Trust in Agents: a case study of Attio Workflows
The UX of handing over the work, not the control
Three months ago, I wrote that CRMs had started to think. The article ended on a four-stage timeline - from classic CRM, to AI features, to chat interface, to the AI eventually doing the work for you:
When talking about the future of AI, Notion mapped the same journey; thought partner, to assistant, to teammate, to running the whole system:
At the end of both, the product ends up doing the work, not just helping with it.
And, as of today - 24 June 2026 - it feels like most products have arrived đ So many tools have an agent now. Which sounds great, except thereâs a catchâŠ
The more you ask something to do, the more it can get wrong. If I hire an intern, do I hand them the company card on day one?
Nope.
So, the real question isnât whether a product can build an agent, itâs how they get customers to trust it enough to feel âOKâ with trying it out.
Enter: Attio Workflows.
Look closely and the trust is developed in four ways:
Discovery: You gotta drive people to the feature first
Defaults: design defaults that match the user
Proof: showing the working
Sign off: clarity on draft states, and when theyâre live
Starting with discovery - because if they donât get there, nothing else matters.
First, drive users to the feature
You canât trust a feature you never find đ«
Many new features die because theyâre hidden two layers deep in the navigation.
The first thing I notice is where Workflows is pushed for the first time: in the onboarding checklist.
âCreate a workflowâ sits at #4 of 6 in the getting-started checklist, above sequences and apps in the side navigation.
Positioned above Invite team members and the trial countdown in a neat module, the list itself comes with completion-bias đ§ i.e. the psychological urge to finish tasks once started.
This is the more prominent position for workflows, as within the normal side navigation of the dashboard, itâs nestled under âAutomationsâ.
As a result, I arrive as a new user, Iâm exposed to the feature, and I give it a spin.
Next up: design for the empty space
Empty space in a product has a similar effect as writers block: itâs hard to get going.
When a user is trying to do something new, they think:
Whatâs possible?
How do I use this?
I canât think of a use case?
How do I get going?
Whatâs a good first thing to try?
Thatâs why the product needs to design for indecision, newbies and writers block.
When I get to the workflows page thereâs a chat box with:
Describe your workflow
A beginner-level prompt - nice. But still hard for some. SO, thereâs four example actions:
Book discovery
Lead triage
Deal nudge
Health check
The interaction design is nice.
Once you hover on a prompt, it fills in the box. Once you click, the grey text turns white.
Hovering over all the suggested prompts I start to see what a good prompt looks like.
Whatâs nice about these is:
It teaches by example, not instruction: Instead of a tooltip explaining, the placeholder is a perfectly-formed prompt: trigger (every Monday), condition (no activity in 14 days), action (create a task with next steps). I learn as I go.
The prompt is written in customer language, not system language. âFind open deals with no activity⊠create a task for the owner to re-engage.â Thatâs workflows described in the reps words.
Itâs low-stakes enough to try an agentic action: only two of these actions are customer facing, and theyâre all painful, manual parts of the deal process. An easy and tempting first step to take.
When thereâs a new behaviour to learn, showing beats telling and doing beats both.
So, I pick âdeal nudgeâ and watch it get to work.
Show the working
This is the moment the agent could feel like a black box. Iâve handed something over and Iâm waiting, hoping it understood me.
Instead, the chat shows me everything itâs doing. Under a âThinkingâ dropdown, each step appears as it happens:
I need the Deals object attributes. Let me look that up.
SQL query it runs
The number of results it gets back (e.g. 1 row in 115ms)
Whether it has enough information to carry on
It even shows the failures. One query comes back âFailed to execute SQL queryâ. The agent adjusts, runs a cleaner one, and gets a result. Many products would hide a fail but Attio leaves it in to show that it can learn and get better, makes it feel real.
Whilst something thatâs perfect and flawless is tempting, it can look staged. A lilâ stumble looks like itâs actually working it out live.
This is the labour illusion đ§ when you can see the work being done, the system feels more trustworthy, even if the wait is the same.
Trust in AI is not about perfect, itâs about seeing the work and signing it off, or course correcting if needed. I trust it because I watched it get something wrong, then fix it.
By the time itâs finished, I can click a drop down to see all the steps. Not that I get them all, but it feels cool.
On to the final step.
Lastly, allow sign off
As the chat gets to work in the right hand pane (thinking, showing me the working, letting me in) on the right the page starts to⊠build.
By the time the chat is finished, the agent has built the full workflow on the canvas:
A recurring schedule (fire every Monday morning)
Find records (open deals with no activity)
A dotted loop over each stale deal
A custom agent to write the re-engagement note
Create a task, assigned to the deal owner
Across the top of the editor thereâs a banner: This workflow has unpublished changes, with a single Publish changes button. Thereâs a Draft / Live toggle in the corner.
Across the dark UI this blue banner with the bright âPublish changesâ button is the most striking thing on the page.
Itâs calling out to me âyou have control to sign this off hereâ.
The key thing this UI is telling me is that Iâm handing over the effort, not the control.
The layout helps too. Itâs the workflow builder I already know: trigger â step â step â done, left to right on a canvas, boxes and arrows. The agentic part is new, but the format is one Iâve seen for years in CRM tools.
The UI is not necessarily needed - but with these complex workflows, UI is trusted more over natural language. The UI of the workflow is the thing that creates trust by showing me the steps.
Itâs the same move Attio made last time; a CRM that looks like a CRM. They didnât ask me to learn a new way of working, they let an agent work inside ways Iâm used to.
In conclusion:
Funnily enough I was going to call this when your CRM starts to act, because the last Attio growth dive I did was called when your CRM starts to think.
But âactâ was the wrong word.
The CRM doesnât act on its own. It drafts, it shows its working, waits. Then it builds the thing in front of me, stops one button short and asks: good?
And the reason I am happy to answer âgoodâ comes down to the four moves we walked through:
Discovery: the feature surfaced in the onboarding checklist got me there in the first place.
Defaults got me started. Four example pills turned a blank box into one click, so I never had to face the empty space.
Proof got me to trust it. The âThinkingâ dropdown showed the working (a failed query, a fix) so I watched it get something wrong and course-correct.
Sign-off gave me control. It built the whole workflow on a canvas I already understood, then handed me one bright button: Publish.
In the days where itâs easier to build than ever, anyone can build. The hard part for us in product is now around trust.
In partnership with Attio. Thanks for the early look at Workflows đ€đ»