AI Agents in Google Workspace & Microsoft 365 2026: Govern Safely

Last month a Pune-based CA firm called me in a panic. One of their junior accountants had turned on Copilot in Microsoft 365 and, within a week, the AI had happily summarized a client's confidential balance sheet into a shared chat that three interns could read. Nobody leaked anything on purpose. The tool just did what it was designed to do: surface information the user technically had access to. The problem was that the firm's file permissions had been a mess for years, and the AI simply exposed that mess at machine speed.

This is the uncomfortable reality of 2026. Both Google and Microsoft have shoved AI agents into the office suites your team already uses. Gemini sits inside Gmail and Docs. Copilot lives in Outlook, Teams and Excel. Microsoft's own data shows Copilot adoption crossing hundreds of thousands of organizations, and Indian SMBs are switching it on faster than they're thinking about the consequences. An AI agent that can read every file, draft every email, and query every SharePoint site is a productivity dream and a compliance nightmare wearing the same shirt.

This post is about getting the dream without the nightmare. I'll walk you through practical AI agent governance for SMBs: how to deploy Gemini and Copilot without accidentally leaking client data, breaking DPDP compliance, or discovering three months later that your intern read the payroll sheet. You'll get a governance checklist, a real deployment walkthrough, a comparison of what the two suites actually control, and the mistakes I've watched businesses make firsthand.

Key Takeaways

• AI agents inherit your existing file permissions. If your Google Drive or SharePoint access is sloppy, the AI will expose that sloppiness instantly. Fix permissions before you deploy.
• Run a phased rollout: pilot with 5-10 users, audit what the AI can see, then expand. Never flip it on org-wide on day one.
• India's DPDP Act 2023 makes you responsible for personal data the AI processes. You need consent records, data classification, and a documented purpose.
• Both Microsoft 365 Copilot and Google Workspace Gemini keep your data out of model training by default on paid business tiers, but the admin settings differ and defaults change.
• Budget roughly ₹2,100-2,900 per user per month for the AI add-on on top of your base licence. For a 25-person team that's a real line item.
• Assign one person as the AI governance owner. Ungoverned AI is worse than no AI.

Why AI agents in your office suite are a bigger risk than regular software

Regular software does one thing. An accounting tool touches accounts. A CRM touches customer records. An AI agent inside Microsoft 365 or Google Workspace touches everything the logged-in user can reach, and then it reasons across all of it.

Ask Copilot "summarize what happened with the Reliance account last quarter" and it will pull from emails, Teams messages, meeting transcripts, Excel sheets and SharePoint documents in one shot. That's the magic. It's also the risk. The AI doesn't know that the compensation spreadsheet in a shared folder was only meant for the two founders. It knows the user can open it, so it uses it.

The three specific risks I see repeatedly with Indian SMBs:

• Over-permissioned files. Most SMBs have "Anyone at the company can view" folders full of contracts, salaries, and client PII. The AI reads all of it.
• Shadow AI. Employees pasting client data into free ChatGPT or random browser extensions because IT never gave them an approved tool. This is rampant and invisible.
• Compliance blind spots. Under the Digital Personal Data Protection Act, you're the data fiduciary. If an AI processes a customer's Aadhaar or health record without a lawful basis, that's on you, not on Microsoft.

If you're still deciding between the two suites before you even get to governance, our breakdown of Google Workspace Gemini vs Microsoft 365 Copilot cost is a good starting point.

What data protections do Copilot and Gemini actually give you?

Let me clear up a common confusion. On the paid business and enterprise tiers, neither Microsoft nor Google uses your prompts or business content to train their foundation models. Your data stays inside your tenant's compliance boundary. That's contractually stated. The catch is that "not training on your data" is not the same as "your data is safe from your own users."

Here's how the two suites compare on the controls that actually matter for governance.

Governance control	Microsoft 365 Copilot	Google Workspace Gemini	Why it matters
Respects existing file permissions	Yes, uses Microsoft Graph permissions	Yes, uses Drive sharing settings	Bad permissions = exposed data via AI
Data used to train foundation models	No (Business/Enterprise)	No (Business/Enterprise)	Protects your IP and client data
Sensitivity labels / DLP	Purview sensitivity labels, mature	Google DLP + labels, decent	Prevents AI from surfacing tagged secrets
Audit logs of AI activity	Yes, via Purview audit	Yes, via Admin/Vault logs	Needed for DPDP accountability
Data residency in India	Available with India data centres	Available with India region storage	Reduces cross-border transfer risk
Indicative add-on price (per user/month)	~₹2,900 (Copilot)	~₹2,100-2,500 (Gemini Business)	Real cost for a 25-user team

The honest takeaway: Microsoft's governance stack via Purview is more mature if you're a firm handling regulated data (finance, healthcare, legal). Google is simpler and cheaper to reason about for a lighter-touch services business. Neither one governs itself. You have to configure it. If you need help picking, our team does exactly this under IT consulting, and we handle both Microsoft 365 licensing and Google Workspace licensing for Indian businesses.

