so you have to check it each time, and checking is the job you were trying to stop doing — you have not removed the work, only changed what kind it is.
Four projects, and why I am telling you the number
Since the feedback automation went in, four projects have reached completion and all four have a completed feedback form behind them. In all the years before it, they completed none.
Four is a small number and I am not going to dress it up — projects there can run to eighteen months, so four completions is simply what the period produced, and I would not draw a trend from it. But the comparison that matters is not four against some larger number, it is four against zero. The old process did not run slowly; it did not run.
Check what your Microsoft licence already does
All of that runs on SharePoint lists, Microsoft Forms, Teams, Outlook and Power Automate. Every connector involved is what Microsoft classes as a standard connector, and those are available to Microsoft 365 users under use rights you already hold — so they bought nothing to make it work. A genuine amount of useful internal plumbing sits inside that boundary and most businesses running on Microsoft have never looked.
The catch nobody selling this mentions is what happens when a route has to touch something outside Microsoft. Premium and custom connectors need a standalone licence, at AU$22.40 per user per month on an annual commitment at Microsoft's current Australian list price, excluding GST. That is not expensive, but it arrives exactly when you try to connect your accounting system — there is no Microsoft-published Xero connector at all, so that leg is third-party or custom before you start.
Where AI does earn its place
None of this says AI has no role here; it is an argument about where it goes.
Take the phone branch. Somebody rings the client and talks for twenty minutes, and the useful material is in what was said rather than in the boxes on the form. A model could transcribe that call and hand the project manager a summary — a job no amount of plumbing will ever do. The same applies to the one step I never replaced: somebody used to write up what the feedback said, and reading a year of free-text answers to find what clients keep raising is unstructured work with no right answer you can check by arithmetic.
Both of those are a single step inside a route that is otherwise entirely fixed, with a person reading the output before it goes anywhere. The route stays checkable, the model earns its place on the parts that were never mechanical, and the whole process still runs perfectly well on plain automation whether or not you ever add one.
Whose job is it once I have gone?
That contract has now finished, which makes this a more useful story than it was a month ago.
I handed the automations over to a named person before I left, which is the right way to do it and I would do it again. But a handover transfers responsibility, not motivation. The routes are still there and they still work; every friction I described at the top has gone, and asking a departing client for feedback is now one click and a read-through. Whether anybody runs them from here is up to that business, and I do not yet know the answer.
That is not me being gloomy about a client I liked working with — it is the honest limit of what this kind of work does, and I would rather say so than sell you the version where the software fixes everything. An automation removes effort, which was genuinely what stood in the way here; what it cannot remove is ownership. One click is still a click, and somebody has to decide to make it, remember it exists, and care whether the answers come back.
So the question worth asking before you build anything is not whether a job can be automated, because the answer is usually yes and it is not very interesting. It is whose job the thing becomes once it exists, whether that person knows it is theirs, and — the one most people skip — whether anyone would notice if it quietly stopped running. Answer those three and it will probably still be going in a year; leave them, and what you have bought is a very efficient way of not doing something.
An AI Readiness Day is a day in your business working out which of these you actually have — where AI fits, where it should stay out, who signs off on what, and a costed roadmap of what is worth building. Often several of those turn out to be plain automations rather than AI builds, which is a cheaper answer than the one you were expecting. $3,500 + GST, and the fee comes off a Workflow Sprint if you book one within 30 days.
Sources, checked 12 October 2026: Microsoft Learn — standard connectors reference (SharePoint, Microsoft Forms, Microsoft Teams and Office 365 Outlook all listed as standard) and Power Platform licensing FAQ (standard connectors available to Microsoft 365 users; a standalone plan required for premium, on-premises and custom connectors); Microsoft's Australian Power Automate pricing page for the Premium per-user price. Client details anonymised.
