Why We Built theywork365.ai: 15 Years of Microsoft Projects and a Completely Different Model

Most AI projects in the enterprise space share an uncomfortable trajectory. They start with a compelling demo. A small pilot produces encouraging results. Then the project stalls somewhere between proof-of-concept and production rollout, and six months later the steering committee quietly shelves the budget line. According to McKinsey’s 2024 State of AI report, only 11% of enterprise AI pilots make it to full-scale deployment. We’ve watched this pattern play out more times than we can count.

After 15 years and more than 200 Microsoft 365 implementations across industries ranging from manufacturing to financial services, we stopped asking “how do we build the AI?” and started asking “why do these projects keep dying?” The answer wasn’t technical. It was structural.

That’s what led us to build theywork365.ai.

In brief

• Only 11% of enterprise AI pilots reach full deployment, per McKinsey 2024 — the problem is structural, not technical
• Three root causes kill AI projects: upfront cost lock-in, shadow AI adoption, and data leaving the organizational perimeter
• theywork365.ai deploys custom AI agents inside your existing Microsoft 365 tenant, with no upfront project cost
• A pay-per-action model means ROI is measurable from the first week of operation
• All data stays inside the customer’s tenant — no third-party cloud exposure

What 15 Years of Microsoft Implementations Taught Us

Enterprise clients on Microsoft 365 consistently face the same obstacle when they try to adopt AI at scale: the cost structure works against the outcome. (McKinsey & Company, 2024) Gartner estimates that AI project failure rates remain above 85% when measured against original business objectives, and cost overrun is the single most cited reason.

The typical engagement looks like this. A software vendor proposes a six-figure discovery and build phase. The organization commits. Months of workshops produce requirements documents. Then the engineering phase begins and the scope grows. By the time a first working version is ready to test, the business context has shifted, stakeholders have changed, and the original ROI model is obsolete.

We’ve been brought in to rescue three separate AI projects in the past two years alone. In every case, the company had spent between €80,000 and €150,000 before a single user had touched the system. In one case, the internal sponsor had left the company and the project had no owner. The investment was essentially written off.

The other pattern we kept seeing was what we’ve come to call the “governance gap.” AI governance frameworks almost always arrive late in these projects. Security reviews, compliance assessments, and data handling policies get added after the architecture is already defined. That’s when the expensive rework starts.

What Structural Problems Actually Kill AI Projects?

Three issues come up repeatedly. Each one is solvable, but only if you design around them from the beginning rather than patching them in later.

Upfront Project Costs That Block ROI

Traditional AI implementation follows a consulting model: you pay for discovery, design, build, and integration before any value is delivered. According to Deloitte’s 2025 Technology Trends report, the average enterprise AI project takes 14 months from initiation to production. That’s 14 months of investment with zero measured return.

The result is a selection bias. Only large organizations with deep reserves can absorb that kind of risk. Mid-market companies on Microsoft 365 — many of them running on Business Standard or E3 licenses — have the same operational pain points but can’t justify the upfront spend. They either skip AI entirely or run informal, ungoverned pilots that create more problems than they solve.

According to Deloitte’s 2025 Technology Trends report, the average enterprise AI project takes 14 months from initiation to production. Organizations that committed large upfront budgets reported a 67% rate of scope change during that window, directly eroding the original ROI case. The structural implication: cost lock-in precedes value delivery by over a year in most deployments. (Deloitte, 2025)

Shadow AI: What Happens When Employees Bypass Official Tools

When the official AI rollout stalls or doesn’t meet day-to-day needs, employees find their own solutions. Microsoft’s 2025 Work Trend Index found that 78% of employees bring their own AI tools to work without IT approval. They paste customer data into ChatGPT. They run contract summaries through consumer Gemini. They build personal automations that bypass access controls.

Shadow AI is not a behavior problem. It’s a governance design problem. Employees are using these tools because the officially sanctioned path is too slow, too generic, or simply absent. If you address the governance symptom without addressing the underlying need for practical AI tooling, you get compliance theater: employees sign policies and then continue doing what they were doing anyway.

This matters beyond compliance. Every time an employee pastes a customer record or internal document into a consumer AI service, that data leaves the organizational perimeter. Most consumer AI services use submitted content for model training by default. The exposure risk is not hypothetical.

Data Leaving the Organizational Perimeter

The third problem follows directly from the second. Organizations that don’t provide embedded AI tooling inside their existing environment create pressure for employees to move data outward. The EU AI Act, which began applying enforcement provisions in August 2024, establishes clear organizational liability for uncontrolled AI use that processes personal data or high-risk business information.

Microsoft’s own research indicates that only 39% of organizations have formal data handling policies that cover generative AI use by employees. The rest are operating on informal norms that don’t hold up under regulatory scrutiny. The organizations we work with can’t afford to discover their exposure after an incident.

What We Built Instead

theywork365.ai is our answer to all three problems, developed from the ground up on the Microsoft 365 stack we’ve been building on for 15 years. It’s not a consulting engagement. It’s a platform that deploys custom AI agents inside your existing Microsoft 365 tenant.

The agents are vertical. They don’t try to do everything. An IT support agent triages helpdesk tickets, resolves known issues autonomously, and escalates only what requires human judgment. An HR agent handles onboarding documentation, answers policy questions, and routes leave requests. A procurement agent validates purchase orders against approved vendor lists and flags exceptions for review.

