Back to Explore
NewsVibeFix editorialUpdated Nov 18, 2026

Microsoft and Salesforce Shake Hands: CRM Moves Into Copilot, but Only One Piece Is Ready Today

On 17 November 2026 Microsoft and Salesforce announced Copilotforce at Ignite: Salesforce customer data and sales skills move into Copilot, Teams and Fabric. The plugin is in public preview with 38+ skills, but only Data Federation is generally available — the rest is preview or pilot.

Four professionals gathered around a laptop in a modern office

In short: old rivals shake hands — but check which pieces you can actually use today

On 17 November 2026, Microsoft and Salesforce announced a partnership called Copilotforce at Microsoft Ignite: Salesforce customer data, sales workflows and business actions are moving into Microsoft Copilot, Teams and Fabric. It sounds like a line from a press release, but there is a real fact inside it. Salesforce in Copilot, a plugin, opened in public preview the same day with more than 38 prebuilt sales skills. At the same time, only half of the data connection between Salesforce Data 360 and Microsoft Fabric is actually ready to use. This article is about that gap: of the four things announced, only one is generally available today. The rest are a public preview, another public preview, and a pilot. Get that straight first, and then we can talk about why this matters.

Some background first. Microsoft and Salesforce are the oldest rivals in enterprise software: one owns the most widely used office suite in the world, the other the most widely used customer relationship system. For a decade they fought over the same question — which application employees actually live in all day. Copilotforce answers: stop fighting, users live in both. Salesforce COO Miguel Milano put it plainly in the announcement, saying the deal brings together the number one CRM and the number one productivity surface. Judson Althoff, CEO of Microsoft Commercial Business, said AI is only as powerful as the data and context that grounds it. In plain language: Copilot was missing customer data, Salesforce was missing office context, and this time each side gets what it needs.

What actually moved into Copilot: a plugin, a set of skills, and two data pipes

The partnership has four parts, and they should not be blurred into one. The first is Salesforce in Copilot, a plugin that entered public preview on 17 November and is listed in AgentExchange and the Microsoft Copilot plugin directory. It ships with more than 38 prebuilt sales skills, named after the daily work of a salesperson: Pipeline Review, Close Planning, Deal Review, Customer Health, Renewal Radar and Call Follow-Up. Salesforce's own announcement adds three everyday workflows: meeting preparation, which combines account history and meeting transcripts into a single executive brief; deal-health analysis, which surfaces hidden risks and customer sentiment before a deal stalls; and pipeline review and action plans, which turn complex pipeline data into ready-to-execute close plans inside Word and PowerPoint. In other words, Copilot stops being just an assistant that writes your emails and starts being able to read your customer ledger — and act on it.

The second part is Sales Cloud in Teams, still in pilot with enterprise customers, with an open beta coming later in 2026. It does something very specific: sellers can see pipeline and quota progress, get deal notifications and find briefing materials without leaving Teams; interactive adaptive cards let them update supported opportunity details, with changes saved back to Salesforce; and they can share secure records in deal rooms to align cross-functional teams. Salesforce's announcement gives a number that explains this step: Teams has more than 400 million active users. Salesforce is not opening another window inside Teams. It is putting the CRM into the chat room where the sales team already spends its day.

The third and fourth parts are two data connections, both "zero copy" links between Salesforce Data 360 and Microsoft Fabric — meaning teams can query and analyse data across both platforms without building pipelines or maintaining duplicate datasets. The third, Data Federation, is generally available now: customers can directly query and analyse supported Microsoft Fabric OneLake data from inside Salesforce Data 360. A service agent can see a customer's recent purchases, a marketer can build a segment from spending behaviour, and an AI agent can answer from current data, all without moving the source data out of Fabric. The fourth, Salesforce Data 360 Data Sharing, runs in the other direction, sharing curated Data 360 customer intelligence into OneLake for analytics, reporting and AI workloads. It is still in public preview. Salesforce's Informatica unit rides along too, with a new plugin for Copilot that lets business users discover relevant data, check data quality and verify how trustworthy the underlying data is, using natural-language queries.

The technical foundation holding the four parts together is what Salesforce calls AIforce: a live interface layer that brings Salesforce capabilities into external AI environments. Microsoft's counterpart is Work IQ, the intelligence layer that powers Copilot and pulls context from Teams meetings, Outlook emails and documents across SharePoint and OneDrive. Put together, the picture is this: Work IQ knows what the customer complained about in a meeting and what was promised in an email, Salesforce knows the customer's status in the system and which sales stage they are in, and combined, the seller sees not two disconnected records but one answer to the question "what should happen next?"

