Skip to content

Microsoft Copilot, Google Gemini or a Custom AI Agent: How to Choose

AI Agents and Chatbots · · 6 min read

By Chief Executive Officer
Two colleagues at an open-plan desk look at a laptop screen and talk it through

Your Microsoft 365 or Google Workspace bill probably already has an AI assistant on it, or a quote to add one. A few months in, the questions arrive. Is this enough? Do we pay for everyone? Or do we need something built around how we work?

Both answers can be right, for different jobs. This custom AI agent vs Copilot guide compares Microsoft Copilot, Google Gemini and a custom agent on setup, data access, control and cost, so you can decide per use case instead of per vendor.

What Copilot, Gemini and a Custom AI Agent Each Do

Copilot and Gemini are assistants that live inside the office tools your staff already use. They draft, summarize and search. A custom AI agent is software built for one job in your business: it reads your systems, follows your rules and takes steps, such as chasing an invoice or filing a claim. If the difference between a chatbot and an agent is new to you, our guide to chatbots and AI agents covers it.

The assistants are strong at what they were designed for. Microsoft's documentation says Copilot works with content in Microsoft Graph that the user already has permission to access, and that prompts and responses are not used to train its foundation models. Google says Gemini only retrieves Workspace content the user can already open and does not train on customer data without permission. Both are vendor statements, but they describe the same sensible design: the assistant sees what the person sees.

What the Evidence Says About Results

The best public trial is from the UK government. Between 30 September and 31 December 2024, the Government Digital Service ran a trial of Microsoft 365 Copilot with 20,000 government employees. Its report found an average of 26 minutes saved per day (GDS, June 2025). Read the caveats too: participants estimated their own time saved in ranges, 17% of users did not notice any clear time savings, the report saw limits with complex, nuanced or data-heavy work, and it could not tell how the saved time was spent.

Business adoption is slower than the headlines. In a 2025 Gartner survey of 187 IT leaders, 40% were piloting Microsoft 365 Copilot and only 5% of those who had finished a pilot were moving to a larger deployment that year (techpartner.news, June 2025). Of 127 respondents who had piloted or rolled it out, nearly half rated it "some value, shows promise." Gartner's analyst also said the effort to manage risk was significantly greater than most organizations planned for, and linked it to scattered content, excessive sharing and weak governance. An earlier Gartner survey of 132 IT leaders found 64% said information governance and security risks took significant time and resources (Computerworld, 2024).

The wider picture matches. Forrester predicted in October 2025 that enterprises will defer a quarter of planned AI spend into 2027, with fewer than one-third of decision-makers able to tie AI to financial growth. And in Gallup's survey of 23,717 US employees in February 2026, half used AI in their role at least a few times a year, but only 41% said their organization had integrated AI tools. People are using these tools. Firms are still working out how to get value from them.

Where Off-the-Shelf Assistants Run Out

None of this makes Copilot or Gemini a poor choice. It shows where their design stops.

  • Data they can reach. Out of the box they are strongest on mail, documents, chats and meetings in their own suite. Both vendors offer connectors to other systems. Google lists permission-aware search and grounding on third-party data in its Gemini Enterprise editions, and Microsoft says Graph connectors return external data only if the user has permission to it. The work is in the setup: someone has to connect, test and maintain each source.

  • Workflow fit. An assistant helps a person do a task faster. It does not own a process end to end, such as matching purchase orders to invoices and posting exceptions to your ERP. For that you need steps, rules and a record of what happened.

  • Permissions that are already messy. Because the assistant follows existing permissions, years of loose sharing become visible quickly. Gartner's analyst said deployment plans often assume a clean data environment, which is rarely true, and that fixing it can mean upgrading licenses or adding governance tools, which raises total cost.

  • Cost shape. As of October 2026, Microsoft Copilot Business is listed at $21 per user per month as an add-on to a Microsoft 365 plan, with some advanced features billed by usage. A first-year promotional price of $18 applies to existing Microsoft 365 Business customers who commit to a year and sign up between July 1 and December 31, 2026; monthly billing is listed at $25.20 (Microsoft). Microsoft has renamed Microsoft 365 Copilot as Microsoft Copilot, so you will see both names. Gemini is now bundled into Workspace Business and Enterprise plans, which Google raised by a couple of dollars per month in 2025 (9to5Google). Both charge per person, so cost grows with headcount even if only a few people benefit.

