Why 400,000 Copilot Agents Still Do Not Know Your Company

In its January 2025 earnings call, Microsoft dropped a number that sounded like the future had arrived: more than 160,000 organisations had used Copilot Studio to build over 400,000 custom agents in a single quarter, more than double the quarter before¹. By late 2025, the company put the running total past one million custom agents across SharePoint and Copilot Studio¹¹. The agent era, it seemed, was already here.

And yet ask the operations lead at a mid-sized German manufacturer whether their Copilot knows how their company codes an invoice, which customer always needs a second delivery note, or why the Tuesday shipment to Lyon is handled differently, and the honest answer is no. It can search the files. It can draft an email. It cannot run the work. Four hundred thousand agents, and almost none of them hold a real memory of the business they serve.

This piece is for the CTO, operations leader, or Geschaeftsfuehrer who bought into the copilot promise and is quietly wondering why it plateaued. The gap is not the model. It is the difference between grounding on documents and building a Company Brain - a living memory of how your company actually works, on top of which AI employees take over the routine. Here is what separates the two, and why one keeps getting better while the other keeps starting over.

The 400,000-Agent Illusion

The headline numbers are real, and they are impressive. What they measure is how easy it has become to create an agent, not how much any agent understands. A Copilot Studio agent can be spun up in an afternoon by pointing it at a SharePoint site. That is a feature. It is also the reason so many of these agents are interchangeable and shallow.

• Volume is not depth - 400,000 custom agents built in one quarter across 160,000 organisations averages fewer than three agents per organisation, most of them thin document wrappers¹.
• Seats do not equal outcomes - Microsoft passed 20 million paid Copilot seats by 2026¹², yet the same year an internal push nicknamed “code red” formed around lagging real-world adoption and impact²⁴.
• Pilots stall - An MIT study of over 300 deployments found roughly 95 percent of enterprise generative-AI pilots delivered no measurable financial return⁵.
• The investment is enormous - That 95 percent sits on top of an estimated $30 to $40 billion in enterprise AI spending, most of it aimed at sales and marketing rather than the back-office work where returns are highest⁵.
• Ease of creation breeds sprawl - The very thing that produced 400,000 agents now produces an ungoverned mess that Gartner has a name for: agent sprawl¹⁰.

The illusion is that a large count of agents means a company has adopted AI. What it usually means is that a company has a large count of stateless helpers, each answering questions about documents, none of them running the work. To see why, you have to understand what these tools actually do under the hood.

What Grounding Actually Does (and Does Not Do)

Both Microsoft Copilot and Salesforce Agentforce lean on the same core technique: grounding, also called retrieval-augmented generation. It is genuinely useful, and it is also the source of the ceiling. Understanding it precisely is the whole point.

How grounding works

When you ask Copilot a question, it does not answer from a private model of your company. It runs a search over content you are allowed to see, retrieves the most relevant snippets, and passes them to the language model alongside your prompt for that single response². Salesforce Agentforce does the same over Data Cloud, using RAG and hybrid search to pull the right record or knowledge article at the moment of the query⁷.

• Retrieval at query time - The system fetches relevant files, emails, records, or knowledge articles when you ask, not before².
• Permission-trimmed - Copilot only grounds in content you already have access to, respecting existing Microsoft 365 permissions⁴.
• Stateless by design - The retrieved context lives for one response. It is not stored as a growing model of your business².
• No training on your data - Microsoft states plainly that your tenant data is not used to train the foundation models, and your prompts do not improve them³.
• Only as good as the documents - Grounding inherits the quality of what it retrieves. If the process lives in someone’s head and not in a file, grounding cannot find it²³.

What grounding is great at

• Search and summarise - Find the contract clause, summarise the thread, pull the latest numbers from a deck.
• First-draft generation - Draft an email, a slide, a status update based on existing material.
• Personal productivity - Help one person move faster through their own inbox and files²¹.
• Q&A over a known corpus - Answer FAQ-style questions when the answer genuinely exists in a document.

Where grounding hits its ceiling

• Tacit process knowledge - The unwritten rules, the “we always do it this way for this client” that lives in people, not files.
• Learning from corrections - When you fix its output, grounding does not remember the fix. Tomorrow it makes the same mistake⁵.
• End-to-end execution - Grounding informs an answer; it does not own a multi-step process across systems and days.
• Accumulated context - It cannot build a picture of a customer, a supplier, or a workflow over time because it holds nothing between sessions¹⁴.

None of this makes grounding bad. It makes it a read layer. The problem starts when leaders expect a read layer to behave like a colleague who remembers.

Copilot vs Company Brain: The Core Difference

A generic copilot and a Company Brain can look similar in a demo. They diverge the moment you care about the same process running a hundred times, correctly, without a human re-explaining it each time. The difference is memory that persists and improves.

The clearest way to feel the difference is a single invoice. Ask a grounded copilot to code it and it will guess from the document. Ask an AI employee running on a Company Brain and it applies the coding rule your controller taught it last month, routes it to the approver who signs for that cost centre, and flags the one supplier whose credits always need a second look - because the brain remembers all three.

