Paper Instruments

AI agents for work: which tools to weigh and when each one fits

This guide compares nine AI agents for desk work on what each does well, the models behind it, where it runs and what it costs: a configured professional agent, general agents from the model labs, an agent inside Microsoft 365, personified bots and hands-off cloud agents.

Checked on

At a glance

ProductBest forAvailable onStarts atFree access
Claude CoworkGeneral agent you extend with role pluginsDesktop app for Mac and Windows, webPaid plans from $20 a monthNone for Cowork: paid plans only
ChatGPT WorkAgent work inside the ChatGPT subscriptionDesktop app for Mac and Windows, web, mobileIncluded with plans, Plus is $20Limited desktop access
Microsoft 365 Copilot CoworkTenant-wide work in Outlook, Teams, WordInside Microsoft 365, under admin controls$21 a user a month list plus creditsNone: needs a Copilot license
Perplexity ComputerAssigned research tasks ending in cited reportsWeb and a Mac appPro $20 a month, credits 100 for $1Free search, limited Computer turns
Ultramarine WorkProfessional work in over 130 modes, HR includedDesktop app for Mac and Windows, plus the web$5 a month7 days free on Starter and Plus
ManusHands-off cloud tasks ending in decks or dashboardsIn the browser, on its own cloud VMFree plan, Pro from $20 a month300 credits a day free
Grok BotNamed role Bots that keep working in the backgroundDesktop app for Mac, Windows and LinuxIncluded with X Premium, $30 a month7-day trial credits on some accounts
Gemini TaskBrowser tasks, form filling and signed-in lookupsGemini app on web and mobile, driving ChromeGoogle AI Pro at $19.99 a monthNone: Pro or Ultra required
LindyRecurring Slack and Teams coordination and outreachSlack, iMessage, Gmail and phoneFrom $49.99 a user a month$50 in credits for 7 days

The tools to weigh

  1. 1. Claude Cowork

    Anthropic

    Good for
    Anthropic’s general agent for work with files, folders and the browser. Its plugin marketplace adds role plugins, including Anthropic’s HR plugin with hiring and people-ops skills, and a plugin can bundle several skills. People who live in Claude can hand it survey files and get a draft benchmarking workbook.
    How it works
    A general agent from Anthropic: desktop app for Mac and Windows, web, and cloud sessions with a built-in browser.
    Runs on
    Claude models chosen by plan: Opus 5.5, Sonnet 5.5, Haiku and Fable (November 2026).
    Price
    Paid plans only: Pro $20 a month ($17 annual), Max $100, Ultra $200. Team seats $20 to $125.
    Choose it when
    You want one model family you extend yourself with plugins, working in your own folders, without configuring a practice first.

    Sources: Anthropic support: Cowork overview, Claude plans and pricing

  2. 2. ChatGPT Work

    OpenAI

    Good for
    OpenAI’s agent for multi-step work inside the ChatGPT app, launched July 9, 2026: run deep research on a market rate, then draft the pay communication and the leadership summary, or turn an HRIS export into a filtered analysis. It works on desktop, web and mobile, and Business adds shared workspaces for a whole team.
    How it works
    A general agent inside ChatGPT: desktop app for Mac and Windows, web and mobile, with connectors and a browser.
    Runs on
    OpenAI’s own models, GPT-5.6 (Sol, Terra, Luna), with GPT-6.1 Sol rolling out in Work and Codex (November 2026).
    Price
    Included in every plan from Free to Enterprise. Plus is $20 a month and Business $25 monthly or $20 annually, and Free or Go stays desktop-limited.
    Choose it when
    Your team already runs on ChatGPT Business, so one subscription covers drafting, coding and delegated tasks under Business controls.

    Sources: ChatGPT plans and pricing, OpenAI Learn: pricing and usage limits

  3. 3. Microsoft 365 Copilot Cowork

    Microsoft

    Good for
    Cowork carries out multi-step tasks across Microsoft 365: it drafts documents in Word, prepares Outlook mailings, posts in Teams and works with tenant files. Released in June 2026 and built with Anthropic on the same agent harness as Claude Cowork, it works inside your tenant under Purview compliance and audit logs.
    How it works
    An agent inside Microsoft 365, reached from the Copilot app, that works across the tenant under admin controls.
    Runs on
    A model picker with Auto by default, plus GPT 6.1 Sol, Claude Opus 5.5, Claude Sonnet 5.5, Grok 4.7 and others (November 2026).
    Price
    Needs a Microsoft 365 Copilot license ($21 a user a month list, $18 promo to 2026-12-31) plus usage-based billing in Copilot Credits.
    Choose it when
    Employee files live in SharePoint and Teams, and IT wants rollout, governance and Copilot Credit limits managed from the Microsoft admin center.

