> 📥 The full field report is available as a PDF. Download it at the top of the page — with the compliance traffic light, the sector-research example and a 5-step getting-started guide.
> 🧭 Part 3 of 4 in the series AI agents anno 2026. Four tools all called AI agents: OpenClaw, Claude Cowork, Perplexity and Hermes. Today: Personal Computer — the one I use for heavy research, and the one I hear several people already run side by side with Cowork.
1. Field report, not hype
I was sitting with a CIO at a large Danish infrastructure company. He needed the AI story in place for his 65-person IT unit. I was giving a 60-minute keynote for them.
I gave Perplexity Personal Computer one sentence: build a thorough research report on generative AI and AI agents in transport, rail and infrastructure companies 2026; find 14 documented cases; classify by maturity; be conservative on evidence; mark explicitly where sources are weak; deliver as an internal working document with clickable citations and a source-gap section.
Half an hour later the report was in the folder. 13 sections. 14 cases (Deutsche Bahn, Italferr, DSB, SMRT in Singapore, Cubic + Imperial AIDA Lab, among others). A maturity matrix on two axes. An EU AI Act timeline with the Omnibus delay sketched out. A section on the client's own starting point. And an explicit source-gap section that listed precisely where the evidence was vendor-based or conference-reported.
It was overwhelming. I had not written a single prompt beyond the first. And in truth, not even that. It was written by Claude. Some people at Anthropic now say you shouldn't prompt as much. You should build systems that prompt themselves.
2. What works: research that lands as a finished deliverable
The narrative arc was born out of the research report — not in an empty Keynote document, but in the folder.
Sector research with sources that hold. Perplexity scanned the field and delivered a report with citations for every material claim. It found 14 cases in rail and transport, assessed maturity (pilot, scaled, production) and separated verified evidence from conference-reported numbers. SMRT in Singapore — the Jarvis system with 30 years of operational data, launched April 2026 — became one of my three headline cases. I wouldn't have found it manually in under two days.
Maturity matrix that becomes slides. I asked for a placement of typical AI use cases on two axes: maturity and relevance to the IT unit. It produced a matrix and a table with concrete pilot recommendations. Then I gave Claude access to the same folder, which built the slides.
Client-specific integration. When I told Perplexity about the customer's existing digital foundation, it wrote a section that connected the sector cases to their real starting point. "Your starting point is stronger than you think" became one of my slide titles — straight from the report.
Talking points with warnings built in. Each case came with a small "traps to avoid" block. On SMRT: don't say "full operation", it's pilot. Don't name a specific LLM if the vendor doesn't disclose one. Don't claim MKBF is improved by Jarvis until it's measured. Exactly what I needed not to oversell in front of an audience that knew rail operations better than I did.
3. The difference from Claude Cowork
Cowork is an assistant you set to work in your folder. Personal Computer is a research specialist that uses your folder as input and delivers back into it. Cowork is broader. Personal Computer is deeper at what it does best. They don't necessarily compete — they complement each other.
Personal Computer orchestrates across several underlying models and Perplexity's own research stack. You normally don't pick the model yourself. The point isn't the model name, it's that the tool tries to choose the right capacity for the task.
Detour: Opus 4.8 and the quarantine
I chose to use Claude Opus 4.8 the other day and ran out of tokens and credits. Now I'm in quarantine until mid-June. Haha! But many agents running in parallel on their own is smart. It's the opposite of sitting with five chat windows open and hand-sorting outputs.
One thing I've particularly appreciated: it documents what it did. Each deliverable comes with a short note on which sources it pulled, which files it read, and where it was uncertain.
And then there's the sandboxing. Personal Computer works in an isolated action environment, asks for approval on sensitive actions and gives better traceability. That matters. But it's not the same as every error being rollback-able, or processing being local. More on that below.
4. What it's not so good at
Research strength is not workflow strength. Here's the honest part.
Personal Computer is a strong research specialist. It is not a general assistant. It does not clean folders for you. It does not build robust working templates. It does not execute multi-step workflows across apps the way a broader desktop agent does. It doesn't necessarily build the keynote itself either — it delivers material, structure, arguments and sources. You still build the presentation, or let another tool (e.g. Claude Opus 4.8) build it further.
Ask Personal Computer to build a course platform, a blog series or a campaign, and it can go sideways. Ask it to find answers, assess sources and deliver a documented working basis, and it's in a class of its own.
It resets between sessions. No reliable persistent memory across everything you've worked on. On the customer project, that meant feeding it the client context in each major session.
Trick: the handover file
You can shrink the problem by giving it access to README files in the folder it's working in. I often use a short handover file — context, goal, sources, assumptions, prior decisions written down. Then it can continue without reinventing the project. Have Claude Cowork write that README and you have both tools in play in the same material, possibly in different folders so they don't step on each other.
It also doesn't polish the last layer. It has put a wrong year on a citation for me. It has produced a chart where the X-axis didn't logically make sense. The classic agent experience: 95% impressive, 5% embarrassing. And the 5% is your responsibility — especially in front of a customer who knows their sector better than you.
The app is new. Personal Computer was broadly released in Perplexity's Mac app in May 2026, and it shows. Delivery to the folder has taken longer than expected. One long session crashed mid-task. It's a young app. Know that.
And there's the platform. The Personal Computer feature with access to local files and native Mac apps is Mac-only right now. On Windows you can still use Perplexity in the browser and other Perplexity products, but the local desktop agent with Mac file access is not the same.