Copilot, Gemini and Custom Agents Compared

Microsoft Copilot

Google Gemini

Custom AI agent

Setup effort

Low to start; real work is permissions cleanup and training

Low inside Workspace; connectors need configuration

Highest: scoping, build, testing, integration

Data access

Microsoft 365 content the user can open, plus connectors

Workspace content the user can open, plus connectors in Gemini Enterprise

Exactly the systems and records the job needs, through its own account

Control

Admin settings and agents the vendor allows

Admin settings and vendor features

Your rules, approval steps and logs; you own the code

Cost shape

Per user, per month, plus a base license

Per user, bundled into the Workspace price

Build cost, then hosting and model usage; not tied to headcount

Best for

Drafting, summaries and search across a Microsoft office

The same in a Google office

A repeatable process across several systems

Many firms end up with both. The assistant helps individuals with writing and search, and a custom agent runs the one process that costs the most hours.

When a Custom AI Agent Makes Sense

A custom agent earns its cost when the work is repetitive, spans systems that are not all Microsoft or Google, and has rules you can write down. Think of order entry from emailed PDFs, claims intake, invoice chasing or support triage tied to your own order data. It also fits when you need an audit trail, or when the agent must act at night without anyone opening a document.

Build only with controls in place. Give the agent its own login and the least access it needs (see our guide to AI agent access control), and keep a person approving anything involving money or customers (see human-in-the-loop AI). Then start small. A four to six week pilot with a number you can measure is easier to judge than a company-wide rollout; our guide to scoping an AI pilot shows how.

A Decision Checklist

  1. Is the task mostly writing, summarizing or searching documents and mail? Start with the assistant you already license.

  2. Does the work live in an ERP, CRM or industry system the assistant does not cover? Lean custom.

  3. Do you need the task done without a person starting it? Lean custom.

  4. Would only a handful of people use it? Per-seat pricing is hard to justify; look at usage first.

  5. Are your file permissions clean? If not, fix that before any assistant goes wide.

  6. Can you name one measurable result, such as hours per week or days to collect payment? If not, you are not ready to buy anything.

  7. Do you need logs and approvals you control? Lean custom.

  8. Could a pilot answer the question in six weeks? Run it before you commit.

How to Decide and Where to Start

Pick the process that costs your team the most hours, write down what it reads and changes, and test the assistant you already pay for against it. If it fits, you are done. If it keeps hitting the edge of one suite, you have a case for a custom agent, and a small pilot will prove it with your own numbers.

Our AI agent development team can review the process with you and tell you honestly which route fits. Book a call and bring your most time-hungry workflow.

Common questions

Is a custom AI agent better than Microsoft Copilot?

Not in general. Copilot suits drafting, summarizing and searching Microsoft 365 content. A custom agent suits a repeatable process that spans several systems and needs your own rules, approvals and logs. Many firms use both.

What data can Copilot and Gemini see?

Microsoft says Copilot surfaces only data the user has permission to view, and Google says Gemini retrieves only Workspace content the user can open. Both can be connected to other systems with connectors, which need setup and upkeep.

How much do Copilot and Gemini cost?

As of October 2026, Microsoft Copilot Business is listed at $21 per user per month on top of a Microsoft 365 plan. A first-year promotion at $18 (annual commitment, existing Microsoft 365 Business customers) runs July 1 to December 31, 2026, and monthly billing is $25.20. Gemini is bundled into Google Workspace Business and Enterprise plans, whose prices rose by about a couple of dollars per user per month in 2025. Check current pricing before buying.

When should we build a custom AI agent instead?

When the work is repetitive, crosses systems such as an ERP or CRM, follows rules you can write down, needs an audit trail or must run without a person starting it. Prove it first with a small pilot with a measurable target.

Keep reading.

More on the same subjects, picked by topic and tag.

Free project estimate

Book a call with a founder.

Discuss your software or AI project with a founder in a free 30-minute call. We will review your goal, identify the main cost drivers and agree what to scope next.

What you get back

  • An indicative budget range
  • A realistic timeline
  • What moves the cost up or down
  • A reply within 4 business hours
  • Answered by a founder
  • NDA on request, before you share details

Tell us what you need

A few lines is plenty. You get a number before any call.

Fields marked * are required.