    Sources: Microsoft Learn: Microsoft Copilot overview, Microsoft Learn: Copilot Cowork common questions, Microsoft 365 Copilot pricing

  4. 4. Perplexity Computer

    Perplexity

    Good for
    Perplexity’s research computer: assign it a question and it plans, browses and reports back with citations. An HR manager can ask for the state-law leave comparison behind a handbook chapter or a cited market review, and Perplexity publishes credit ranges per task type, from 350 to 950 credits for a research report.
    How it works
    A task computer from Perplexity in the browser and a Mac app: it plans, browses and returns a cited report.
    Runs on
    Perplexity’s own harness over models from other labs, including GPT-5.6 and Claude models, with exact routing per task not published (November 2026).
    Price
    Metered in credits, 100 for $1. Pro at $20 a month gains a weekly Computer allowance as it rolls out, and Max at $200 includes 10,000 credits a month.
    Choose it when
    The deliverable is research itself, such as a cited comparison of state leave laws, and you want to estimate the spend from published per-task credit ranges.

    Sources: Perplexity: How credits work, Perplexity blog: Introducing Perplexity Computer

  5. 5. Ultramarine Work

    Paper Instruments

    Ultramarine Work fits people whose job produces files every week. An HR manager in its Human Resources mode gets 18 skills, from hiring and handbooks to investigations, leave cases, pay ranges and people reports, and the agent picks the skill each step needs. Its desktop app works in folders on your computer. If your files live in Microsoft 365 and IT wants admin control, Copilot Cowork fits better, and personified bots suit standing Slack routines.

    Ultramarine Work runs Ultramarine 1 Flash, our model, post-trained on GLM-5.3-Flash for knowledge work. The Human Resources mode adds what a model alone lacks: 18 skills carrying the practice’s methods, review standards and traps, from compensation benchmarking to pay equity and investigations. Paper Instruments writes the skills and trains the model, so they improve together. Standard is the default way of working, Assisted adds a frontier model as an advisor, and Advanced hands the whole task to a frontier model.

  6. 6. Manus

    Manus

    Good for
    Manus runs tasks on its own cloud virtual machine and hands back a finished output: it can browse job boards, then produce a recruitment summary, deck or dashboard in one go. Manus 2.0, from September 28, 2026, added Cascade, an in-house agent harness, and a desktop studio for reviewing the run.
    How it works
    A cloud agent that works in its own virtual machine: you give it a task and it returns the finished file.
    Runs on
    Runs Manus 2.0’s Cascade harness, with the current core model not disclosed (November 2026).
    Price
    Free plan with 300 credits a day, Pro from $20 a month with 4,000 credits or from $40 with 8,000, Team from $20 a seat, and annual billing about 17% lower.
    Choose it when
    The finished artifact matters most and the job can run hands-off on its own machine before you review it.

    Sources: Manus blog: Introducing Manus 2.0, Manus help: pricing and credits

  7. 7. Grok Bot

    SpaceXSI

    Good for
    Grok Bot gives you named Bots, each with its own role and memory, that share one persistent cloud computer: one Bot tracks benefits carrier deadlines, another takes meeting notes and follow-ups, and work continues while your laptop is closed. Sensitive actions such as sending or purchasing stop for approval.
    How it works
    Named role Bots sharing one persistent cloud computer, in a desktop app and from X, Slack and mobile.
    Runs on
    No model picker: it chooses per task from SpaceXSI and third-party models and does not show which one it used (November 2026).
    Price
    Included with The X Subscription Premium ($30 a month), Super and Ultra, or a legacy paid Cursor plan.
    Choose it when
    Work must keep running without you: routines, monitoring and coordination across Bots, with approvals gating anything sensitive.