5. Does data leave your machine?
Personal Computer feels local because it works on your Mac, can access local folders and can operate in your apps. But local access is not the same as local processing. Perplexity describes Personal Computer as a feature that connects local files, apps and browser context with Perplexity Computer, so the agent can work with the context on your desktop. Assume by default that relevant content from the files and apps you grant access to can be sent to Perplexity's service and processed in their model and service infrastructure.
On the customer project
Everything I gave Perplexity access to was public material: BIM program, digital asset management partnership, O&M tender, press releases, sector reports, public sector cases. For that kind of research Perplexity is almost purpose-built. The picture changes if you put client data, personal data or confidential documents in the research folder.
On Perplexity Free, Pro and Max AI Data Retention is on by default. Perplexity states data may be used for AI training unless the user actively turns it off. Opt-out is forward-only. Rule of thumb: red or amber-red for client data, confidential business data and personal data unless an explicit risk assessment has been done and organizational controls are in place.
On Perplexity Enterprise Pro the baseline is different. Perplexity states enterprise data is not used for training or fine-tuning, and agreements with third-party model providers prohibit training on enterprise data. But Enterprise use is not automatically compliant. It requires concrete review of DPA, sub-processors, retention, access controls, data transfers and the specific data category.
Sensitive personal data, full KYC folders, patient data, HR cases and heavily regulated material is still red without a separate risk assessment. I have not found sufficient public documentation for a full EU Data Boundary or EU-only processing for Personal Computer. Until that is concretely documented in your contract, regulated use cases should be treated conservatively — likely because they use several LLMs, not all of them hosted in the EU.
The practical difference
Sandboxing is about what the agent is allowed to do on your machine. Compliance is about which data the agent gets access to, where it's processed, who can process it, how long it's stored, and whether the vendor contract matches your data category.
My recommendation is simple: give Personal Computer its own research folder. Use it for public material, sector research, industry analyses, keynote prep, documents without client-sensitive content. Keep personal data, KYC, draft contracts, internal strategy papers and confidential customer documents out — unless your organization has approved the setup.
6. Who it's for
If you do research-heavy work where the output needs to be a finished document, it's hard to beat right now.
| Role | What it solves |
|---|---|
| Consultants | Preparing for an industry they don't live in every day. |
| Analysts | Delivering market briefings. |
| Journalists | Digging before an interview. |
| Lawyers | Initial, source-backed research on public material (not a substitute for legal judgment). |
| Investors | Understanding a sector before due diligence. |
| Leaders | Building briefings before board meetings. |
For them it is the strongest research tool I've tried. For a small team without heavy security process it's fast to get started. In a real enterprise context the tool is technically easy but the approval can still take time: DPA, SSO, retention, access controls, audit logs and internal AI policy must be in place.
For financial services, healthcare and other regulated organizations it's not only a data-residency question. It's a full compliance question: DPA, sub-processors, data transfers, retention, access controls, audit logs, DPIA and the specific data category. Until EU-only processing and enterprise controls are documented in the contract, it's no to sensitive data.
For Microsoft houses
Perplexity has launched a Microsoft 365 integration in preview. It gives Perplexity access to Teams threads, mails and SharePoint files. In my view this is one of the most interesting challenges to Microsoft's own research feature, because Perplexity tries to combine web research, enterprise sources and model choice in one workflow. But that's also where the data-residency and compliance question gets hardest. I'm watching it closely.
One last thing: I've had real value from Cowork side by side with Perplexity Computer, so the "combo pack" is also an option. Material for another article, perhaps.
7. How to get started
Personal Computer is the opposite of my agent swarm from issue 1. Where OpenClaw needed technical hands and an isolated environment, and Cowork needed a folder and a sentence, Personal Computer needs a folder and a question. But because the threshold is low, it's easy to forget the important part:
Personal Computer touches your real files, and relevant context can leave your machine. Choose the plan deliberately. Choose the folders deliberately. And always proofread.
1. Download the Perplexity Mac app. Personal Computer requires at least a paid plan. Always check current pricing and access — Perplexity changes packages fast.
2. Create a research folder on your Mac. E.g. "Perplexity-research". Not your whole Desktop — one folder. Keep client data and personal data out of it.
3. Pick a concrete task from this week. A keynote you're preparing, a briefing, a sector you need to learn. The output should be a document, not a workflow.
4. Write the prompt as if briefing a senior consultant. Three things: goal, sources, format. Add one more: explicitly ask for a source-gap section. "Mark where the evidence is weak." That one line lifts the output from useful to reasonably reliable.
5. Read everything afterwards. Always. Especially citations, numbers, dates, names and regulatory claims. The 5% mistake is always somewhere. That's your job, not the machine's.
Do that and you have the fastest research assistant I've yet tried. And one more real agent alongside Cowork. Not a replacement. A supplement.
8. Next issue
Number 4 in the series, the last one for now: Hermes Agent from Nous Research. The opposite extreme of Personal Computer. Open-source, runs locally on your own machine, persistent memory, built with sandboxing close to the core. Three months under test, preliminary assessment. What it can do, where it's strong, and why I installed it myself instead of waiting.
Stefano Vincenti — GenAI strategist and architect. 25 years in IT and digital transformation. Co-founder of BotTellMe. External lecturer at ITU and DIS Copenhagen. Partner at TryZone. Subscribe to the newsletter and get the next issues directly.
This guide is a professional snapshot as of June 2026 and constitutes neither legal advice, GDPR assessment nor IT security advice for your specific situation. Compliance, data residency and security assessments are general orientation points based on publicly available sources. Final assessment for your data, processes and use cases must be made by your own compliance function, DPO, lawyer or IT security function. Tools, prices, terms and security advisories change quickly — always verify current conditions before deciding.