Microsoft’s 35.8% problem with Copilot is a category problem
Why paid Copilot stalled at 4.4% of the M365 base and what every CIO running Salesforce, ServiceNow, Oracle, SAP, or Workday should do about it.
On May 5, Tom Warren from The Verge posted on X what every Microsoft customer was already thinking but wouldn’t say:
Tom was reacting to a deleted post by Jacob Andreou, Microsoft’s new EVP of Copilot. The wider context: Windows and Apps teams (Notepad, Snipping Tool) had quietly stripped Copilot branding from Windows 11 a few weeks earlier.
Also, a few days earlier, Satya Nadella mentioned that “the seat-based pricing is just entitlement to some consumption”.
It might look like Microsoft’s AI bet is wobbling.
However, the correct reading is probably that Microsoft is the first enterprise software vendor-turned-hyperscaler publicly admitting that prompt-as-a-product has a structural ceiling and that the next phase of enterprise AI value will not be unlocked by Copilot in every product, but by agents inside every process.
The AI adoption numbers that should be on every executive dashboard in 2026 Q2
Microsoft now has over 20 million paid Microsoft 365 Copilot seats, and that would sound fantastic if not the fact that over 450 million people use Microsoft M365 every month. That means that only 4.4% of them use paid Copilot. Three years after huge launch and tens of billions of capex spent, fewer than 1/20 Microsoft 365 commercial users pays for Copilot.
According to Recon Analytics' survey of 150,000+ U.S. respondents published in the report titled “AI Choice 2026: Why Licenses Don’t Equal Adoption”1 Microsoft Copilot's workplace conversion rate stands at just 35.8%, while ChatGPT converts at 83.1% workplace usage. Gemini is at 34%. A 47-point gap, so whatever is going wrong did not happen in the model layer.
Of users who tried Copilot and stopped, 44.2% cite distrust of Copilot’s answers as the primary reason - higher than ChatGPT (40.6%) or Gemini (42.8%). Recon’s accuracy NPS for Copilot deteriorated from −3.5 in July 2025 to −24.1 in September, recovering only partially to −19.8 in January 2026.
So, the natural question is why is M365 Copilot so poorly perceived, considering it’s essentially ChatGPT (OpenAI’s models) integrated into an enterprise product? It’s the same model but there is huge difference - ChatGPT is used by knowledge workers on a blank page where intent is declared in natural language. Microsoft Copilot meets that same worker in every possible place and every possible app, from a SharePoint site built in 2008, an Excel file with seven hidden tabs, in a Word document with suggestion to sum up a document not yet written, to a CRM form designed in 2013.
I took part in two workshops in April for clients who planned to resign from Copilot (several hundred users each), quoting “lack of usability” as the main reason. We ran a series of meetings during which we showed what Copilot Cowork can do and how Copilot custom agents could be embedded into processes, rather than forcing users to go to Copilot chat windows. It worked. The customers stayed and engaged us in further process redesign to increase Microsoft Copilot adoption. It’s twelve months after they bought hundreds of licenses from a “licensing partner” disguised as an “AI company”, during agreement renewal. A bit late.
The 35.8% problem is a category problem, not a Microsoft problem
It is a process problem masquerading as a Microsoft problem. Salesforce has the same problem, ServiceNow has it too. So does HubSpot, Oracle, SAP and Workday - you name them. Every system-of-record vendor is now shipping agents over data models designed before the agents existed. They and their customers will quickly discover that the agent’s adoption ceiling is set by the data model and the business workflow, not by the LLM.
Klarna learned the dark version of this lesson the expensive way in May 2025, when its CEO, Sebastian Siemiątkowski, publicly walked back the full AI customer-service replacement and started hiring humans back: “investing in the quality of human support is the way of the future for us.” The lesson is not that AI failed.
The lesson is that AI cannot rescue a process that was never redesigned for it.
Companies that will win value from agentic AI in the next 24 months are the ones that stop treating AI as a feature and start treating it as a forcing function for redesign. The ones that don’t will be running M365 Copilot pilots, wondering why their 35.8% never moved.
Why no amount of prompt engineering will close the 47-point gap
Here is the thing nobody wants to say out loud at AI conferences:
enterprise work is not prompt-shaped - it is process-shaped.
A typical B2B seller does not want to ask questions in a chat window - he/she has a 16-step renewal workflow embedded in a CRM and spanning through Teams and SharePoint lists.
A claims adjuster does not have a question - he/she has a regulator-defined sequence with mandatory legal capture and a hard SLA.
A contact-center agent does not have a question/response need - he/she has a script, a compliance checkpoint, and a quality scorecard.
The blank page that makes ChatGPT magical for individuals is exactly what makes Copilot frustrating for enterprise users.
