Copilot vs. a Real AI Assistant: What Microsoft Won't Tell You

Microsoft says 70% of Fortune 500 companies use Copilot. What they don’t mention: only 3.3% of Microsoft 365 users actually pay for it. That’s 15 million paid seats out of more than 450 million. The other 96.7% tried the free chat and moved on.

According to Gartner’s 2025 survey, 40% of IT leaders are still “piloting” Copilot. Only 5% have moved to full deployment. Nearly half of those who tried it rated it “some value, shows promise” — corporate-speak for “not worth $30 per user per month.”

The numbers tell a story Microsoft’s marketing team would prefer to skip.

What Copilot Actually Does Well

Credit where it’s due. Copilot has real strengths that no competitor can easily replicate.

It lives inside Word, Excel, PowerPoint, Outlook, and Teams. You don’t install a separate app or switch contexts. Ask it to summarize a meeting in Teams, draft a reply in Outlook, or generate a formula in Excel, and it pulls from your Microsoft Graph data — your files, your emails, your calendar.

For simple tasks — meeting summaries, email drafts, basic document generation — Copilot is competent. Not transformative, but competent. With 430 million M365 commercial seats worldwide, it has a distribution advantage that no startup can match.

If your workflow is “type prompt, get output, move on,” Copilot works fine.

Where Copilot Falls Short

1. The Amnesia Problem

Copilot does not remember you. Every session starts from zero.

Microsoft’s own documentation confirms it: “Each new chat starts without personal recall. Continuity comes from organizational data stored in Microsoft 365, not from learned user memory.” In practice, this means you re-explain your role, your preferences, and your context every single time you open a chat.

Copilot’s architecture is retrieval-first, not memory-first. It searches your Microsoft Graph for relevant documents and emails, then uses that as context. But it doesn’t build a model of who you are, what you care about, or how you work. It doesn’t know that you’re the CFO who always wants numbers in a specific format, or that “the Berlin project” refers to a specific client engagement from last quarter.

Microsoft did introduce a “Copilot Memory” feature. Users report that it’s inconsistent, frequently broken, and sometimes simply stops working even when toggled on. The Microsoft Q&A forums are full of paid subscribers asking why their memory feature doesn’t persist across sessions. One Copilot Pro subscriber wrote: “The overall behavior is as if memory were disabled, even though it appears active in the interface.”

A tool that forgets everything between conversations isn’t an assistant. It’s a chatbot with good branding. (This is the same knowledge drain problem that costs companies millions — except here, the AI is doing it to itself.)

2. One Size Fits Nobody

Copilot is a single generic model applied across every M365 application. The same AI that drafts your emails also writes your financial reports and summarizes your sales pipeline.

There’s no specialization. No deep expertise in any domain. It’s an intern who’s read everything on your SharePoint but hasn’t actually worked on any of your projects.

Microsoft offers Copilot Studio for building custom agents. In theory, you can create specialized assistants for different functions. In practice, the experience is rough. A brutally honest review from an MVP developer described it as “a platform of contradictions.” Connected agents can’t run their own tool servers. Version control between Teams users is broken. Agent lifecycle management produces “vague SQL errors.” Building enterprise-grade multi-agent systems on Copilot Studio requires workarounds on top of workarounds.

Real specialization means having dedicated AI agents for different domains — one that deeply understands your calendar and scheduling patterns, another that knows your email communication style, a third that specializes in research and document analysis. Not the same generic model wearing different hats.

3. Data Residency Theater

Microsoft completed its EU Data Boundary in February 2025. Customer data at rest stays within the EU. That sounds reassuring until you look at the details.

In-country data processing for Copilot — meaning your prompts and the AI’s responses are handled by servers within your country — is a different story. The first wave (end of 2025) covered Australia, the UK, India, and Japan. Germany, along with ten other countries, was pushed to 2026.

Until then, when a German employee types a prompt into Copilot, that prompt travels to wherever Microsoft’s LLM infrastructure happens to be. For companies subject to the GDPR with strict compliance requirements, “your data is stored in the EU but processed elsewhere” is not a comfortable answer. (We unpack the full regulatory picture in our article on GDPR and AI compliance.)

4. The Real Price

Copilot Enterprise costs $30 per user per month. That’s on top of your existing Microsoft 365 subscription. For a company of 200 employees, that’s $72,000 per year — for a tool that Gartner says most organizations can’t measure ROI on.

The headline price also doesn’t include Copilot Studio credits ($200 per pack of 25,000 credits), vertical SKUs for Sales and Service ($50 per user), or the change management effort required to get people to actually use it. One executive on LinkedIn described rolling out 4,000 Copilot seats at $1.4 million annually. Three months later, 47 people had opened it. Twelve had used it more than once.

What a Real AI Assistant Looks Like

These aren’t theoretical complaints. They point to a design question: what should an AI assistant actually do?

Start with memory. Your assistant should know who you are across sessions — your role, your preferences, your projects, your communication style. Not because it searched your SharePoint, but because it genuinely remembers working with you yesterday and last month. When you say “update the Berlin proposal,” it should know which document you mean without you pasting a link.

Then consider specialization. A dedicated email agent that understands your communication patterns is a fundamentally different tool from a research agent that synthesizes information across sources. One model wearing different hats is not the same as purpose-built specialists coordinated by a supervisor.

An assistant should also work proactively. It should prepare your morning briefing, triage your inbox before you open it, and flag the three things that actually need your attention today. Copilot sits idle until prompted. That’s a search engine with a chat interface.

And if you’re a German company, data residency means data residency. Not “EU data boundary with processing elsewhere.” Your prompts, your responses, your memory — within German infrastructure. Not promised for next year. Available now.

Making the Decision

Copilot makes sense in a specific scenario: you’re a large enterprise, fully committed to M365, with a change management team to drive adoption, and your needs don’t go beyond summarization and drafting.

Outside that profile, the math gets harder. If you need persistent memory, specialized agents, German data processing today, or you’re a mid-sized company watching $30/seat disappear into unused licenses — Copilot’s value proposition breaks down.

	Copilot	amaiko
Memory	Session-only, resets each chat	Persistent across all interactions
Architecture	Single generic model	Multi-agent specialists
Data Residency	Germany processing: 2026	German hosting: now
Pricing	$30/user/month + M365 license	EUR 24.99/user/month, all-in
Proactive Work	Waits to be prompted	Morning briefings, inbox triage
M365 Integration	Native (strongest advantage)	Teams-native, deep M365 access

The Bottom Line

Copilot is Microsoft’s answer to the question “how do we put AI in every M365 subscription?” It’s a competent tool for one-shot tasks inside Office apps. It is not a persistent, intelligent assistant that grows with you over time.

If your team needs more than a stateless prompt box — if you need an AI that remembers, specializes, and acts before you ask — the alternatives are worth a serious look. amaiko is one of them: Teams-native, persistent memory, multi-agent specialists, hosted in Germany, EUR 24.99 per user. No amnesia. No workarounds.