The Learning Gap: Why Doc-Grounding Plateaus

MIT gave the failure pattern a name, but you can watch it happen in any office. A team rolls out a copilot, sees a burst of enthusiasm, then slowly stops using it for anything that matters. The reason is structural, not cultural.

• The same mistake, forever - Because the model does not train on your data³, a correction you make today is gone tomorrow. Grounding re-reads the document; it never remembers your fix.
• Garbage-in stays garbage-in - Grounding inherits your document quality. Analysts consistently trace the 95 percent pilot-failure rate back to fragmented, poorly governed data²³.
• Context evaporates between sessions - Enterprise work is multi-step and multi-day. A stateless tool cannot hold the thread, so a human keeps carrying the institutional knowledge¹⁵.
• Capability is not the bottleneck - Researchers found that less capable agents embedded in a real workflow, with humans carrying context, outperformed a more capable agent that had no way to access that context¹⁴.
• The plateau is predictable - Without a memory layer, usefulness caps out at “better search.” That is real value, but it is not transformation, and it is not what the budget was approved for⁵.

The fix is not a better model. Frontier models are already good enough. The fix is a place for the company’s know-how to accumulate and improve - which is what a Company Brain is.

What a Company Brain Actually Is

A Company Brain is not a chatbot and not a bigger document store. It is a persistent, structured memory of how your company works, built from three things a generic copilot never keeps: people-knowledge, processes, and outcomes. It sits on top of your existing systems and grows with the work.

The three layers it captures

• People-knowledge - The tacit expertise your team carries: the exceptions, the judgement calls, the “ask Maria about the Austrian orders.” The brain captures it as it surfaces in real work.
• Processes - The actual sequence of steps for each routine task, including the branches and workarounds that never made it into an SOP.
• Data and decisions - The records from your CRM, ERP, and inboxes, plus the decisions made against them and, crucially, the corrections that followed.

What makes it a brain, not a database

• It writes, not just reads - RAG retrieves. A Company Brain also records outcomes and corrections, so it has a write layer, which is what turns retrieval into learning¹⁴.
• It connects - It links a customer email to the CRM record to the ERP order to the last three exceptions, so context is a graph, not a snippet.
• It is shared - One memory serves every AI employee and every use case, instead of 400,000 disconnected agents each starting cold¹⁰.
• It compounds - Agent-memory research in 2026 shows the biggest gains come from handling exactly this: histories where facts accumulate, change, and relate over time¹⁶.
• It is governed - Because memory is centralised, access control, audit, and data residency apply once, to one layer, not to a swarm of bots.

If you want the deeper mechanics of how persistent memory differs from retrieval, we go into it in AI Agent Memory. The short version: a Company Brain is the thing that makes an AI employee worth trusting with a process, because it remembers how the process actually runs here.

AI Employees on Top of the Brain

A Company Brain on its own is memory. The value shows up when AI employees use that memory to do the work - connected to the same tools your team already lives in. This is where “answers a question” becomes “owns the process.”

Where they connect

• Email - Read incoming requests, draft and send replies grounded in the brain’s memory of that customer, escalate the ones that need a human.
• Teams - Take instructions in chat, post updates, ask a colleague when confidence is low.
• SharePoint - Use documents as one input, but write the durable knowledge back into the brain, not into another orphaned file.
• CRM (Salesforce, HubSpot) - Update records, log activity, prepare the next best action based on the full history, not a single retrieved snippet.
• ERP (SAP and others) - Create and match orders, code invoices, reconcile - the transactional core where mistakes are expensive.

What an AI employee owns end to end

• Accounts payable - Capture the invoice, apply the coding rules the brain learned, match to the PO, route to the right approver. We break down this exact flow in The AI Employee in Accounts Payable.
• IT and internal service desk - Triage tickets, resolve the known ones, and get better at routing as the brain learns your systems.
• Customer operations - Answer, update, and follow up using the customer’s full remembered context, not a cold lookup.
• Order and quote handling - Turn an email into a validated order or quote, applying the pricing exceptions the brain remembers.
• Onboarding - Give a new hire an AI colleague that already knows the process, which we cover in Onboarding AI Employees.

The Daily Feedback Loop That Grounding Cannot Have

The single biggest difference between a grounded copilot and a Company Brain is what happens after a mistake. For a copilot, nothing happens - the mistake is not remembered. For a Company Brain, the correction is the fuel. This is the mechanism, step by step.

1. The AI employee acts - It codes the invoice, drafts the reply, updates the record, using the brain’s current memory.
2. A human reviews the edge cases - Low-confidence actions are flagged for a person. The rest flow through with an audit log.
3. The correction is captured - When the controller re-codes a line or the rep rewrites a paragraph, the brain records what changed and why.
4. The correction becomes a rule - Next time the same pattern appears, the brain applies the learned rule instead of repeating the error.
5. The knowledge compounds - Over weeks, thousands of small corrections encode the exact know-how no document ever held.

