Work-Day AI: What It Can Do for You

Workflows

Inbox triage and drafted replies

It reads what came in overnight, sorts it, and drafts the replies you would have written.

This is not "answer my email." It is: sort what came in overnight into three piles (needs me, needs a draft, noise), and have the drafts waiting when I sit down. The agent reads the inbox on a schedule, files or labels the noise, flags the handful of messages that need a human decision, and writes a reply for everything else in the tone you have shown it. You review the drafts, correct them, and send the ones that are ready. The aim is to spend less time sorting and starting replies from scratch.

The first week is teaching, not automating. Connect the mailbox, show it ten past replies that sound like you, and tell it what counts as noise for your business (the vendor newsletters you keep, the ones you don't). Run it with every send behind an approval and leave it there. Save recurring corrections in the instructions and compare the time spent reviewing with the time you used to spend drafting.

What to watch. This is the workflow where prompt injection lives: an agent reading untrusted email all day can be steered by text inside a message. Review what it changed as well as what it drafted, and keep sending behind an approval.

The weekly numbers

A report assembled from your spreadsheets or system of record, delivered on a schedule.

Every small business has a number it wants on Monday morning and assembles by hand on Sunday night: bookings, jobs completed, receivables, hours, whatever the week is measured in. The agent reads the spreadsheet or the system of record, builds the summary you have shown it once (same layout every week, this week against last, one line on what moved), and delivers it before you are awake. It reads the source without changing it and saves the report separately for you. That makes it a useful first workflow; a small weekly run may fit the usage already included in your plan.

The first week is the source, not the report. If the numbers live in a sheet that three people edit by hand, tidy it first: one row per thing, one meaning per column, no notes typed into number cells. Then show the agent one finished report as the example, choose where it should save the result, and schedule it. For Workspace, build those steps in Studio and check that it can read the source and create the report you need. Add the "what moved" line only once the plain summary has been right three weeks running.

What to watch. Garbage in. Check totals against the source: a tidy sheet can still be summarized incorrectly. Reliable inputs and a repeatable calculation matter more than polished commentary.

Proposals and client documents

A new proposal from your templates, your past work, and this client's notes.

A new proposal should start from your last three, your template, and this client's notes, not from a blank page and not from the vendor's idea of what a proposal is. The agent pulls the pieces that fit, drafts the scope and the schedule in your structure, and leaves you the two paragraphs only you can write. The same applies to statements of work, renewal letters, and onboarding packets: anything you produce repeatedly with a stable shape and a client-specific middle.

The first week is your template and your past work. Put the last few proposals and the template it should follow where the agent can reach them, and tell it which sections are boilerplate, which are always rewritten, and which numbers it must never invent. Save the template and instructions for reuse, and update them when you correct a draft. Memory can help, but it does not guarantee improvement from one proposal to the next.

What to watch. Teach it your template once, or every draft will drift toward generic. And read the numbers: scope and price are the two places a plausible draft does damage.

The periodic document in your voice

A monthly update built on industry research, that reads like you wrote it, ready to send to your clients.

A monthly note to your clients on what is happening in your industry, sourced, short, and written the way you would write it, is the workflow two of our clients asked for in almost those words. The agent runs the research on a schedule (trade press, regulators, the two competitors you actually watch), drafts the issue against your style profile and last month's edition so it reads as a series, and hands you a draft with sources to approve and send. It uses more of the building blocks than any other card here: reach for the research, continuity for the voice, output for the piece, autonomy for the schedule, sharing for delivery, and governance for review before sending.

The first week is entirely about voice. Before the first run, write the style profile down: three past pieces you are proud of, the words you never use, how you open, how you close, how long it runs. Then fix the sources list and the sections. Save the profile with the workflow so each issue uses the same guidance. In Workspace, connect those steps in Studio; try one complete issue before scheduling the next.

What to watch. Keep the research sources available for review and check the claims that matter to your clients before sending. The finished piece need not be crowded with citations. Measure the usage of a complete issue before choosing the schedule.

