Updated 21 May 2026 — Microsoft has now enabled Anthropic Claude models by default inside Word, Excel and PowerPoint for eligible Microsoft 365 Copilot tenants, GPT-5.5 thinking and ChatGPT Images 2.0 have landed in Copilot, Microsoft 365 E7 (Frontier Suite) and Agent 365 reached general availability on 5 May 2026, the new Copilot Cowork cross-skill agentic experience went live on 5 May, and a three-year Microsoft 365 Copilot SKU is now available in CSP. The off-the-shelf box has grown again — so the "when do I need custom?" question deserves another refreshed answer.
It's one of the most common questions we're asked: "Should we just roll out Copilot (or Gemini in Workspace), or do we need something custom?" It's a reasonable question, and the answer is almost always "both, but for different things." Understanding which tool belongs where will save you significant money and avoid the frustration of deploying an expensive AI product that doesn't deliver on its promise.
For most of 2025 this was a clean binary — Copilot on one side, custom on the other. Twelve months on, that framing is breaking down. Three May 2026 changes pushed the boundaries further than April did:
Net effect: more workflows that used to require a custom build can now be composed inside Copilot Studio, Cowork or an Agent 365 agent. Custom is not obsolete — but the threshold has moved.
The major vendors — Microsoft (Copilot in M365), Google (Gemini in Workspace, formerly Duet AI), and a few others — have each launched productivity AI assistants. These are AI tools woven through existing office suites. They use your organisation's data as context, apply large language models, and help you draft documents, summarise meetings, generate presentations, and query data in natural language.
Pricing typically ranges from £15–30 per user per month (Microsoft 365 Copilot Premium is currently around £24.70 per user per month in the UK; Google Workspace Gemini sits in a similar band). As of May 2026, Microsoft 365 Copilot also lets you choose between OpenAI's GPT-5.5 thinking and Anthropic's Claude Opus 4.6 as the underlying model — and crucially, Claude is now on by default inside Word, Excel and PowerPoint for eligible tenants. The line between "choosing a model" and "choosing a platform" is now much blurrier than it was twelve months ago, and that matters when you're comparing it against a custom build.
From 15 April 2026, Microsoft restricted Copilot Chat inside Word, Excel, PowerPoint and OneNote for users on the free Copilot Chat tier. Full in-app Copilot now requires a paid Microsoft 365 Copilot (Premium) licence. If your AI strategy was "everyone gets Copilot Chat for free and we'll figure it out from there," that strategy stopped working last month. It's a useful forcing function: are you actually getting value from Copilot, or were you just hoping you would? If it's the latter, targeted custom tooling almost always looks better than scaling Premium to the whole team.
Less talked about, but strategically bigger, is what happened inside Copilot Studio in April. Microsoft took multi-agent orchestration to General Availability alongside the open Agent-to-Agent (A2A) protocol, integration with Microsoft Fabric and the Microsoft 365 Agents SDK, and a meaningful expansion of governance controls, prompt authoring and model choice. Power Platform's April 2026 feature update layers on top, putting more AI capability directly into Power Apps and Power Automate flows.
In practical terms, the off-the-shelf box was already bigger than it was at the start of the year — and the May 2026 changes (Claude by default, E7/Agent 365 GA, Copilot Cowork) expanded it again. Workflows that used to require a custom build can increasingly be composed inside Copilot Studio — especially if your data lives in Microsoft Graph, Dataverse or Fabric. That doesn't make custom obsolete; it shifts the line. The new question to ask is "can Copilot Studio + an agent do this well enough?" before you ask "do we need to build something ourselves?"
For knowledge workers doing varied desk-based work, a well-deployed productivity AI can genuinely save 30–60 minutes a day — if they actually use it. Industry research (Mole Valley Chamber, British Chambers of Commerce, 2026) suggests only around 11% of UK SMEs use AI extensively for automation, and just 16% of UK businesses use AI in any meaningful way versus 88% globally — so the gap between "deployed" and "used" remains the real story.
It can't integrate with systems outside its ecosystem. If your business runs on a CRM, an ERP, a bespoke database, or any system that isn't part of the platform, the AI can't see it, can't read from it, and can't write to it — unless you build a connector, at which point you're doing custom work anyway. Cowork narrows this gap for the Microsoft estate, but the world outside the Graph is still a custom problem.
