Automating repetitive tasks works, as long as you pay the right price. That price isn't building it. It's watching it. At Pillet Grenié Bureau, we automated three of our own tasks: meeting notes, technical SEO and recurring emails. All three deliver. And each one taught us how an automation can break without making a sound.
Automatic meeting notes: from transcript to assigned tasks, with no retyping
Every meeting we hold is recorded and transcribed. The transcript lands in our system on its own, through an n8n workflow. An AI model then classifies the meeting: client or internal, then discovery, scoping or ongoing project. It pulls out a summary, the decisions made and the actions to take, each with an owner. Finally, it writes the notes in our usual format.
An unknown participant? A prospect record is created. An existing client? Their record is updated. Every action becomes an assigned task. Nobody copies out their notes anymore. That's not a small thing: according to a ProcessMaker study cited by Ideematic, office workers spent 10% of their time on manual data entry in 2023.
Two rules made the difference. First: when in doubt, the model picks the earliest stage. Treating a client like a prospect in discovery costs nothing. Sending them a proposal too early can cost you the relationship. Second: nothing becomes a sales proposal without a human signing off. That click takes a few seconds. A wrong proposal sent automatically can lose you a client.
SEO automation: what a machine does very well, and what it will never do for you
We ran our own site through a tool-assisted technical audit. In one day, it found what we had missed for months. Our service pages reused the homepage title, so Google saw them as duplicates. They were also missing from the sitemap. That sitemap had been frozen at launch, so every article published afterwards stayed invisible. And our pages were rendered far from our visitors. From France, an article took 1.2 to 1.6 seconds before it even started to display.
All of that can be fixed and measured. After the fixes, that delay dropped to between 0.2 and 0.5 seconds. The sitemap now regenerates every hour, with no one touching it. That's what a machine does very well: detect, check, repeat.
What a machine will never do is make a search engine choose you and cite you. The same audit proved it. Our name was spelled four different ways. We showed up in none of the ten searches a prospect might type. No directory, no review and no press article mentioned us. That work is about identity, reputation and content. No script does it for you.
Automated emails: why "sent" doesn't mean "received"
On a client's website contact form, the sending tool showed "delivered" for every test. Yet nothing arrived, in any of the five inboxes we tested. We started by looking for the fault in the wrong place.
What we learned from that case: "delivered" only meant the recipient's server had accepted the message. What happened next was out of the sending tool's hands. The spam filter can quarantine it without warning. A forwarding rule can point to a dead address. Here, the host's spam filter swallowed the messages right after accepting them. The dashboard stayed green, and prospect enquiries vanished.
The second trap: the email that goes out twice. Our monthly newsletter is triggered by a scheduled task. The first version checked whether that month's issue had already gone out, then sent it. An independent review caught the problem before the first send: the platform can launch the same task twice at the same moment. Both runs check, find nothing, and each sends the newsletter. Every subscriber gets it twice. Now each run creates a draft, then rereads the list of sends, and only the oldest draft goes out. When in doubt, nothing goes out: better a late newsletter than a duplicate one.
What an automation really costs: monitoring, not building
Building an automation takes a few days. Monitoring it lasts as long as it runs. And the failures that cost the most are silent.
The sneakiest is the scheduled task that never runs. We look for it in every system review. A typical case we check for: a cleanup job meant to delete data after 30 days, tied to another feature. The day that feature is switched off, the cleanup stops with it. No error surfaces and the tests stay green. The result: you keep data longer than your privacy policy promises. Our rule: every recurring task gets its own trigger.
Another leak: retries. If two AI calls are bundled together and the second fails, the retry runs the first one again too. You pay twice. Worse, the AI can return a different result the second time, and an assessment already on record changes without anyone noticing.
Hence our metric. The famous "10 hours saved a week" rests on no serious study, as this analysis points out. The same source cites a Workday survey: 85% of employees say AI saves them one to seven hours a week, but nearly 40% of that time goes back into fixing mistakes. So we track something else: the share of what we pay for that produces a usable result. The rule fits in two lines:
successful run, result used → useful spend
failed run → wasted spend
A failure recovered later is still wasted spend: you paid twice for one result. Under 10% wasted spend, all is well. Above 25%, fix the workflow, don't raise the budget.
How we automate a repetitive task for a client
The first decision is never about a tool. We look at the process as it actually exists. If it's fuzzy, we stabilise it before automating it. As Onilys puts it, automating a messy process only reproduces its flaws. Then we pick a single task: the one whose result is easiest to check.
The safeguards go in before launch. We add a human sign-off before anything that touches the client relationship. We test the whole journey end to end, on real inboxes, without trusting the provider's green light alone. And we set up an alert for when nothing happens but something should have. It's the only way to spot a task that has stopped running.
Take the contact form mentioned above, on a service company's brochure site. We reworked the sending address, then checked delivery inbox by inbox. The result: prospect enquiries arrive again, and the client now knows what to check if the spam filter changes its mind.
That's the heart of our integration and automation work: less spectacular building, more quiet monitoring. An automation is only worth something if you can trust it on a Tuesday morning without going to check.