Before the call

A one-page brief for each of today's meetings: who is coming, where things stand, and what you promised.

The calendar says you are meeting a prospect at ten, a client at two, and a vendor at four. Each morning the agent looks ahead, gathers what you already know about each one (your recent emails, the notes from the last meeting, the open items, the proposal you sent) and what is new in public (their website, recent news), and leaves a one-page brief for each meeting: who is coming, where things stand, what you promised, and two questions worth asking. It pairs with the next card: one prepares you for the conversation, the other turns it into follow-up.

The first week is the template and the sources. Decide what a brief should hold and how long it can be, show the agent one you would have liked to have, and tell it where your client history lives. Start with external meetings only and run it each morning. Ask it to say "nothing found" rather than fill a gap, and check the briefs against what you know before relying on one for a meeting you have not prepared for yourself.

What to watch. Public web pages are untrusted text; the brief should summarize them, not follow instructions in them. Keep briefs internal, and check that one client's details never land in another client's brief. If your calendar holds personal or sensitive appointments, limit which meetings it reads.

After the call

Your note-taker's transcript turned into the follow-up email, the action list, and the internal note.

If your note-taker produces a transcript, this workflow gives it a job. After a client call, the agent reads the recording's transcript (from Zoom's own notes, Otter, Fireflies, or whatever you run) and produces three things: a follow-up email to the client in your voice, an action list with an owner and a date on each line, and an internal note that lands where you track projects. You review, confirm the owners and dates, and approve the email. The client gets a recap the same afternoon, which is the kind of thing clients remember.

The first week is plumbing and a template. Point the agent at where the transcripts land, show it one follow-up you have actually sent so it learns the shape, and tell it where the internal note should go. Read every recap before it goes out. Ask it to flag missing owners and dates instead of inventing them, and save your team’s names and roles in the instructions.

What to watch. The transcript is untrusted text too; treat it like an email. The follow-up goes to a client, so send stays behind an approval. Check the action list against the transcript, especially commitments and deadlines.

Forms and data entry across web apps

The renewal, the portal, the vendor form that has no integration, filled in while you do something else.

The insurance renewal, the licensing portal, the vendor onboarding form, the supplier site that has no export: the jobs where you copy the same twelve fields from one screen into another. The agent opens the site in a browser it controls, fills in what it can from the documents you have given it, and stops at anything it should not touch. This is exactly the work people most want gone, and exactly where credentials and payments live, so it is also the card with the most tension in it.

The first week is scope. Pick one form, give the agent the source document, and watch it do the job once with every step visible. Use the platform’s secure sign-in flow and explicitly require review before submission. Microsoft Cowork browser work requires local Edge and admin enablement. Workspace Studio does not provide general browser control, and personal-only Spark is not a substitute for a business account. Once the fill is reliable, the job becomes "review and submit" rather than "type."

What to watch. Keep passwords out of task messages. A site may block automation or require human input, and repeated attempts can add usage. Keep submission behind an approval whenever a form files, pays, or binds.

Paperwork intake and renewal dates

Contracts, certificates, and licenses filed where they belong, with the dates that matter in one tracker.

Every business collects paperwork with a date buried in it: the lease, the vendor contract with an automatic renewal, the certificate of insurance a subcontractor sent, the business license, the equipment warranty. The agent watches the folder or inbox where these arrive, renames each file to your convention, files it where it belongs, and adds a row to a tracker: what it is, who it is with, when it starts, when it ends, and the notice deadline if there is one. The renewal stops being a surprise, which is the whole point.

The first week is the backlog and the rules. Write down your naming convention and folder structure, then let it work through a small batch of existing documents while you check every row. Have it mark anything it is unsure of instead of guessing, and limit it to folders you have chosen. Once the backlog is clean, schedule it for new arrivals and add reminders ahead of each notice deadline.

What to watch. Scans and photos are where extraction goes wrong; check dates and parties against the page. Keep the original file untouched and treat the tracker as a summary, not the record. This card stops at tracking: anything that pays, signs, or cancels stays with you.