You can't control its behaviour with the precision custom gives you. Even with Copilot Studio's expanded governance controls and the new Cowork/A2A model, you still can't match the level of enforcement you get in a purpose-built tool. Having Claude or GPT inside Copilot Premium changes the tone and reasoning quality of outputs — but it doesn't change what Copilot can do, which is still bounded by the Graph and the apps it sits inside.
The outputs aren't auditable in the way compliance requires. For regulated industries, these tools' general-purpose nature makes auditability difficult — even with the new multi-agent audit logging, which is enterprise-grade rather than industry-specific. The Claude-by-default change also introduces a new sub-processor relationship that needs documenting in your records of processing.
It won't automate end-to-end processes at scale. Copilot Studio agents and Cowork help individuals and small teams work faster, and Agent 365 raises the ceiling. But none of them yet replace a workflow that needs to run reliably, on a schedule, against your CRM or finance system, for the next three years.
For most of 2025, this piece lived as a binary — Copilot on one side, custom builds on the other. In May 2026 that's no longer accurate. A third category has firmly established itself between the two: agentic AI products like Microsoft Copilot Cowork (GA 5 May 2026), Microsoft Agent 365 (also GA 5 May 2026, bundled in the new E7 Frontier Suite), and Anthropic's standalone Claude Cowork (now GA on macOS and Windows with OpenTelemetry monitoring and role-based access controls).
These aren't productivity suites, and they aren't bespoke builds. They're configurable agents that execute multi-step tasks across your tools with human checkpoints. They're more capable than traditional Copilot, more flexible than a custom single-purpose tool, and typically cheaper than commissioning one. For SMEs deciding what to do next, they belong in the conversation alongside "subscribe to Copilot" and "build something custom" — particularly for cross-app coordination work that used to be custom-only.
The honest framing: off-the-shelf suits individual productivity, agentic AI suits delegating multi-step work within your existing stack, and custom still wins when the process is specific, high-volume, regulated, or requires deep integration with systems outside Microsoft or Anthropic's reach.
| Scenario | Off-the-shelf | Custom AI |
|---|---|---|
| Summarise a meeting | ✓ Off-the-shelf — native integration | Overkill |
| Process 200 invoices/month | Cannot do this end-to-end | ✓ Custom — built once, runs automatically |
| Draft a client proposal | ✓ Off-the-shelf — Claude-by-default in Word is genuinely strong here | Possible, but unnecessary |
| Answer customer queries from your knowledge base | Possible, but inconsistent | ✓ Custom — reliable, on-brand, auditable |
| Analyse CRM sales data | Can't access external systems | ✓ Custom — scheduled automated analysis |
| Onboard new employees | Partial — documents only | ✓ Custom — structured, consistent, connected |
| Prep for a client meeting | ✓ Off-the-shelf — Copilot “Prepare” surfaces prior context | Unnecessary unless CRM integration needed |
| Coordinate a new deal across Outlook, Teams, SharePoint, Excel | Agentic (✓ Copilot Cowork or Claude Cowork, both GA) | Custom only when deep CRM/ERP integration is required |
| Run a recurring agent across the Microsoft estate | ✓ Agent 365 (GA 5 May 2026, bundled in E7) | Custom only if you need behaviour outside the Graph |
For most UK SMEs, the optimal AI strategy combines all three. Roll out productivity AI to knowledge workers who will benefit from general productivity gains — and now that Claude is default-on inside Word, Excel and PowerPoint, the model debate inside Copilot has effectively been settled for you. You'll still need to decide which users genuinely warrant a paid M365 Copilot Premium licence rather than enabling everyone by default, and whether E7 (E5 + Copilot + Agent 365) makes sense for your team. Layer agentic tools (Copilot Cowork, Agent 365, or Claude Cowork) on top for the cross-app workflows that used to need custom glue code. Then identify your two or three highest-value, most repetitive specific processes and build custom AI tools for those.
The question we ask every client: "What does your team do every week that takes significant time, follows a consistent pattern, and produces a predictable output?" Those are your custom AI opportunities. Everything else is probably off-the-shelf — or, increasingly in May 2026, agentic — territory.
If you're weighing up whether a productivity AI subscription is worth it for your business, whether a Copilot Cowork or Agent 365 pilot makes sense now they're GA, or wondering what a custom tool for a specific process might cost and deliver, that's exactly the kind of conversation we have on a discovery call — and the AI Growth Lab gives you a way to try one custom automation against a real process before committing to anything bigger.
Book a 60-minute conversation — no sales pitch, just a clear assessment of where productivity AI makes sense and which processes could be custom-built.
Get in Touch