Bar chart: Accenture has more than 740,000 Copilot users and 55,000 Salesforce users; Copilotforce is being validated first inside its own global sales organisation

Permissions are the real story here

If you only read the feature list, Copilotforce looks like another routine alliance between giants. What deserves a closer look is how the two companies handle permissions — because that is why this kind of deal used to fail. Microsoft's announcement is specific: through authenticated Salesforce MCP access, sellers can securely retrieve customer information and take supported business actions within their existing Salesforce permissions. Salesforce continues to govern data access and business actions, while Microsoft applies its existing enterprise compliance controls to communications and documents. Supported writeback actions to the CRM require explicit user confirmation and respect existing validation rules. Salesforce's version says the same thing in different words: granular permissions, sharing rules and audit trails are preserved in both environments.

Translated: Copilot can look at your customer data and draft follow-up actions for you, but it cannot act beyond your authority. It can suggest changing an opportunity's status, but only with your explicit confirmation, and the change is logged and auditable. The most dangerous moment for an AI assistant is when it holds data from across systems and nobody can say clearly what it is allowed to touch. Copilotforce solves this by leaving the question of "who can do what" exactly where it already was — inside each company's existing permission system — instead of inventing a new one for the sake of a demo. That is also why two old rivals could agree to this at all: neither has to give up its own control plane. Microsoft does not touch Salesforce's permission model, Salesforce does not touch Microsoft's compliance rules, and each guards its own wall.

There is another lesson here for anyone building AI assistants and agents. Over the past year, most enterprise AI announcements have been about what the assistant can do. Copilotforce spends a large part of its announcement on what it cannot cross. For buyers, that is actually the more credible signal: a product willing to put its permission boundaries in the launch announcement has at least thought about what will happen on day one. Of course, there is still a distance between permission design on paper and permission in production — the details of MCP authentication, what the audit logs look like, and how cross-system permission conflicts get resolved are all left unanswered. Those answers will have to wait for feedback from the first pilot customers, and that is exactly what Accenture is for.

Accenture, the test bed: 740,000 against 55,000

Accenture plays two roles in this partnership: design and implementation partner, and the first customer — "customer zero" in the announcement's words. It is deploying the integrations inside its own global sales organisation and sharing what it learns to help clients get value sooner. Its scale gives the pilot real weight: it has scaled Copilot to more than 740,000 people and is one of the largest Salesforce users in the world, with 55,000 users. The ratio of 740,000 to 55,000 says something by itself — inside Accenture, far more people use Copilot than use Salesforce. Copilotforce is first and foremost for the people who spend all day in Copilot and Teams and rarely open Salesforce at all.

This arrangement is also a form of risk management. Instead of pushing the full package to every joint customer at once, Microsoft and Salesforce are letting Accenture run it through its own sales team first. Accenture CEO Julie Sweet says the combined strengths of the three companies will reinvent sales with AI, giving sellers more time with customers. Take that with a grain of salt: feedback from the pilot customer will only arrive over the coming months, and there are no public answers yet on how the plugin performs in public preview, how the Teams pilot handles real sync latency, or how stable Data Federation is under heavy query loads. The one thing that is certain now is that the test bed is built — and it is not small.

The 38 skills: what do they actually save a salesperson?

The number 38 means little by itself. What matters is that the skills cover the most grinding hours of a salesperson's day. Reconstruct a working day from the two announcements: in the morning, run Pipeline Review to see which opportunities deserve attention today; before a meeting, let it combine account history, past emails and meeting transcripts into a single brief instead of hunting through three systems; after the call, use Call Follow-Up to turn the conversation into follow-up drafts and CRM records; near the end of the quarter, let Renewal Radar flag renewals that may slip away and Customer Health tell you which customer's mood has already turned. The selling point is fewer switches between the CRM, the inbox, meeting notes and chat.

But here is a cold splash of water: saving "switches" is not the same as saving "judgement". One word keeps appearing in the announcements: "grounded" — outputs anchored in real customer records and conversation context, not pretty sentences invented by the model. The word appears so often precisely because both companies know what enterprise users fear most: a confident-looking deal analysis that cites an email that never existed, or mixes up two customers' numbers. The number of skills can be printed in a launch announcement; the trustworthiness of the output can only be proven by people trying it on real deals. The right way to use a public preview is to test it against an account you know well: hand it a familiar customer and check whether the brief it assembles contains factual errors; let it draft a follow-up email and check whether it misstates a commitment. A skill that fails that test is decoration, however long the menu is.

