Walk into the IT meeting at any $20-100M brand running on Microsoft and you will hear the same sentence: "We bought Copilot. Almost nobody uses it." The licenses are provisioned. The launch email went out. The Slack channel got three reactions. Then the conversation moved on to a procurement question about hardware.

This is the central problem with Microsoft Copilot for business in 2026. It is the largest enterprise AI rollout in history, with billions in paid seats sitting on desks doing nothing. The technology works. The reason it sits idle is not the technology. It is everything around the technology: the missing connectors, the missing per-department playbooks, the missing context layer, the missing training, the missing accountability. The license is a starting line, not a finish line.

This is the implementation playbook for the mid-market operator who knows the licenses are not paying for themselves yet. It covers what is actually inside the Microsoft AI stack in 2026, how the pieces connect, what they cost at $20-100M revenue, how to decide between build and buy and partner, and how to run a 90-day rollout that ends with the licenses actually being used. It introduces the AiSolv SOLVE Framework, the five-stage method we use on every engagement.

The $50 billion problem nobody talks about

Microsoft has sold more than 70 million Copilot seats. At list price that is more than $25 billion in annual recurring revenue from one product line, growing fast. Independent surveys keep finding the same uncomfortable number: somewhere between 60 and 75 percent of paid seats see less than one meaningful interaction per week. Across the install base, that is tens of billions of dollars of software no one is opening.

The mid-market version of the story is the one we see in nearly every initial discovery call. Leadership signed a $30 per seat per month enterprise contract for 400 employees. That is $144,000 a year. The IT team rolled out the licenses, sent a launch email, scheduled an optional lunch-and-learn, and considered the work done. Three months later, the CFO asks why renewal is on the AP queue, and the only honest answer is that two people in marketing are using it daily, a handful of sales reps occasionally, and nobody else has logged in.

The reason this happens is structural, not human. Microsoft sold a productivity layer that only works when it has context. The context layer (connectors to your CRM, your help center, your call recordings, your SOPs, your product documentation, your tickets) does not exist by default. Without it, Copilot becomes a slower way to write an email. With it, Copilot becomes the operator's interface to every system in the company. The gap between those two outcomes is exactly the work that nobody scoped when the license was bought.

Mid-market is uniquely exposed here. Enterprise IT teams have dedicated AI Centers of Excellence handling rollout. SMBs do not have enough surface area for the gap to matter. The $20-100M operator sits in the worst middle: enough licenses to make the bill painful, not enough internal AI maturity to make them productive. This is the wedge the rest of this playbook addresses.

What you actually get with Microsoft AI

The Microsoft AI stack is six distinct products that share branding and overlap in confusing ways. Treat them as a stack with layers, not as a single product, and the architecture becomes obvious.

The six layers

  1. Microsoft 365 Copilot: the productivity layer that lives inside Word, Excel, Outlook, Teams, PowerPoint, and Loop. This is what most people mean when they say "Copilot."
  2. Copilot Studio: the customization layer where you build your own branded copilots and agents. This is where the real work happens for mid-market.
  3. Azure OpenAI / Azure AI Foundry: the build layer where developers call frontier models (GPT-4o, GPT-5, Claude through partner endpoints) inside your tenant with enterprise data controls.
  4. Microsoft Fabric: the data layer that unifies your data lake, warehouse, real-time analytics, and Power BI under one capacity SKU.
  5. Power Platform AI: the workflow layer covering Power Automate (workflows), Power Apps (lightweight apps), and the AI Builder service.
  6. Dynamics 365 AI: the CRM layer covering Copilot for Sales, Copilot for Service, Copilot for Finance, and Copilot for Customer Insights.

Tying all of this together is Microsoft IQ, Microsoft's freshly rebranded Graph plus knowledge platform. IQ is the most important thing in the stack and the least understood. We cover it as its own layer below because the entire stack stops being valuable the moment IQ is absent.

Most mid-market brands already pay for at least three of these six layers without realizing it. The E5 license bundle includes the Microsoft Graph data plane, the Azure tenant includes Azure OpenAI access (subject to approval), and Dynamics customers already have most of the Copilot for Sales surface. The question is rarely "what should we buy." It is "what are we already paying for that we have not turned on."