• Corrections are cheap, retraining is not - No model retraining is required; the memory layer captures the rule directly¹⁶.
• The loop needs the work - This is why a Company Brain has to sit on top of real systems and real tasks, not off to the side.
• Confidence rises with use - As the rule base grows, more actions clear the confidence bar and fewer need review.
• It is measurable - Correction rate per hundred actions is a clean KPI that should fall month over month.

Agent Sprawl vs One Brain

The same ease of creation that produced 400,000 agents produces a governance problem when nobody is counting. Gartner calls it agent sprawl, and it is the predictable end state of a build-fast, remember-nothing model.

• Explosive growth - Gartner projects the average Fortune 500 company will run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025¹⁰.
• Almost no governance - Only 13 percent of organisations believe they have adequate agent governance in place¹⁰.
• Concrete risks - Sprawl exposes companies to misinformation, oversharing, and data loss as agents multiply without oversight¹⁰.
• Uniform control backfires - Gartner also warns that applying one blunt governance policy across every agent will itself cause failures; agents need context-appropriate control²².
• Duplication everywhere - Each team rebuilds the same shallow agent because none of them share a memory of the company.

A single governed brain is not just more capable than sprawl. It is dramatically easier to secure and to document under GDPR and the EU AI Act, which becomes fully applicable in August 2026²⁰. For a deeper take on keeping that memory under your control, see The Sovereign Company Brain.

The 90-Day Path From Copilot to Company Brain

You do not rip out Copilot to build a Company Brain. You pick one process that a grounded copilot cannot actually run, and you give it a memory. Here is the sequence that works.

Phase 1: Pick the process (Weeks 1-3)

1. Find a memory-hungry process - Choose a routine, high-volume task full of exceptions that a copilot keeps getting wrong: AP coding, order entry, ticket triage.
2. Map the real workflow - Sit with the people who run it. Capture the branches and exceptions that never reached an SOP.
3. Baseline the numbers - Time per item, error rate, rework. This is what the feedback loop will improve against.

Phase 2: Build the brain and the AI employee (Weeks 4-8)

4. Connect the systems - Wire the brain to email, Teams, SharePoint, CRM, and ERP for that process. No new platform for the team to learn.
5. Seed the memory - Load existing rules and records, then let the brain observe live work to capture the tacit parts.
6. Run in parallel - The AI employee shadows the process, humans review every action, and each correction lands in the brain.

Phase 3: Hand over the routine (Weeks 9-12)

7. Raise the autonomy - As the correction rate falls, let high-confidence actions flow through without review.
8. Measure against baseline - Show the time saved and the falling error rate to the people who approved the pilot.
9. Reuse the brain - Extend the same memory layer to the next process. This is where the copilot model cannot follow, because it never built the memory.

How Superkind Fits

Superkind builds the Company Brain and the AI employees that run on it. The starting point is never a generic product you adapt to - it is your processes, your systems, and the know-how your team already has. Not another off-the-shelf tool that only knows the internet.

• Company Brain as the core - We build a living memory of your people-knowledge, processes, and data, so AI employees act on how your company actually works, not on a retrieved snippet.
• AI employees, not chat windows - Purpose-built roles - AP clerk, service agent, order handler - that own routine work end to end, with humans on the exceptions.
• Sits on top of your stack - Connects to email, Teams, SharePoint, Salesforce, HubSpot, and SAP. No rip-and-replace, nothing new for the team to learn.
• Gets better every day - Every correction your team makes becomes a durable rule in the brain, so the same process runs sharper next week.
• Live in weeks - A first use case goes into production in weeks, not a six-month rollout, and the memory layer then extends to the next one.
• One governed layer - A single, access-controlled brain instead of hundreds of ungoverned agents, which is far simpler to secure and document.
• Runs in your infrastructure - Data stays inside your environment with encrypted connections, audit logs, and GDPR-ready controls.
• Outcomes, not licences - Pricing is tied to the process outcome and measurable ROI, not per-seat fees for a tool nobody adopts.

To be clear: keep your Copilot. It is good at what it is good at. Just stop asking it to run the processes it structurally cannot remember, and put a Company Brain underneath those. For the economics of that decision, see What a Company Brain Costs.

Decision Framework: Copilot, Company Brain, or Both

The choice is not either-or. Most companies should run both, matched to the job. Here is how to decide which tool a given need belongs to.

If you cannot yet name a process your copilot fails to run, you probably do not need a Company Brain today. The moment you can - and most operations leaders can name three - the tool for that job is memory, not more grounding.

Frequently Asked Questions

Henri Jung

Co-founder of Superkind, where he helps SMEs and enterprises deploy custom AI agents that actually fit how their teams work. Henri is passionate about closing the gap between what AI can do and the value it creates in real companies. He spends most of his time with operations leaders who tried a generic copilot, hit the plateau, and needed something that actually remembers how their company runs. He believes the Mittelstand has everything it needs to lead in AI - it just needs the right approach.