From helpful answers to finished work

Work Musings · General AI

There’s a frustrating version of AI productivity: you use it more, get more summaries, and somehow end up with more work on your plate.

Now you have your emails, your Slack messages, and two pages of AI-generated reminders about everything you need to do. Very organized. Still your problem.

In the early days of our go-to-market AI team at Canva, a lot of our work came down to one question: how do we get reps the information they need, when they need it?

Meeting prep was a big focus. So were post-call follow-ups. Get the context together, make it useful, and put it in front of someone at the right moment.

As more of that preparation becomes possible inside tools like Claude and Codex, I see less value in our team centrally building every one of those workflows.

At Canva, we’ve already moved toward AI execution. Our focus is on actual business impact and fully automating work, including delivering sales collateral and building an agentic pipeline.

I’ve asked the same thing of my own AI workflows: how much of the work can this actually finish?

At the beginning of this year, a lot of the value in my own AI workflows came from gathering information and preparing something useful. That might be where you are, too: better meeting notes, a helpful brief, a clearer picture of what needs doing.

The next step is extending that workflow through the work itself.

Gather: Bring together meeting notes, account history, emails, Slack conversations, and previous commitments. This still matters enormously. Missing the conversation where the plan changed can send the whole workflow in the wrong direction.

Prepare: Produce something concrete: the follow-up, the relevant setup instructions, or proposed updates to your task list.

Review: Show the actual work, where it will go, and anything unresolved. Number the actions so you can approve, reject, or edit them together at a useful moment.

Execute: Carry out the authorized work through the connected tools and confirm what happened. That could mean saving a Gmail draft or updating Jira. The boundaries you’ve set still apply, including anything you always want to do yourself.

Improve: Regularly review corrections and places you had to step in. Use them to improve the instructions and checks for the next run.

The review point depends on the action. An already-authorized draft save can happen before you review the wording. A shared-record change might need your approval first. The sequence should reflect those boundaries.

Imagine finishing a customer call and finding a researched follow-up waiting in your drafts. This is an illustrative workflow, not exactly how we do things at Canva.

It needs access to the call notes, relevant documentation and calendar, plus a Gmail connection that can create and update drafts. Check those capabilities in your setup before adding the instructions. A skill describes the process; it doesn't grant access or start an automation.

Say you need to follow up about setup and a technical session. Before writing anything, the AI should use Glean or your internal documentation to understand the process. Which setup instructions apply? What needs to happen before a technical session? A generic answer can miss how your company actually works.

It should also check your calendar. If you already have a follow-up scheduled with that person, the email should reference it instead of asking them to book another meeting. If it can’t confidently match the person or event, it should flag that for you.

For this workflow, saving and updating the Gmail draft is already allowed. I’m reviewing the work it has prepared, rather than approving every save. The boundary is still firm: always save the follow-up to my Gmail drafts. Never send anything directly. Any Slack message stays proposed text for me to review and send myself, too.

The review could look like this:

You can say, “Shorten 1 and save the update. Skip 2, we’ve already arranged that.” The AI makes the changes in the right place. Approving a draft doesn’t override the rule against sending.

The finish line is the updated draft in Gmail, with any unresolved question visible. The workflow should confirm that the save succeeded, rather than stop after displaying the email in chat.

Here’s a real example from my own work: Jira.

I hate being in Jira. Any moment I’m spending inside the Jira interface is a moment too long for me. Its MCP, though? I really like it.

I use workflows that save meeting action items into Jira, help organize my day, and update my to-do list. Jira is where the work gets recorded. I don’t need to spend my day in its interface to keep that record useful.

The morning workflow goes further than telling me what’s on the list. It reviews my Jira to-dos, then AI attempts to help complete those action items as far as it can.

I've also had a lot of success using the Salesforce and Snowflake MCPs for quick data analysis and building reporting I need. Depending on your setup and connectors, the AI can do more than update the task that says “build the report.” It can help do the analysis behind it.

For a reporting task, I'd want the result, the query or source records behind it, the date range, and any assumptions that could change the answer.

Some tasks still need my judgment, someone else's response, or access the workflow doesn't have. It should bring back what it completed, what it prepared, and what is blocked. A task only gets marked complete when the actual result supports that.

Pick the finish line before building the skill: the draft saved, the right Jira issue updated, or the report delivered with supporting evidence. Then define what can happen automatically and where you need to decide. Here's how I'd set up the customer-call example.

1. Give the skill your actual process.

For the follow-through example, I’d start with instructions like these:

Add an example of a follow-up you like so the AI has something more useful than “write in my voice.”

2. Start it when the work becomes useful.

Try the skill on a few calls, correct the output, then automate the repeatable part. That could mean running when notes become available, if your platform supports it, or preparing a batch before your usual review time. For a Jira workflow, the start of your day is a natural moment to review the queue and begin working through it.

Reflect the draft-only boundary in the tool permissions or execution controls wherever supported. Written instructions alone don't enforce it.

3. Design how it brings you back in.

Choose one review surface in your AI tool. Keep action numbers stable and show the actual work, its destination, and anything you need to decide. Bring routine items together at a useful time instead of interrupting for every small step.

Your feedback should result in the work being updated. “Shorten 1” should save the shorter Gmail draft. For Jira, an authorized change should appear on the relevant issue. Ask it to confirm the result and flag anything it couldn’t complete.

4. Improve the workflow once a week.

Keep the corrections and failures in a small feedback log. Once a week, ask AI what keeps coming up and what it would change in the skill. Review those suggestions before adopting them.

“Use the customer’s product version when choosing the setup guide” is a useful correction. So is “check the calendar before asking for another meeting.” Try the revised instructions on a previous example to see whether they address the problem.

Count the time you spend correcting results and maintaining the workflow, too. Keep the weekly check short enough to earn its place: fix a recurring problem, test the change, and leave working instructions alone.

That’s my north star for personal productivity now: how can I actually have AI completely take this off my plate?

I won't get all the way there with every workflow. But the question makes me decide what information is needed, what a finished result looks like, and which decisions I want to keep.

Take one recurring task from your queue and extend the workflow through the last useful step your access allows. On the next few runs, compare the remaining review and repair with doing the task yourself. Include the setup and maintenance effort when deciding whether to keep it.

Robert