Microsoft 365 Copilot: the productivity layer

Microsoft 365 Copilot is the most visible part of the stack and the most often misjudged. It is a thin layer that wraps GPT-class models around Microsoft 365 documents and surfaces. By itself, it is a faster way to draft an email or summarize a meeting. Wired to your real corporate data, it becomes the natural language interface to your business.

What Copilot actually does today

Out of the box, M365 Copilot ships four high-leverage capabilities that pay for themselves at mid-market: meeting summaries and action item extraction inside Teams, ad-hoc Q&A across your own files and emails through Business Chat, draft generation inside Word and Outlook, and formula and pivot generation inside Excel. The fifth capability, slide generation in PowerPoint, is honestly still not good enough to ship to a client.

What it does not do without help is reach outside Microsoft 365. By default Copilot cannot see your Salesforce records, your Zendesk tickets, your Stripe events, your Notion or Confluence pages, your Klaviyo flows, or your help center articles. Without those connectors, every prompt that needs business context returns either a generic answer or a refusal. This is the gap most enterprise rollouts fail to scope.

How mid-market should approach the rollout

The pattern that works is to provision the license to a tightly scoped pilot group (10-30 power users across one functional team), wire the connectors that matter to that team, and only then expand. The pilot group should be marketing, ops, or finance, not engineering. Engineers are the worst pilot for Copilot because they have better tools already. The marketing manager who writes ten emails a day is the ideal user.

Pricing is $30 per user per month on the M365 Copilot SKU as of 2026, with a 300-seat minimum on the enterprise plan. There is a Copilot Pro consumer SKU at $20 per month that is not the same product. Make sure your procurement team is buying M365 Copilot, not Copilot Pro.

Copilot Studio: the customization layer

Copilot Studio is where Microsoft AI stops being a productivity SKU and starts being a platform. It is a low-code environment for building branded copilots and agents that live inside Teams, on a public website, inside a Power App, or as a standalone bot. For mid-market brands this is where the most leverage hides.

What you can build in Copilot Studio

The canonical use cases we see at $20-100M operators are: an internal "ask anything" assistant that answers HR, IT, and policy questions for employees, replacing a Confluence search that nobody uses; a customer support copilot that retrieves answers from the help center and the product documentation, deflecting 30-60 percent of ticket volume before a human ever sees it; a sales enablement copilot that lets reps ask "what are the latest pricing exceptions for healthcare," instead of pinging Slack; and a finance and procurement copilot that handles vendor questions, PO status, and AP follow-ups without escalating.

Each of these is a distinct copilot with its own scope, its own data sources, its own guardrails, and its own success metrics. The brands that get the most out of Copilot Studio treat each copilot as a product, with an owner, a roadmap, and weekly observability reviews. The brands that get the least treat them as a one-off chatbot project that ships and is forgotten.

Where Copilot Studio gets dangerous

The risk in Copilot Studio is the same risk we see in any low-code AI platform: it is easy to ship a bot that confidently says the wrong thing. The defenses are the same defenses we use on any production AI system. We always pair a pre-LLM classifier and post-LLM judge with the copilot, and we always wire it into an AI agent observability layer so we can see every prompt, every tool call, and every failure before the customer sees it.

Azure OpenAI and Azure AI Foundry: the build layer

When the use case outgrows Copilot Studio, the next layer up is Azure OpenAI (now consolidated under Azure AI Foundry as of late 2025). This is where developers build custom AI systems that run inside the customer's Azure tenant with enterprise-grade data governance, regional deployment, and the same security model as the rest of the Azure estate.

What you get with Azure OpenAI

Azure OpenAI gives you GPT-4o, GPT-5, the o-series reasoning models, embeddings, fine-tuning where it makes sense (which is rarely), Whisper for speech, and DALL-E for image generation. All of it runs under your Azure subscription, billed through your existing Microsoft enterprise agreement, with data residency you control, no consumer training on your prompts, and the same identity layer (Entra ID) as the rest of your stack.

For mid-market brands subject to regulatory constraints, this is non-negotiable. We have shipped Azure OpenAI as the underlying inference layer for clients in financial services, telehealth, and supplements specifically because the OpenAI consumer endpoint would have failed compliance review on day one. The same conversation applies to any operator considering HIPAA-compliant AI implementation: Azure OpenAI is usually the right answer for the inference layer.