There is another tension worth noting, hidden inside Salesforce's own product line. For the past year Salesforce has been pushing Agentforce — its own agent platform, pitched as agents doing the work for the enterprise. This partnership puts Salesforce capabilities and skills inside Microsoft's Copilot, and even the Teams integration works alongside Agentforce Coworker. One company's agents running inside a rival's assistant would have been unthinkable five years ago. It signals a shift that is already under way: the agent race is moving from "whose assistant" to "whose assistant can reach the most business systems". Salesforce's choice is to let Microsoft own the entry point while keeping the execution of business actions and the judgement of permissions in its own hands. If that division of labour works, more enterprise software vendors will make the same choice: the platform owns the front door, each vendor owns its own actions.

Why now: the agent race is becoming a context race

Zoom out, and Copilotforce does not arrive out of nowhere. Microsoft has Copilot, Salesforce has Agentforce, and both spent the past year telling the story of agents entering the enterprise. But every version of that story runs into the same bottleneck: however smart an agent is, without business data it is an empty shell. Customer data lives in Salesforce, operational and financial data lives in Fabric, meetings and emails live in Microsoft 365 — no single vendor's assistant can see more than one corner of the puzzle. Copilotforce is essentially both companies admitting this: rather than each growing its own assistant, first stitch the data together.

This judgement also has practical value for people building their own projects. The old way of doing enterprise integration was to move data from one system to another — build pipelines, run syncs, tolerate delays and duplicates. Zero-copy connections represent a different path: the data stays put, the query goes to it. Data Federation being generally available shows this path already works between systems of Microsoft's and Salesforce's size. For developers building AI projects of their own, the pattern is worth remembering: your agent does not necessarily need to own the data. What it needs is a channel that is allowed to see the data, with permission rules that both ends recognise. MCP plays exactly that role here: authenticated access, not a master key.

Only one piece is ready today — do not let the launch event carry you away

Back to the gap from the start, which is also the most practical part of this story. Of the four components, only Data Federation is generally available. Salesforce in Copilot is in public preview, Data 360 Data Sharing is in public preview, and Sales Cloud in Teams is still in pilot, with an open beta coming later this year. Both companies say so themselves at the bottom of their announcements: pricing and packaging are subject to change, regional availability is governed by customer agreements, and customers should base purchasing decisions on capabilities that are commercially available today. In other words: do not budget against the full blueprint from the launch event. Plan against the one piece that works today.

If you are a joint customer of both products, the pragmatic approach has three steps. First, try the plugin in public preview from AgentExchange or the Copilot plugin directory, run it with one sales team for two weeks, and focus on whether its deal analysis and meeting briefs are actually accurate — not on how many skills it lists. Second, before counting on the Teams piece, wait for news of the open beta; there is no public data yet on how the pilot handles sync and permissions. Third, and easiest to overlook: check whether your own Salesforce permission model is clean. Copilot only magnifies the permissions that already exist, and whatever mess already lives in that model will only look worse once an AI assistant is using it.

Looking ahead, the real test comes next year. The Teams open beta is due later this year, there is no published timetable for the plugin to move from public preview to general availability, and no pricing or packaging has been announced — joint customers cannot even get a quote yet. There is also a question Microsoft will have to answer for itself: Copilot is Microsoft's flagship product, and the better Salesforce skills run inside it, the more Microsoft has to decide whether Copilot is its own assistant or an open platform anyone can build skills into. This time the answer leans towards the latter — the announcement stresses Copilot's "open and interoperable" platform. If that direction holds, Copilotforce is not a one-off alliance but a precedent, and more business systems will move into Copilot the same way. When that happens, today's argument about "is the plugin any good" will turn into a harder question: who writes the rules of the platform. That is the more difficult problem.

The real bet behind Copilotforce is not on any single feature. It is on a judgement: the next round of enterprise AI competition will not be won by the assistant that talks best, but by whoever can stitch the context scattered across systems into one trustworthy, permissioned picture. Microsoft and Salesforce have chosen to stitch it together rather than apart. If that step works out, salespeople will switch apps less often and see customers more often. If it does not, it will be just another loud name from Ignite. The difference will only be visible next year — in Accenture's pilot report, in how fast the plugin moves from preview to general availability, and above all in whether the first customers are willing to let an AI assistant press the confirmation button that writes back to the CRM. The simplest test for everyone else is this: a year from now, are Accenture's salespeople still using it? Pilots end. Habits do not lie.

Views 0Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
A developer types at a dual-monitor workstation; the left screen shows a dark code editor, the right a testing dashboard with green checkmarks and bar charts
News
Did Your Agent Get Better or Worse? AWS Open-Sources UAEF to Measure Agents from Every Framework with One Ruler

On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.

Developer WorkflowProduct LaunchTesting & Quality