Marketing collateral from one source of truth

The post, the one-pager, the carousel, and the email, all from the same brief.

One brief in, many pieces out: the post, the one-pager, the carousel, the email, all saying the same thing in the shape each channel needs. The agent works from a written brief and your style profile, drafts each piece, and lays out the visual ones (a one-pager, a slide, a carousel) so you are editing rather than assembling. This is the workflow we used to produce this guide series: the research became the guide, the guide became a post, the post became a carousel, all from one source document.

The first week is the brief. If the message is not written down, there is nothing to derive from, and each piece drifts. Write the brief, decide the channels, and give the agent one example of each piece you have liked. Platforms with a design surface will hand back finished visual pieces; the others will hand back copy you lay out elsewhere.

What to watch. Consistency is the whole point, so review the pieces side by side rather than one at a time. Check the relevant text and visual-generation allowances; they are not always shared. Budget for revisions as well as the first draft.

Setup effort by platform

Building block Claude Cowork ChatGPT Work Grok Bot Gemini for Workspace and Workspace Studio Microsoft 365 Copilot and Copilot Cowork Muse
Continuity Built in Built in Configure Configure Configure Configure
Output Built in Built in Built in Built in Built in Configure
Autonomy Configure Configure Configure Configure Configure Configure
Reach Configure Configure Configure Configure Configure Configure
Sharing Built in Built in Configure Built in Built in Configure
Governance Configure Configure Configure Configure Configure Configure
Extensibility Configure Configure Configure Configure Configure Configure

Grid notes

Continuity, Claude Cowork
Chat and Cowork memory, plus project instructions and context.
Continuity, ChatGPT Work
Projects and account memory settings support repeat work.
Continuity, Grok Bot
Persistent bot context; keep procedures in explicit instructions.
Continuity, Gemini for Workspace and Workspace Studio
Saved templates, Gems, and Studio flows carry instructions.
Continuity, Microsoft 365 Copilot and Copilot Cowork
Use project context and saved instructions for repeat work.
Continuity, Muse
Remembers across tasks; training on your conversations is on by default.
Output, Claude Cowork
Documents, spreadsheets, slides, and visual pieces.
Output, ChatGPT Work
Documents, spreadsheets, slides, and web content.
Output, Grok Bot
Documents, spreadsheets, and slides previewed in chat.
Output, Gemini for Workspace and Workspace Studio
Docs, Sheets, and Slides; automated output depends on Studio steps.
Output, Microsoft 365 Copilot and Copilot Cowork
Word, Excel, and PowerPoint output.
Output, Muse
Business document output not yet documented.
Autonomy, Claude Cowork
Cloud schedules; local tools still need the desktop online.
Autonomy, ChatGPT Work
Cloud schedules and supported app events; sources must be cloud-accessible.
Autonomy, Grok Bot
Routines run on the persistent cloud computer.
Autonomy, Gemini for Workspace and Workspace Studio
Studio runs scheduled flows and triggers from Workspace and supported outside apps.
Autonomy, Microsoft 365 Copilot and Copilot Cowork
Cloud schedules and supported events; local Edge needs the device.
Autonomy, Muse
Background work on its cloud computer; Mac app tasks need the device.
Reach, Claude Cowork
Connected folders and apps; local access needs an online desktop.
Reach, ChatGPT Work
Connected apps and uploaded files; local access is a separate mode.
Reach, Grok Bot
Connect apps or upload files; optional local commands need approval.
Reach, Gemini for Workspace and Workspace Studio
Workspace files and services; outside apps through Studio integrations in beta.
Reach, Microsoft 365 Copilot and Copilot Cowork
OneDrive, SharePoint, Outlook, and supported connections.
Reach, Muse
Small-business connectors; one account per Gmail or Outlook connector.
Sharing, Claude Cowork
Shared artifacts and projects; confirm the destination for delivery.
Sharing, ChatGPT Work
Shared workspace results; delivery depends on the connected app.
Sharing, Grok Bot
Deliver through an authorized app or browser connection.
Sharing, Gemini for Workspace and Workspace Studio
Native Workspace sharing and Studio email actions.
Sharing, Microsoft 365 Copilot and Copilot Cowork
Native M365 sharing and Teams.
Sharing, Muse
No shared workspace or team features published.
Governance, Claude Cowork
Configure approvals and usage limits; audit access depends on plan.
Governance, ChatGPT Work
Configure approvals and spend controls; audit coverage depends on plan.
Governance, Grok Bot
Teams rules and approvals; stronger controls require Enterprise.
Governance, Gemini for Workspace and Workspace Studio
Review flow actions; approval requirements depend on admin policy.
Governance, Microsoft 365 Copilot and Copilot Cowork
Admin spending policies, audit, and labels; Cowork DLP rolling out.
Governance, Muse
Owner approvals for posts, messages, purchases; no admin controls.
Extensibility, Claude Cowork
Skills, plugins, connectors, and team distribution.
Extensibility, ChatGPT Work
Plugins and reusable instructions; check each app’s supported actions.
Extensibility, Grok Bot
Skills, routines, and connectors under team policy.
Extensibility, Gemini for Workspace and Workspace Studio
Studio steps, webhooks, and integrations; Skills replace Gems in flows.
Extensibility, Microsoft 365 Copilot and Copilot Cowork
Skills, plugins, and Copilot Studio for custom workflows.
Extensibility, Muse
Custom connectors; no team distribution.