    Sources: SpaceXSI docs: Grok Bot FAQ, SpaceXSI docs: X Subscription pricing

  8. 8. Gemini Task

    Google

    Good for
    Google’s Task agent, renamed from Spark in October 2026, drives Chrome on your desktop or a remote browser on your phone: it checks benefits portals, fills forms and gathers figures from sites you are signed in to, holding up to 15 tasks at once while you work.
    How it works
    A browser-controlling agent inside the Gemini app on web and mobile, with a remote browser for phone handoffs.
    Runs on
    Google’s own Gemini models, where usage is compute-based and refreshes every 5 hours (November 2026).
    Price
    Needs Google AI Pro at $19.99 a month or Ultra, plus a personal Google account, and it is unavailable in the EEA, UK, Switzerland and Nigeria.
    Choose it when
    The work is browser-shaped, such as form filling and signed-in lookups, and Google AI Pro is already on your personal account.

    Sources: Gemini help: about the Task agent, Gemini subscription plans

  9. 9. Lindy

    Lindy

    Good for
    Lindy is a personified teammate you hire for roles like recruiting, and it lives where people work: it screens candidate responses, schedules interviews, writes meeting notes, triages inboxes and runs scheduled routines, with 3,000-plus integrations and approvals required before anything with outside impact goes out.
    How it works
    A personified teammate that lives in Slack, iMessage, Gmail and phone, with skills, routines and approvals before external actions.
    Runs on
    Model agnostic: choose from most major models per task (November 2026).
    Price
    Free: $50 in credits for 7 days. Team from $49.99 a user a month for 5,000 credits, top-ups at $10 per 1,000.
    Choose it when
    The job is recurring coordination in Slack, such as reminders, meeting follow-ups and candidate scheduling.

    Sources: Lindy pricing

Pricing and how to try it

Plans start at $5 a month, and Starter and Plus begin with 7 days free. On our internal PI Knowledge Work benchmark, Ultramarine 1 Flash costs about $0.13 a task, about a tenth of Claude Opus 5.5 or GPT-6 Sol at comparable scores.

Start free trial See every plan

Questions

Are AI agents safe to use with employee data?

They can be, with the precautions you apply to any cloud vendor: read its data page before you load an HRIS export. Ultramarine Work changes files only in the project’s folder, and the model runs on Paper Instruments’ servers, so the content it works on is sent there to be processed. Paper Instruments is SOC 2 Type II, and our security and compliance details are on our compliance page. Keep your HRIS as the system of record and give agents exports instead of live access.

What is the difference between a mode and a plugin?

A plugin extends a general agent, and one plugin can bundle several skills and connectors, while a mode configures the agent for a whole practice before work starts. Anthropic’s HR plugin gives Claude Cowork hiring and people-ops skills when it is enabled. The Human Resources mode configures the whole practice and stays selected for the workstream, and the agent picks the skill each step needs from its 18.

Can an AI agent work directly in our HRIS or payroll system?

None of them holds a login to your HRIS or payroll system. Ultramarine Work works on files you export, never signs, files, sends or posts to a system of record, and its skills have it cite the source file for its numbers. Copilot Cowork reaches Microsoft 365 apps under admin rules, and payroll entries stay manual on all of them.

How much usage does a merit cycle task actually cost?

Ultramarine Work publishes each plan’s monthly credits, with top-ups at $10 for 5,000 credits. Perplexity meters 100 credits for $1, Manus gives 300 credits a day free, and Claude and ChatGPT meter usage inside plan windows that differ by tier, so the number of tasks a subscription covers varies by product and plan.

Should I run pay equity analysis with an AI agent?

The study itself belongs with your counsel, though agents can prepare the files. The Human Resources mode’s pay equity skill calls for regression or cohort comparisons of comparable jobs, privilege handling with counsel and suppressing small cells. None of these products makes the remediation decision, and reports need a human review before anyone acts on them.

Desktop app or web app: which one should I use?

Both, on one account, because work and credits stay in sync. The desktop app works in folders on your computer, which suits HRIS and survey exports sitting on a local drive, and the web app covers uploads and connected cloud storage. ChatGPT Work starts on the desktop too, while Manus and Perplexity Computer run their work in the cloud.

Sources

Checked on