The frontier vs fine-tuning question

The single most expensive mistake we see in this layer is the temptation to fine-tune. Vendors push it, internal data science teams push it, and the result is almost always six to nine months of work that gets beaten by a well-prompted frontier model in production. Our position is unambiguous: at mid-market, fine-tuning is rarely the right call. Use the frontier model, invest in retrieval and prompt engineering, and ship in six weeks instead of nine months. We pull the brake every time a fine-tuning conversation starts.

Microsoft Fabric: the data layer

Microsoft Fabric is the single most important and most misunderstood piece of the stack for mid-market. It is the unified data platform that combines what used to be Synapse, Data Factory, Power BI, Real-Time Analytics, OneLake, and now an ontology layer into one capacity-billed SKU. Without Fabric (or an equivalent unified data layer), every AI use case downstream is bottlenecked by data plumbing.

Fabric pricing for $20-100M brands

The pricing model is straightforward once you understand the capacity SKU concept. Capacity SKUs range from F2 (the smallest, $260 per month at pay-as-you-go) up to F2048. The "magic SKU" most mid-market operators land on is F64, which sits at roughly $8,400 per month at PAYG and $5,000 per month on a one-year reservation (a 41 percent discount). The F64 tier is the threshold at which you stop needing per-user Power BI Premium licenses, which is almost always where the math flips in favor of reserving capacity.

The pragmatic adoption pattern we recommend, and that Microsoft's own field engineers recommend in informal sessions, is to start at F2 with PAYG for hands-on experimentation, then activate the 60-day F64 trial when you are ready for a real stress test, then commit to a one-year F8 or F16 reservation once the workloads are stable. Most mid-market operators land between F16 and F32 in steady state unless they have a heavy Power BI footprint. Storage during pause is $23 per terabyte per month, so leaving capacity paused on weekends is genuinely cheap.

Why Fabric matters for AI specifically

Two things make Fabric the linchpin of the AI story. First, OneLake is the single source of truth that downstream Copilot, Copilot Studio, and custom Azure OpenAI apps all read from, which means you stop maintaining separate data layers for analytics and AI. Second, vectorization is included in the SKU cost. Anything that Fabric has purview over is automatically available as a vector index without separate Azure AI Search billing. For brands that previously paid $2,000 or more per month for a separate managed vector store, this alone often pays for the F64 reservation.

Power Platform AI: the workflow layer

Power Platform is the most overlooked piece of the Microsoft AI stack and, for mid-market operators, often the highest-ROI layer in the first 90 days. It contains three products, Power Automate (workflows), Power Apps (low-code apps), and AI Builder (the AI models you wire into both), that together replace most of what brands today are doing with N8N, Zapier, Make, or stitched-together internal scripts.

Why Power Automate beats the alternatives at mid-market

We have written separately about why we have moved off N8N for AI workflows on most production engagements. The same logic that pulls us off N8N pulls us toward Power Automate at brands already on the Microsoft estate. The tooling sits inside the same tenant as M365 and Dynamics, the auth model is Entra ID end-to-end, the DLP policies your security team already configured for SharePoint apply automatically, and the per-workflow pricing is bundled in most enterprise plans.

The use cases we ship most often: routing inbound support tickets through a classifier before they hit the human queue; turning every booked sales meeting into a structured CRM update with action items and next steps (this is the same pattern we cover in AI sales automation for revenue teams); flagging at-risk POs by joining ERP data with carrier reliability data; and auto-drafting the weekly executive summary from finance, sales, and marketing dashboards.

The trap to avoid

The trap is that Power Automate makes it too easy to ship flows with no observability. We mandate that every production Power Automate flow logs every run, every input, every output, and every failure to a central observability layer. The flows that operations teams trust are the ones they can audit, and the ones they can audit are the ones that get used.

Dynamics 365 AI: the CRM layer

Most $20-100M brands run on either Salesforce or Dynamics. Roughly half the brands we work with are on Dynamics already and do not realize how much AI capability is bundled into their existing seats. Copilot for Sales, Copilot for Service, Copilot for Finance, and Copilot for Customer Insights are not separate SKUs at most license tiers (though some advanced features require add-ons), and they slot directly into the workflows reps and CS teams already use.

The most under-used capability we see in this layer is the meeting-to-CRM update inside Copilot for Sales. The product captures the call, summarizes the discussion, drafts the follow-up, and writes structured fields back to the opportunity record. The vast majority of brands that have it turned on are using only the summary feature. The structured-write-back capability is what closes the loop on the rep-not-updating-the-CRM problem, and it is sitting unused under the same license.