A real deployment: how a Gurugram logistics firm rolled out Copilot safely

Let me give you a concrete example, with the names changed but the numbers real.

A 40-person logistics and freight-forwarding company in Gurugram wanted Copilot for their ops and sales teams. Their motivation was solid: sales reps were spending two hours a day drafting quotes and follow-up emails to clients across Delhi NCR and Mumbai. The founder had already bought 40 Copilot licences at roughly ₹2,900 each. That's about ₹1.16 lakh per month, a serious commitment for an SMB.

When we ran the pre-deployment audit, we found the problem. Their SharePoint had a folder called "Company Docs" set to "Everyone in organization can edit." Inside it: a salary sheet, a signed customer NDA scan, and GST filings with the firm's PAN and bank details. Every one of the 40 users, including two contract data-entry staff, could have asked Copilot to fetch salary figures.

Here's what we did, in order:

1. Froze the rollout. We disabled Copilot for all but a 6-person pilot group before touching anything else.
2. Ran a permissions audit. Using SharePoint access reports, we listed every folder shared broadly and flagged 14 folders with sensitive content.
3. Applied sensitivity labels. Payroll, contracts and GST docs got a "Confidential - Finance" label via Microsoft Purview. Labelled content is excluded from Copilot responses for users without rights.
4. Rebuilt sharing groups. We replaced "Everyone" access with role-based groups: Sales, Ops, Finance, Leadership. This took two days and was frankly overdue anyway.
5. Set up audit logging. We turned on Purview audit so every Copilot query touching flagged content is logged. This is your DPDP paper trail.
6. Wrote a one-page acceptable-use policy. Plain Hindi and English. No pasting client PII into external tools, report anything odd, AI drafts must be human-reviewed before sending.
7. Ran the pilot for three weeks, reviewed the logs, then expanded to all 40 users.

The result: the quote-drafting time dropped from two hours to about 35 minutes per rep, and the finance data was locked down properly for the first time in the company's history. The permissions cleanup was worth more than the AI, honestly.

Common Mistake: Businesses buy AI licences first and think about permissions never. The AI is not the project. The permissions cleanup is the project. If you do nothing else from this article, run a "who can see what" audit before you enable any AI agent. It's the single highest-value hour you'll spend.

What does DPDP Act compliance require when AI processes personal data?

India's Digital Personal Data Protection Act, 2023 is now the law that matters, with rules being phased in. If your AI agent processes personal data of customers or employees, you're on the hook as the data fiduciary. Here's what that means in practice for an SMB.

• Lawful purpose and consent. You need a documented reason for processing personal data, and where consent is the basis, a record of it. An AI summarizing a customer complaint email is processing personal data.
• Data minimisation. Don't let the AI hoover up more than needed. Sensitivity labels and access controls enforce this.
• Purpose limitation. Data collected for order fulfilment shouldn't be repurposed by AI for, say, marketing profiling, without fresh consent.
• Breach notification. If AI exposure causes a personal data breach, you must notify the Data Protection Board and affected individuals.
• Data principal rights. Customers can ask what data you hold and demand correction or erasure. Your systems, including AI-accessible stores, must support this.

The practical move: keep your AI's data access inside India-region storage where both providers offer it, log everything, and keep sensitive personal data (Aadhaar, financial, health) behind sensitivity labels that the AI respects. This is also where a virtual office address for GST and company registration matters for firms formalizing their compliance footprint across states.

Your AI agent governance for SMBs checklist

Print this. Tape it to the wall of whoever owns IT. This is the checklist I use when setting up AI agent governance for SMBs, sequenced so you're never exposed.

Before you enable anything

• Name one governance owner. Founder, ops head, or your IT partner. One throat to choke.
• Confirm you're on a business/enterprise tier where data isn't used for model training.
• Run a permissions audit. List every broadly shared folder and file.
• Classify your data: Public, Internal, Confidential, Restricted.
• Apply sensitivity labels to Confidential and Restricted content.

During rollout

• Start with a pilot of 5-10 users from one department.
• Enable audit logging (Purview for Microsoft, Vault/Admin logs for Google).
• Publish a one-page acceptable-use policy in the languages your team reads.
• Set India data residency where offered.
• Disable or block unapproved external AI tools and browser extensions.

After rollout, ongoing

• Review AI audit logs monthly. Look for AI touching Confidential data.
• Re-run permissions audits quarterly. Access creep is real.
• Train staff on what not to paste into AI. Repeat it.
• Keep consent and processing records updated for DPDP.
• Reassess licence count. You may be paying for users who never touch it.

The staff training piece is where most SMBs quietly fail. Governance tools stop accidents, but educated employees stop 90% of them before they happen. Our guide on AI skilling for Indian SMB teams covers how to upskill without disruption.

Should you use built-in AI or build your own agents?

A question I get a lot: "Should we just use Copilot, or build a custom AI on top of open-source models for more control?" It depends on your data sensitivity and your appetite for engineering.

