Michael Wilson.
Available for product design roles Dallas — ← All work

Independent audit · Designing for AI · Aside

Aside’s agent did the job. The friction sits around it.

An unsolicited audit of an AI browser’s first run, first task, and upgrade path. It includes 22 findings with evidence and fixes, a measured setup timeline, and five concept screens.

Rolesolo, self-initiated ProductAside (AI browser) PlatformWindows TestedOct 6–7, 2026

Unsolicited audit, October 2026. I have no relationship with Aside. The company didn’t commission or endorse this audit, and every concept screen is my proposal, not Aside’s current UI.

Concept screen, not Aside's current UI: a finished agent task. Under the prompt, a line reads Worked for 12m 16s beside 145 credits, 29% of your month, followed by four timed milestones, a link to the new Google Sheet, a Figma, Framer and Webflow pricing table, and a card showing 71% of the month's credits left with a Compare plans link.
In this concept for quick win 2, the finished task shows what it cost, in credits and as a share of the month
12 min
One sentence to a cited, four-tab Google Sheet
9 asks
Before the agent does anything, six of them about secrets or money
29%
One task’s share of the free month, never shown with the task

Summary

Aside is an AI browser whose agent carries out tasks across websites. I screen-recorded setup, narrating as I went, then gave the agent one real job in one sentence, to compare Figma, Framer and Webflow pricing in a Google Sheet. Twelve minutes later I had a cited, four-tab Sheet in my own Google account.

The friction sits around the agent, not in it.

Setup makes nine asks before any work. Mid-task, I had to sign in to Google and report back. The run used 29% of the month’s free credits, and the result never said so. And where friction belongs, a default takes it away. When I asked the agent to email my own address, it sent the email without asking first.

Quick win 1

Make the Chrome import recoverable

The error appears only after Import, with no reason and no Skip, and cost my run 69 seconds. Check first, name the cause, and wait instead of failing.

Quick win 2

Show what a task costs

One task used 29% of the free month, and the result never said so. Show the cost on the result, not a live meter, and price Pro in tasks.

Strategic

Ask for trust when it pays off

All nine asks come before the first task, and six concern secrets or money. Start with one real task, then let the work trigger each ask.

Read the full audit (PDF) 24 pages · every finding with its evidence and fix, except one security finding shared only with Aside

My role & method

One reviewer,
five recordings.

This was self-initiated, with no brief and no access beyond a free account. Over three Windows sessions on October 6–7, 2026, I narrated five screen recordings. Each finding lists its evidence, and most cite a timestamp in one of those recordings.

  • Recordings: setup (6:12), one agent task end to end (23:17), a usage check (2:52), a Help Center lookup (0:52), and a permissions test (5:44)
  • Scope: the first run, one core flow, and one conversion surface
  • Severity: High risks the first session or user trust, Medium costs time or clarity, and Low is polish
  • Deliverables: a 24-page report, 22 findings, five annotated concept screens

Timings come from one run, not averages, and the narration quoted here is my own. A solo audit can show where friction sits, not how often it happens.

22 findings by area5 High, 12 Medium, 5 Low. Three of the five High findings sit in setup, before the first task.

Setup, timed

Where setup’s four minutes went.

I timed every step from the import screen to the browser, 4:11 in all. Two steps, the Chrome-close loop and vault setup, took 2:20 of it, or 56%. Every other step took 25 seconds or less.

Seconds per step, in order. One recorded run on Windows. Install and account sign-in happened before the recording.
  1. Pick browser and profile0:00–0:24
  2. Chrome-close loop0:24–1:33
  3. Import runs, “No more integrations”1:33–1:49
  4. Sign-in choice + extension pop-ups1:49–2:12
  5. Vault: password, key, import2:12–3:23
  6. AI subscription + plan3:23–3:46
  7. “You’re all set” checklist3:46–4:11

Both setup recommendations aim here. Quick win 1 removes the loop, and the strategic reorder moves the vault to the moment it pays off.

Recommendations

Two quick wins and one strategic bet.

QUICK WIN 01

Make the Chrome import recoverable

The import fails only after you commit to it, and the error gives no reason and no Skip. In my run, Chrome took about 50 seconds to exit after its windows closed, and nothing said so. The loop cost 69 seconds, at least three failed retries, and a trip to Task Manager.

“Oh, Chrome is still open? I thought I just closed it.”

My narration, setup recording 0:59

What to change

  1. Check before Import. Detect a running Chrome on the profile screen, before the click.
  2. Name the cause. “Chrome is still closing in the background,” and why Aside needs it closed.
  3. Wait instead of failing. Continue on its own when Chrome exits, so nobody has to retry.
  4. Offer to do it. A “Close Chrome for me” button that warns about unsaved work first.
  5. Always leave a way out. A “Skip for now” button, pointing to the import that already lives in Settings.
  6. Warn one step earlier. Add “Quit Chrome first” to the Windows install instructions.
Concept screen: the Import from Chrome step with a Chrome is still open notice. It explains that Aside needs Chrome fully closed to copy sign-ins and history, says the import starts on its own once Chrome exits, and offers Close Chrome for me and Skip for now buttons.
In this concept, the import waits instead of failing. It checks for Chrome first, says why, offers to close it, and carries on once Chrome exits.
The tradeoff

Closing someone’s browser can cost them unsaved work, and a “Leave site?” dialog held Chrome open in my run. So the warning sits right under the button, and Skip is always one click away. The expected effect is that the riskiest moment in setup becomes a short wait that resolves itself.

QUICK WIN 02

Show what a task costs

My one 12-minute task used about 145 of the month’s 500 free credits, or 29%. The finished task said how long it worked, not what it cost, and the Usage page named the task on each of dozens of rows without ever totaling them. At that rate, the free plan covers about three tasks like mine a month.