For an agent to create value at work, several things have to be true at once:
An intent must be declarable in natural language by a human or another agent
The agent needs a permissions-aware environment because it must act in the system of record, with the right scope, on behalf of the right identity
There has to be a deterministic backbone behind the stochastic narrative
This is exactly the architecture that vendors must start converging on. At TrailblazerDX (TDX) 2026 in April, Salesforce open-sourced Agent Script, an agent definition language built as Madhav Thattai, COO of Salesforce AI, had explained - precisely because LLM reasoning wavers as workflows get more complex (an interview at Salesforce Ben from January 2026 titled “Is Salesforce Losing Confidence in LLMs?”) 2
Decagon ships the same idea under the label Agent Operating Procedures.
The pattern across both: LLMs handle communication and intent; deterministic code handles execution and guardrails.
Working with AI is a different shape of work
Capturing voice straight into Dataverse or Salesforce CRM instead of asking a seller to open a form is not a feature - it is a different shape of work. Microsoft’s latest natural voice features, including Hands‑free note capture in the Sales agent in Outlook mobile might be the first official acknowledgment from Redmond that the form itself is the bottleneck.
The customers who insist on preserving the existing form of work will get a beautiful demo and a stalled rollout. The difference between a pilot that converts and one that dies is almost never the model. It is whether the customer was willing to redesign the underlying process.
Sangeet Paul Choudary made the architectural point cleaner than I can in his February 2026 HBR article3:
“AI’s greatest economic impact will come not from automating tasks but from dramatically lowering the ‘translation’ costs that keep teams, tools, and data from working together.”
In an AI-adopting company, the person is the node and AI is the tool. In an AI-native company, the system itself is the node, and work gets redistributed to wherever intelligence, human or artificial, is more effective.
Sarah Wang at a16z, in Big Ideas 20264, calls the same shift the moment:
“the traditional system of record slips into the background as a commodity persistence tier — its strategic leverage ceded to whoever controls the intelligent execution environment employees actually use.”
Microsoft is reorganizing to better align with industry needs (and to catch up?)
On March 17, Mustafa Suleyman moved off day-to-day Copilot product leadership to focus on what Microsoft now labels “Superintelligence.” Jacob Andreou, formerly SVP at Snap, then CVP of Product and Growth at Microsoft AI, was promoted to EVP of Copilot, reporting directly to Satya Nadella.5
Exactly two weeks later, on March 31, Mary Jo Foley wrote her GeekWire post, “What the heck is going on with Microsoft lately?”, which was a polite way of asking the question everybody from the industry was already asking.
Then, on April 27 the Microsoft / OpenAI agreement was restructured:
non-exclusive license through 2032, AGI clause removed
OpenAI free to ship in any cloud, Microsoft’s revenue share to OpenAI ended
OpenAI’s share to Microsoft capped through 2030
The default media quote is that Microsoft is “losing the AI race”. There's also a reading where this is retreat, not strategy. I think the strategic reading is closer to the truth, although both are partially true. I am in the camp that Microsoft is resetting the value model from Copilot in every app to agents in every process. And it is doing it in front of us on the living organism, after M365 Copilot adoption sucked billions and did not bring significant market penetration.
The questions that we need to answer in 2026
The HFS Research and Genpact study released the same day as Microsoft’s earnings call „Autonomy Requires Trust in AI“6 found that 33% of respondents identify business processes not ready for agentic integration as the leading barrier to scaling agentic AI. So, process readiness, not data, not governance, not talent, is the number 1 obstacle.
The important questions that come to my mind:
Which 3 of your top 10 business processes will not exist in their current form in 24 months?
Where is your system of record investing - in better forms, or in agent-mediated intent capture?
Who is the owner of each process you want to extend with agentic AI?
“AI Choice 2026: Why Licenses Don’t Equal Adoption”, Recon Analytics, February 2026, https://www.reconanalytics.com/ai-choice-2026-why-licenses-dont-equal-adoption/
“Is Salesforce Losing Confidence in LLMs?”, SF Ben, January 2026, https://www.salesforceben.com/is-salesforce-losing-confidence-in-llms/
“AI’s Big Payoff Is Coordination, Not Automation”, HBR, February 2026, https://hbr.org/2026/02/ais-big-payoff-is-coordination-not-automation
“Big Ideas 2026: Part 1”, a16z New Media, December 2025, https://www.a16z.news/p/big-ideas-2026-part-1
“Microsoft Reorganizes Copilot Team, Names Jacob Andreou EVP Reporting To CEO Nadella”, CRN, March 2026, https://www.crn.com/news/ai/2026/microsoft-reorganizes-copilot-team-names-jacob-andreou-evp-reporting-to-ceo-nadella
“Autonomy requires trust in AI”, Genpact, April 2026, https://www.genpact.com/insight/autonomy-requires-trust-in-ai