For most Indian SMBs, the built-in agents in Microsoft 365 or Google Workspace are the right call. They're governed within a compliance boundary you don't have to build. But if you're processing highly regulated data or want AI that plugs into your custom systems, a purpose-built agent may be worth it. We compare the trade-offs in detail in our post on open-source AI vs paid APIs for Indian SMBs, and if you're weighing model vendors, Claude for Business vs ChatGPT is a useful read.

Where a custom build makes sense, we handle it end to end through custom software development and mobile app development. And if your real pain is phone-based customer support rather than document work, an AI voicebot often delivers faster ROI than an office-suite agent. See how in our piece on AI voicebots for Indian SMBs.

Pro Tip: Don't let sales reps use AI-drafted emails without a human check. I've seen Copilot confidently invent a delivery date that wasn't in any thread, and the rep sent it. The AI is a fast intern, not a fact. Every outbound customer message with commercial commitments needs a human eyeball. Build that into your acceptable-use policy explicitly.

What will AI agents in office suites cost a 25-person Indian SMB?

Let's talk real rupees, because the marketing pages rarely do the math for you. Take a 25-person firm in Bengaluru or Hyderabad.

• Base licences: Microsoft 365 Business Standard is roughly ₹770/user/month; Google Workspace Business Standard is similar. Call it ₹19,250/month for 25 users.
• AI add-on: Copilot at ~₹2,900 or Gemini Business at ~₹2,100-2,500 per user. If you enable it for all 25, that's ₹52,500-72,500/month on top.
• Governance setup (one-time): permissions audit, labelling, policy. Budget ₹40,000-80,000 depending on your data mess.

Here's the smart move most people miss: you rarely need AI for everyone. Give it to the sales, ops and leadership roles that draft and analyse heavily. Skip it for warehouse or purely operational staff. In that Gurugram example, cutting AI licences from 40 to 24 saved about ₹46,000 a month with zero productivity loss. Buy AI by role, not by headcount.

eDarpan can bundle your licensing, governance setup and rollout together so you're not overpaying or under-protected. Our cloud migration and managed services team runs this regularly for SMBs across India, and you can see the full range on our services overview.

Frequently asked questions

Does Microsoft Copilot use my company data to train its AI?

No. On Microsoft 365 Business and Enterprise tiers, Copilot does not use your prompts, files, or content to train the underlying foundation models. Your data stays within your organization's compliance boundary. The same is true for Google Workspace Gemini on paid business tiers.

Can an AI agent see files it shouldn't?

An AI agent can see whatever the logged-in user can already access. It respects existing permissions but does not fix bad ones. If a confidential file is shared with "everyone in the organization," the AI will surface it. This is why a permissions audit before deployment is essential.

Is Copilot or Gemini compliant with India's DPDP Act?

The tools provide the technical controls needed for compliance, such as audit logs, data residency in India, and DLP. But compliance is your responsibility as the data fiduciary. You must document your lawful purpose, maintain consent records, and configure the controls correctly.

How much does Microsoft 365 Copilot cost per user in India?

Copilot is priced at roughly ₹2,900 per user per month as an add-on to an eligible base Microsoft 365 licence. Google Workspace Gemini business add-ons run around ₹2,100 to ₹2,500 per user. Prices vary with billing terms and current promotions, so confirm at purchase.

Should I give AI licences to my whole team?

Usually no. AI agents deliver the most value to roles that draft, summarise, and analyse heavily such as sales, operations, finance and leadership. Buying by role instead of headcount can cut your AI spend by a third or more without hurting productivity.

What is the single biggest AI governance risk for SMBs?

Over-permissioned files combined with shadow AI usage. Sloppy file sharing means the AI exposes sensitive data, while employees using unapproved free AI tools leak data outside your control entirely. Fixing permissions and providing an approved, governed AI tool addresses both at once.

Can eDarpan help set up AI governance for our office suite?

Yes. We handle the full cycle: licensing, permissions audit, sensitivity labelling, audit logging, acceptable-use policy, and phased rollout for both Microsoft 365 and Google Workspace. You can contact eDarpan to scope your deployment.

Bringing it together

AI agents inside Google Workspace and Microsoft 365 are genuinely useful. I've watched them give small teams the leverage of much bigger ones. But the tool that reads everything is only safe when you've decided what it should read. That's the whole game.

Effective AI agent governance for SMBs isn't complicated, it's just sequenced: fix permissions, classify data, pilot small, log everything, train people, and buy by role. Do that and you get the productivity without the panic call I got from that Pune CA firm. Skip it and you're one intern and one shared folder away from a compliance incident.

If you'd rather not figure out Purview labels and DPDP consent records on your own, that's exactly the kind of work we do. Reach out through our contact page or read more about eDarpan and how we help Indian businesses deploy technology that actually behaves. Set up the guardrails first. The AI will thank you, and so will your auditor.

Image credit: AI, Automation, and Human Judgment panel at Wikimania 2026 - 11 by Luisalvaz via wikimedia (BY-SA 4.0), sourced through Openverse.