Past the demo
What an AI browser does when the data gets messy
Most people who try AI browsers never get past the demo. They watch a slick walkthrough, try to replicate it on their own work, and find the tool stumbles the moment the data gets specific. The problem is not the technology. It is that most reviews show you what the tool can do in ideal conditions, not what it looks like when you point it at HubSpot, a blank spreadsheet, and a genuine research brief.
This is Perplexity Comet — a local AI browser built on Chrome — run across three real business tasks, plus the risks worth knowing before you hand it access to anything sensitive.
Local by design
Why running on your machine is the point
Comet is Perplexity’s browser, built on Chromium — the same open-source foundation as Chrome and Edge — so your existing extensions, passwords, and bookmarks migrate across in one step.
The key distinction versus other AI browsers is that Comet runs locally. You do not hand your credentials to a cloud server; the AI layer sits inside the browser on your machine. That matters if your team handles client data, commercially sensitive proposals, or anything you would not want sitting on a third-party server.
Setup takes under five minutes: download the installer from perplexity.ai, run it, import your profile from Chrome or Edge, and you are in.
Three real tasks
HubSpot, cross-tab drafting, and a lead list in Excel
Opening HubSpot in Comet looks identical to Chrome. The difference appears when you open the assistant panel alongside it. From a deals view, you can ask it to filter to open and in-progress deals and write a summary of each. It reads the page, applies the filter, and returns a structured summary — without clicking through individual records or exporting anything. You can also dictate the request, such as “produce a summary of the deals ordered by due date,” with no special setup. For teams chasing pipeline status or prepping weekly reviews, this cuts the mechanical part of the work considerably.
The second task is where Comet shows something genuinely different. The assistant is aware of all open tabs at once. After researching a topic — use cases, community feedback, pros and cons for a content brief — in one tab, it was asked to use that research to populate a working document open in another. It reviewed the research tab and wrote directly into the document: no copy-paste, no switching back and forth, no prompt re-entry. The context holds across the session, not just within a single window. This is most useful for knowledge-heavy work: competitive intelligence, briefing documents, first-draft proposal sections.
The third task was researching primary and secondary schools in a specific area and populating an Excel spreadsheet with publicly available contact details. The assistant ran the research, decided how to structure the spreadsheet, and populated it directly — school names, contact information, and relevant associations in a clean table. No copy-pasting, no opening each website, no reformatting rows. The same approach applies to any outreach or research operation: prospect lists, supplier directories, event attendee research, competitor tracking. If the data is public and the output belongs in a spreadsheet, Comet handles the full loop from search to structured output.
The caveats
The risks you should not ignore
Comet is new and capable. It also has real limitations that matter before you give it access to anything critical.
- Security. Recent audits have flagged prompt injection vulnerabilities. If Comet summarises a malicious web page, it could execute hidden commands from that page’s content. Do not use it on unknown or untrusted sites without understanding the risk.
- Privacy. When Comet connects to your email, calendar, or cloud storage, what it accesses is sent to Perplexity’s servers for processing. Check what you are sharing before connecting sensitive data sources — for client-facing firms, this is not a minor consideration.
- Reliability. AI assistants misread context, miss information, and occasionally produce confident-sounding errors. Human review is not optional. Any output feeding a client deliverable, a CRM record, or a financial model needs a check before it goes anywhere.
- Maturity. Enterprise-grade controls and connectors are still developing. If you need granular permissions, audit logs, or role-based access management, Comet is not there yet.
Adopt it safely
Start with one workflow, not the whole firm
The worst way to adopt a tool like this is to hand it to everyone at once and wait for people to figure it out. The better path follows the same logic as any governed AI adoption.
- Identify the right workflow. Find one person or team drowning in repetitive research or admin — sales or operations are the obvious starting points. Pick two or three specific tasks: email follow-ups, prospect research, competitor tracking, pipeline summaries.
- Run a contained pilot. Test those workflows and measure the time saved — actual recorded time, not estimated. Find the highest-leverage workflows before you scale anything.
- Build the guardrails first. Define what data Comet can access. Set a clear policy on which sites and data sources are in scope. Keep humans in the review loop for any output that touches clients or commercial decisions. Skipping this is how firms end up with avoidable data incidents.
- Evaluate before you expand. One or two people seeing genuine time savings is not the same as a firm-wide rollout making business sense. Calculate the ROI at pilot scale before you commit further.
The honest read
A useful tool, used in a contained way
Perplexity Comet does real work. The cross-tab awareness, the local processing model, and the direct integration with live applications like HubSpot and Excel are genuinely useful for professional services teams where research, data gathering, and document prep consume significant senior time.
The hype around browser agents tends to show you the clean demo and omit the caveats. The caveats here are real: security vulnerabilities under active development, privacy considerations that require proper data handling decisions, and reliability limits that mean human review remains non-negotiable.
Used in a contained, governed way — starting with one workflow, measuring the impact, and building guardrails before expanding — Comet is the kind of tool that earns its place in a working day.
That is the point of tools like this. Not to replace judgement, but to remove the mechanical work that sits underneath it, so the people your firm pays for their expertise are actually spending time on the work that justifies the fee.