AI Strategy / Enterprise
Roughly four out of five enterprises that bought Microsoft 365 Copilot are still in pilot. The licence was never the problem.
of Copilot pilots ever reach production
Four in five organisations bought licences and are still piloting. Only 3% say Copilot delivered significant ROI. 46% of users save up to about 14 minutes a day: real, modest, and nowhere near a business model change.
Gartner, M365 Copilot surveys 2024–2025
Inside Microsoft 365, it reads and writes well. These are genuine wins, and they are worth the licence for the right teams:
These are horizontal productivity gains. They are not process transformation, and no licence tier makes them into process transformation.
Copilot reasons over Microsoft Graph: your tenant content. Outside that perimeter it is structurally weaker, and six gaps account for most stalled rollouts.
Non-Microsoft data
Project files on network drives, scanned drawings, PDFs in a document system, ERP records, site photos. Not in the Graph, not in the answer.
End-to-end process
Copilot advises. It does not reliably act across systems: raise a purchase order, update a register, route an approval. That takes Copilot Studio and Power Platform work.
Domain reasoning
Contract clauses, tender evaluation, safety regulation, works specifications. No industry depth out of the box, and generic drafting is not review.
Cost at scale
Around US$30 per user per month on top of an existing licence, plus governance tooling. Across hundreds of seats that is a material line in the total cost of ownership.
Data residency
A live question under Indonesia's PDP Law and under most group data policies. It shapes what you may connect, not just how.
Measurement
“One use in 28 days” is the adoption metric that ships by default. It is not adoption, and it will not survive an audit of the business case.
Five pilots is not five wins. It is one pilot repeated five times.
Every serious option is a variation on one decision: extend the estate you already pay for, or add something beside it.
Option A
Extend what you already own
Copilot Studio, Azure AI Foundry and Power Platform. Agents grounded on your document system and ERP data, with Power Automate carrying out the actions.
Best when you must stay inside the existing security, compliance and procurement perimeter.
Cost scarce Copilot Studio engineering skill, which is usually the reason it stalls.
Option B
Vendor-neutral orchestration
Workflow platforms that connect to any data source regardless of vendor, so a process spanning several systems can be automated whole.
Best when the workflow genuinely crosses non-Microsoft systems.
Cost a new vendor, a new security review and a new residency question.
Option C
One bespoke agent
A single high-value process built end to end and deployed inside your own cloud. Nothing wasted on processes you were not going to automate anyway.
Best when there is one process with measurable ROI and a named owner.
Cost build and maintenance burden. Somebody has to own it after launch.
Take Option A, with Option C's patterns where Copilot Studio is too limited.
Extend the estate you already pay for. Do not rip it out and start again. The failure mode is not the wrong platform; it is a platform decision made before anyone measured where the work actually goes.
Three to four weeks, one department, and a decision at the end that you can defend. That is the whole scope.
What you get is evidence behind a decision. Not a slide deck, and not a platform recommendation looking for a problem.
The component parts are already running in production somewhere. Grounded agents with per-document access control and a full per-action audit trail. Multilingual engines that switch language mid-conversation. Verification-first pipelines where every output must validate against a source of truth before it is released.
What is missing is your data, not the machinery. A large integrator will staff a discovery phase and a governance workstream. A small one has to show you the working system, because there is nothing else to show.
A free 20-minute Deployment Readiness Check. No deck and no pitch: we look at where your data sits and tell you plainly whether a custom agent is worth building yet.
Book the checkOr write to hello@askunclejifu.com
Gartner, Microsoft 365 Copilot adoption surveys (2024, 2025): 80% piloting or planning, 16% of pilots reaching production, 3% reporting significant ROI, 46% of users saving up to ~14 minutes per day. Forrester, “The Copilot Reality Check”, February 2026: governance and data hygiene decide who scales.
Microsoft-commissioned Forrester TEI studies report large modelled ROI figures. They are vendor-funded and model productivity time, not process transformation. Read them as such.
Copilot pricing as published by Microsoft. Figures are as at the date of writing and should be re-verified before any business case.