What to change

  1. Put the cost on the result. Under “Worked for 12m 16s,” a line like “Used about 145 credits · 29% of your month.” That’s the concept at the top of this page.
  2. Group usage by task. One row per task with its total, expandable to its steps.
  3. Make it the upgrade moment. When a task takes a large share of the month, offer Pro right there, priced in tasks rather than credits.
Concept screen: a Choose a plan dialog that counts each plan in tasks like the user's pricing comparison, which used 145 credits. Free, about 3 tasks a month; Pro at $20 a month, about 10; Max at $200 a month, about 140. Ultrabrowse gets a one-line explanation, and a footer offers to run Aside on an existing ChatGPT or Claude subscription.
In this concept, plans are priced in tasks. Opened from Compare plans on the finished task, the picker measures each plan against the task the user just ran.
The tradeoff

A cost that ticks up during the run works like a taxi meter, and people tend to use less while they watch it. Showing it once, on the result, keeps the information without that pressure.

STRATEGIC

Ask for trust when it pays off

Setup makes nine asks before the agent does anything, and six concern secrets or money, from an unlock password and a 12-word recovery key to a plan. Each ask lands better at the moment it buys something, like the vault at the first login wall and the plan at the first limit.

Today, in my run

Nine asks before the first task

  1. Import from Chrome
  2. Close Chrome: a 69-second loop
  3. Choose how Aside signs in
  4. Create an unlock password
  5. Save a 12-word recovery key
  6. Import from a password manager
  7. Choose an AI subscription
  8. Choose a plan
  9. Pin, default browser, usage sharing
  10. First task
  11. Mid-task: save the Google login

Proposed

One ask before the first task

  1. Import from Chrome, with Aside closing Chrome
  2. First task: a real task on your own sites
  3. Login wall → save it to the vault
  4. Login saved → back up the key
  5. Credits run low → choose a plan
  6. Pro feature → upgrade in place
  7. Later: pin and default browser

The case for the reorder

  1. The import already holds the best fuel. History, cookies and bookmarks arrive in the first two minutes, and Aside’s homepage says that history becomes memory.
  2. Start with one real task. Right after import, open Ask mode with a task on a site the user already visits. Suggest tasks that work on public pages, so the first one doesn’t need an inbox or other private data.
  3. Let the work trigger each ask. Aside already does this once, when the vault offered to save my Google login at the first sign-in. Extend it to plans, and offer Pro when a task takes a large share of the month.

Mid-run, I wondered aloud what this did that a chat assistant couldn’t. The Sheet in my own Google account answered that, which is the argument for putting a real task first.

Concept screen: a new tab in Ask mode. The composer invites a task on a site the user already uses, and below it, under the heading Start with a real task, three illustrative suggestions built from imported history and bookmarks, such as comparing three pricing pages opened yesterday in a Google Sheet. Each shows where the idea came from, an estimated credit cost, and a Run it button.
In this concept, the first tab starts a real task. Right after import, the new tab opens in Ask mode with jobs built from pages the user already visited, each priced before it runs. The suggested tasks are illustrative.
The tradeoff

Moving the asks later means fewer people see them. Someone who never hits a login wall never sets up the vault, and fewer people switch their default browser in week one. The bet is that an ask with a reason converts better per view.

To find out, I’d measure time to first task, first-session task completion, vault saves at the first login wall, and upgrades from the result card.

Also in the report

Less friction in setup, more where it counts.

Most of the audit takes friction out. One finding puts some back.

1  Confirm before the agent acts for you

Aside’s homepage describes confirmation for sensitive actions, but Final confirm is off by default. On the defaults, I asked the agent to email my own address. It wrote “I’m sending this now as requested” and sent it, with no prompt. With Final confirm on, the same request stopped at a card showing the full email, with Cancel and Confirm. From the outside, the fix looks like one default, turning Final confirm on for send, pay, delete and submit. The approval card already exists.

2  Keep the task in view through a sign-in

Despite the cookie import, Google Sheets was signed out. The agent asked me to sign in “in the open Sheets tab,” a tab the sidebar listed under a different name. Its question docked beside Google’s page, then vanished when the page moved on, and I answered 1:23 after it asked. The concept keeps the task docked through every step, names the tab the way the sidebar does, and resumes on its own once the session is live.

Concept screen: a Google sign-in page with the task docked in a side panel. The panel lists the run's steps with Sign in to Google, waiting on you, names the Sign in to Google Sheets tab, shows Watching for your sign-in, and offers I'm signed in and Use a spreadsheet file instead. The sidebar marks the task Needs you.
In this concept, sign-in keeps the task docked, with “I’m signed in” as a fallback and a way around Google

3  Name what’s working

The report closes with twelve patterns worth keeping as the flow changes. Among them, the agent declined instructions a pricing page addressed to AI readers, memory lives in plain files the user can open and edit, and a Sources tab makes the finished Sheet checkable. My verdict on the run itself was “a pretty good job for one shot.”

The deliverable

22 findings5 High · 12 Medium · 5 Low, each with evidence and a fix
5 conceptsEach annotated against today’s screen
5 recordings39 minutes of narrated screen recording

The full report runs 24 pages, covering the summary, the setup timeline, every finding with its evidence and fix, the five concepts annotated against today’s screens, and what’s working. One security finding is left out of the public copy, and I shared it only with Aside.

There’s no outcome metric on this page, on purpose. The audit was unsolicited, and I can’t speak to what Aside does with it. What I can show is the evidence, and the report cites all of it.

Read the full audit (PDF)

What I’d do differently

I’d put two or three first-time users through the same setup before writing a word. Every timing here comes from my own run, which makes the audit precise about where friction sits and silent about how often it happens. A few more sessions would show whether a 69-second Chrome loop is common or just my machine.