For brands not on Dynamics, the same playbook applies through Copilot for Sales sitting on top of Salesforce. Microsoft and Salesforce have a working integration here and the rep-facing surface is similar enough that the rollout looks the same.

Microsoft IQ: the secret weapon

Microsoft IQ is the rebrand of what used to be called Microsoft Graph plus its newer knowledge components, and as of late 2025 it is now four parts: Work IQ, Foundry IQ, Fabric IQ, and Web IQ. Understanding IQ is the difference between a Copilot rollout that works and one that does not. We did not understand this until a Microsoft field session a few weeks ago laid it out plainly, and almost no mid-market operator we have talked to since understands it either.

The four IQs in plain English

Work IQ gives Copilot awareness of all your Microsoft 365 data: calendar, email, SharePoint, Teams, OneDrive, and so on. Critically, it honors the RBAC permissions of the caller. If a salesperson does not have access to a finance folder, Copilot will not surface that data to them either. This is the safety property that makes IQ deployable in regulated environments.

Foundry IQ is for external knowledge sources, the things that live outside Microsoft 365. Your help center, your product docs, your knowledge base, your Confluence, your Zendesk, your call recordings. Foundry IQ indexes and vectorizes them automatically using Azure AI Search under the hood as a managed vector store with re-ranking.

Fabric IQ is the integration with Microsoft Fabric and OneLake. It mirrors your operational data, including SQL managed instances, into the data lake in real time and adds an ontology layer that lets different departments use their own terminology for the same data. Marketing can ask about "leads," finance can ask about "opportunities," and the ontology resolves both to the right underlying table.

Web IQ is the newest addition, adding live web search to the contextual layer so agents can pull current information into responses without leaving the platform.

Why IQ is the linchpin

Every part of the Microsoft AI stack above (Copilot, Copilot Studio, Azure OpenAI apps, Power Automate flows, Dynamics copilots) reads from IQ. Configure IQ well and every product downstream gets better in lockstep. Skip IQ and you ship a copilot that hallucinates because it has no real context to retrieve from. This is the architectural decision that quietly determines whether your Microsoft AI rollout succeeds or sits idle.

Pricing math for $20-100M brands

The number we hear most often from CFOs at this revenue band is "we expected Microsoft AI to be expensive but we did not expect the bill to be this lumpy." The lumpiness comes from the way the SKUs stack. Here is the realistic 2026 cost profile for a 200-person, $50M revenue brand that wants the full stack live, mid-market sized.

Annual run-rate, modeled

  • M365 Copilot: 60 power users at $30 per user per month = $21,600 per year. (You do not need 200 seats. You need the seats for the people who will actually use it.)
  • Copilot Studio: $200 per tenant per month for the messaging surface plus message packs scaled to volume. Budget $10,000-$25,000 per year depending on volume.
  • Azure OpenAI: consumption-based. For a brand running one customer support copilot, one internal assistant, and one sales copilot, expect $1,500-$5,000 per month, or $18,000-$60,000 per year.
  • Microsoft Fabric: F16 on one-year reservation, roughly $1,250 per month or $15,000 per year. (Storage and overage tracked separately, usually adds 10-20 percent.)
  • Power Platform: bundled in most E5 plans. AI Builder add-on capacity adds $500-$2,000 per month at this size.
  • Dynamics 365 AI: most features included in existing Dynamics seats. Premium add-ons add $50 per user per month for the reps who get them.

Sum: somewhere between $80,000 and $150,000 per year for a $50M revenue brand running the full Microsoft AI stack. That is between 0.15 and 0.30 percent of revenue, which is in the noise for the value it delivers when the rollout actually lands. The number that matters more than the absolute bill is the percentage of those licenses actually being used six months in.

The build vs buy vs partner decision

Once the stack is understood and the pricing is sized, the next question is who does the implementation. There are three real options, each with a different cost curve, risk profile, and time-to-value.

Build internally

Hire an internal AI engineer or repurpose an existing senior developer. Realistic loaded cost: $180,000-$280,000 per year. Realistic time-to-first-production-system: 9-12 months. Realistic risk: the person you hire to do this is often a former IT generalist, not an AI practitioner, and the resulting systems suffer from the lack of production AI patterns. Build internally only if you have a director-level technical leader who has personally shipped multiple production AI systems before.