The commercial model reflects what we learned from watching projects fail. The pay-per-action model means you pay for results, not for the project. No large discovery phase, no six-figure build commitment. A governance subscription covers maintenance, model updates, and monitoring. You start seeing measurable activity from the first week.

theywork365.ai operates on a pay-per-action commercial model, charging per agent-executed task rather than per user seat or upfront project fee. In our early deployments, organizations reported a median time-to-first-measurable-ROI of 11 days after go-live. The governance subscription model includes continuous monitoring and model updates, replacing the traditional post-implementation support contract. (theywork365.ai governance and pricing, 2026)

Why 100% Inside Microsoft 365 Matters

The architecture choice we made was deliberate. Every agent runs inside the customer’s own Microsoft 365 tenant. No data touches an external cloud, a third-party model host, or our own infrastructure. This is not a marketing claim — it’s a technical constraint we built the entire platform around.

The components involved are ones your IT team already manages. Azure AI Foundry provides the language model capabilities, running inside the customer’s Azure subscription. Copilot Studio handles the agent orchestration and conversation logic. Entra ID enforces authentication and respects the existing RBAC configuration: an agent only has the permissions the assigned service account has, nothing more. SharePoint and Teams serve as the primary interaction surfaces.

For detailed technical and compliance specifications, the data security documentation covers tenant isolation, data residency, and the audit trail architecture.

What this means in practice: the compliance team doesn’t need to assess a new external vendor relationship. The DPO doesn’t need to sign off on a new data processing agreement for an external AI provider. The security team can audit the agent’s activity logs in the same place they audit everything else. The entire deployment sits inside the governance perimeter you’ve already established.

How Long Does It Actually Take to Go Live?

The deployment process follows four defined phases. We’ve structured it to reach production in weeks, not months. Details on how the deployment works are available on the platform site, but here’s the practical sequence we follow.

Phase 1: Environment Assessment and Setup

We start by reviewing your existing Microsoft 365 configuration: license tier, Entra ID setup, SharePoint structure, and current governance policies. This takes one to two weeks and produces a clear picture of what’s already in place. No surprises in later phases.

Phase 2: Agent Design and Configuration

This is where we define exactly what the agent does. We’re not building general-purpose assistants. We map the specific process the agent will handle, define the decision rules, set the escalation conditions, and configure the integration points. For a standard IT support agent, this phase runs two to three weeks.

Phase 3: Controlled Go-Live

The agent launches in a controlled environment with a defined user group. Activity is monitored in real time. We track which actions the agent completes autonomously, which it escalates, and where it asks for clarification. Adjustments happen in this phase based on observed behavior, not assumptions.

Phase 4: ROI Measurement and Scale

From the first week of live operation, we’re measuring. Tickets resolved without human intervention. Documents processed per hour. Average response time versus the previous baseline. This data feeds the business case for expanding to additional agent types or business areas.

Where Does It Apply?

The available use cases span ten business areas. The agents we deploy most frequently fall into IT support, HR operations, and sales enablement — but the pattern applies across procurement, finance, compliance, legal, customer service, operations, and marketing.

Two concrete examples from current deployments illustrate how the agents actually operate in practice.

An IT support agent deployed at a manufacturing company with 800 employees handles first-line helpdesk requests through Microsoft Teams. It resolves password resets, software access requests, and known configuration issues autonomously. In the first month, it handled 62% of incoming tickets without human intervention. The IT team’s average resolution time on escalated tickets dropped by 40% because the agent pre-populated the ticket with diagnostic information before handing off.

An HR onboarding agent at a professional services firm guides new hires through their first two weeks via Teams. It delivers required policy documents at the right time in the onboarding sequence, answers frequently asked questions about benefits and leave policies, and triggers the system access request workflows automatically. HR reported a 70% reduction in ad-hoc onboarding queries to the team.

Frequently Asked Questions

What is theywork365.ai?

theywork365.ai is a custom AI agents platform built entirely inside the Microsoft 365 ecosystem. Agents perform real operational tasks — triaging support tickets, approving documents, updating CRMs — with a pay-per-action model and no upfront project cost.

How is theywork365.ai different from Microsoft Copilot?

Microsoft 365 Copilot is a general-purpose AI assistant embedded in M365 apps. theywork365.ai builds custom agents designed around your specific business processes. The distinction matters: Copilot helps a person do their work faster. A theywork365 agent handles the work autonomously, with defined governance and measurable output. According to Forrester Research, purpose-built task agents deliver 3-5x the automation yield of general-purpose AI assistants in structured business processes. (Forrester Research, 2025)

How does the pay-per-action model work?

You pay only for the actions your agents execute. There’s no per-user monthly seat and no large upfront project fee. A governance subscription covers maintenance, model updates, and continuous monitoring. The governance and pricing page has the current tier structure and per-action rates.

Does our data ever leave our Microsoft tenant?

No. Every component runs inside your Azure subscription and Microsoft 365 tenant. The language model capabilities run through Azure AI Foundry in your own environment. No data is sent to external services, including ours. The data security page covers the technical architecture in detail.

Which Microsoft 365 license tiers are compatible?

Most standard enterprise configurations work, including Business Standard, E3, and E5. The specific requirements depend on the agents you deploy and which Microsoft services they integrate with. The environment assessment in Phase 1 confirms compatibility before any commitment.

---

Fifteen years of Microsoft projects taught us that the organizations struggling most with AI adoption aren’t short on ambition or budget. They’re working against a cost model and a governance model that were never designed for the pace at which AI needs to move.

theywork365.ai is our attempt to solve the structural problem first. Custom agents, inside your tenant, paying only for what gets done. If you’re an IT manager or digital transformation lead tired of AI pilots that never reach production, that’s exactly the conversation we want to have.