The workflow that is not on the list

If your week does not look like any of the nine, the grid still works. Write down the job in one sentence, name the blocks it needs, and read across. If even one required step cannot be completed on the platform you already pay for, resolve that gap before buying or scheduling anything.

Where this guide picks up

The readiness guide answers one question: should you let one of these agents into your business at all. This guide assumes you have decided yes, or nearly, and answers the next one: what would it actually do for you on a Tuesday. So it is organized by the job, not the vendor. Find the workflow that looks like your week, see what it needs, and the card tells you which platforms we recommend for it today.

The grid shows what adoption takes: "Built in" means ordinary account setup and instructions; "Configure" means connecting apps or assembling a workflow. The workflow cards make a separate recommendation: "Go" with the stated setup and review, or "No-go" where a specific limitation rules the platform out.

How to read a workflow card

Each card has the same parts. What it looks like, in a small business, in plain terms. What it needs, as building blocks. Which platforms we recommend for the job today, alongside the setup they need. What the first week of setup involves. How the cost behaves, since pricing itself lives in the readiness guide. What to watch, which is usually the step where the workflow quietly goes wrong. And where custom work helps when your systems, rules, or handoffs need more than the standard setup. These are opportunities to adapt the workflow to your business, not requirements for every implementation.

The cards assume business accounts, connected cloud sources for scheduled work, and human review before sending or submitting. Google means Workspace apps and Studio, not personal-only Spark; Grok means Cursor Teams. Muse appears for comparison and is No-go on every card until Meta offers a business plan and admin controls. Browser jobs need a separate check of the chosen website and device requirements. A cloud schedule cannot use an offline laptop’s files or browser.

What we don't put on this list yet

Two kinds of workflow are missing on purpose. Customer-facing work (a support inbox answered by an agent, a chat on your site) and anything that touches money (invoicing, payments, payroll, bookkeeping) are both possible on some of these platforms today, and both deserve deeper consideration than a guide can give: what a wrong answer costs when a customer receives it, what a wrong number costs when it is a payment, and who is accountable when it happens. We would not recommend either to a small business without that conversation. It is one we are glad to have.

Grid notes

The grid assumes the common business tools described in the cards. Connecting a supported app is configuration; connecting an unsupported internal system may require custom work. Where our assessment comes from documentation rather than daily use, the card says so. Try the complete job with your own inputs before depending on it. We refresh the recommendations monthly.