Buy off-the-shelf

Pick a SaaS layer (DigitalGenius for support, Salesforce Einstein for sales, etc.) and let the vendor handle implementation. Cost: $30,000-$150,000 per year per use case. Time-to-value: weeks. Trade-off: you do not own the system, you cannot customize beyond what the vendor exposes, and the data flows through the vendor's infrastructure rather than yours. For some use cases this is the right call. For most mid-market AI implementations it leaves too much value on the table. The longer framework we use for these decisions is in the SaaS vendor evaluation checklist.

Partner with an implementation firm

The third option (and the one we recommend in most cases) is to hire an implementation partner who knows the Microsoft estate, builds inside your tenant, and transfers ownership at the end. Realistic cost at mid-market: $50,000-$150,000 per major use case as a project, then $12,000-$20,000 per month for ongoing optimization and additional rollouts. The advantage is that you get production-grade AI architecture from day one, the code lives in your cloud, and the partner has already solved the failure modes that the build-internally path will hit blindly.

Common rollout failures and how to avoid them

Across the engagements we have shipped, the same five failure modes account for nearly every stalled Microsoft AI rollout we have walked into.

The five failure patterns

  1. Licenses without connectors. Copilot is provisioned but no business systems are connected through Foundry IQ. Users get generic answers and stop opening the product within two weeks.
  2. Pilots without owners. A pilot launches with no single person accountable for whether it succeeds. There is no weekly review, no metric, no escalation. Three months in, leadership cannot tell you whether it worked.
  3. No guardrails on customer-facing copilots. A support copilot ships without a classifier in front of it and a judge behind it. It confidently tells a customer something incorrect about a refund policy. The incident kills executive trust and rolls back the entire program.
  4. No observability layer. Power Automate flows and copilots run in production with no central log of inputs, outputs, and failures. When something breaks, nobody can tell what happened. The system goes to "we cannot trust this" within months.
  5. Engineering-led rollout, not ops-led. The project is owned by IT or engineering instead of by the operators who will use it. The resulting system is technically correct and operationally useless.

Every one of these failure modes is avoidable. The next section describes the playbook that avoids them.

The 90-day company-wide rollout playbook: the SOLVE Framework

The AiSolv SOLVE Framework is the five-stage method we run on every Microsoft AI engagement. The name maps to the work: Scope, Orchestrate, Launch, Validate, Embed. The whole framework runs in 90 days for a first major use case, with subsequent rollouts compressing to 30-45 days once the patterns are in place.

S: Scope (weeks 1-2)

Workflow discovery across the operations, marketing, sales, support, and finance teams. We sit with each function for two hours, document the manual work that should be automated, and rank the candidates by hours saved per month and risk of failure. The output is a ranked backlog of 15-30 AI candidate workflows, with a recommended top three for the first 90 days.

O: Orchestrate (weeks 3-4)

Solution architecture for the top three. Which layer of the Microsoft stack each one lives in, what data flows in, what guardrails are required, what success looks like measured in concrete numbers, and what the production observability looks like. Frontier model selection happens here. Fine-tuning gets ruled out almost every time.

L: Launch (weeks 5-8)

Custom build inside the client's tenant. This is where the actual integration to IQ, the Copilot Studio configurations, the Azure OpenAI app builds, and the Power Automate flows get shipped. We always build in the client's environment from day one. No data ever leaves their cloud. The code is theirs.

V: Validate (weeks 9-10)

Production guardrails. Pre-LLM classifiers and post-LLM judges where the system touches customers. A regression test suite that gates any future prompt or model change. Full observability wired in so every prompt, every tool call, every failure is logged. PII redaction at the ingest layer. We always recommend reviewing the same patterns we cover in testing an AI chatbot before shipping here.

E: Embed (weeks 11-13)

Company-wide rollout. Per-department playbooks for what to actually use the system for, role-based context configured through IQ, training sessions for each function, and a 30-60-90 day adoption metric review cadence. This is the stage that traditional consultants skip and that determines whether the licenses get used.

Run SOLVE for the first major use case in 90 days and the second one compresses to 45. The framework is reusable because each stage produces artifacts (the workflow ranking, the architecture document, the test suite, the observability dashboard, the department playbook) that get reused for every subsequent engagement. The brands that get furthest with Microsoft AI are the ones who treat SOLVE as a permanent operating motion, not a one-time project.

Frequently asked questions

What is Microsoft Copilot for business and how is it different from the consumer version?

Microsoft Copilot for business refers to Microsoft 365 Copilot, the enterprise SKU that costs $30 per user per month with a 300-seat minimum and includes access to the M365 productivity surface plus Business Chat across your tenant data. The consumer version (Copilot Pro at $20 per month) is a separate product that does not have tenant data access, enterprise data controls, or admin governance. Make sure your procurement team is buying M365 Copilot, not Copilot Pro.

Do we need Microsoft Fabric to use Microsoft 365 Copilot?

No, M365 Copilot works without Fabric. You need Fabric (or an equivalent unified data layer) when you want copilots and AI agents to read from your operational data outside of Microsoft 365: ERP records, e-commerce transactions, support tickets, billing events. For most $20-100M brands, the first 60 days of M365 Copilot can run without Fabric. By day 90 most operators want it.

Is Azure OpenAI HIPAA-compliant?

Yes, Azure OpenAI is covered under Microsoft's HIPAA Business Associate Agreement when deployed inside an Azure tenant with the appropriate configurations. This is one of the main reasons we use it as the inference layer for telehealth, supplement, and financial services clients. Compliance still requires the broader controls (audit logs, RBAC, data residency, etc.) we cover in the HIPAA-compliant AI implementation guide.

How long does an enterprise Copilot rollout actually take?

For a first major use case shipped to a department, plan for 90 days using a structured method like the SOLVE Framework. Subsequent rollouts (the second, third, and fourth copilots) compress to 30-45 days because the connectors, the observability layer, and the guardrail patterns get reused. Brands that try to roll out Copilot company-wide on day one almost always end up with the licenses-sitting-idle pattern described above.

Should we fine-tune a model for our business or use the frontier model?

Use the frontier model. At mid-market revenue scale, fine-tuning is almost never the right call. The frontier models (GPT-5, Claude 4.6 and 4.7, Gemini 2.5) with proper retrieval and prompt engineering beat fine-tuned alternatives on operator workflows at a fraction of the time and cost. Fine-tuning makes sense for narrow high-volume domains (medical coding, certain compliance checks) where the same prompt runs millions of times. For 99 percent of mid-market use cases, retrieval beats fine-tuning.

How does Microsoft AI compare to ChatGPT Enterprise and Claude Enterprise for mid-market?

The three platforms compete on different axes. Microsoft wins on integration with the productivity stack and Dynamics. ChatGPT Enterprise wins on raw frontier model quality and ChatGPT-native workflows. Claude Enterprise wins on long-context tasks, code, and document analysis. For most $20-100M brands already running on Microsoft, the answer is Microsoft for the core stack with one of the others added for specialty work. The longer comparison is in our enterprise AI platform comparison for mid-market.

What is Microsoft IQ?

Microsoft IQ is the rebrand of Microsoft Graph plus the newer knowledge platform components. As of late 2025 it is four parts: Work IQ (Microsoft 365 data with RBAC), Foundry IQ (external knowledge sources, auto-vectorized through Azure AI Search), Fabric IQ (operational data through OneLake with an ontology layer), and Web IQ (live web search). Every product in the Microsoft AI stack reads from IQ, which makes configuring it well the single most important architectural decision in a mid-market rollout.

Can we use Microsoft AI alongside other AI we already deploy?

Yes, and most mid-market brands should. The Microsoft stack handles the productivity layer and the Dynamics-anchored workflows best. For e-commerce-specific use cases like inventory and demand forecasting, you typically want the patterns covered in our AI for e-commerce operations playbook running alongside the Microsoft stack rather than inside it. The two layers integrate through IQ and through Azure OpenAI.

Bringing it together

The opportunity in Microsoft AI for mid-market in 2026 is not the technology. The technology works. The opportunity is in the implementation. Brands that treat the Copilot license as a starting line, scope the IQ layer carefully, ship one use case at a time using a structured method, and measure adoption monthly are pulling away from the brands that handed out licenses and hoped for the best. The SOLVE Framework is one way to do this. There are others. What matters is having a method, having an owner, and treating each copilot as a product that gets shipped, observed, and iterated, not a tool that gets provisioned and forgotten.

If you are running a $20-100M brand on the Microsoft stack and the licenses are not paying for themselves yet, the gap between where you are and where you could be is roughly 90 days of focused work. The rest is a question of who runs it.