Agentic AI Consulting: How We Built a Department-Scoped Digital Workforce for a Commercial Equipment Distributor
By Ed Hitchcock, Enterprise AI Systems Architect, SupplyTech Solutions
When a $140M commercial equipment distributor asked us to help them deploy AI agents across their operation, the executive team had already watched two pilots stall inside the previous year. Both followed the same pattern: a vendor demo, a wide-scope agent build covering ten departments, six months of integration work, and a launch that no one used. The agents could answer questions about the company. They could not actually do the work.
We took the engagement on the condition that we narrow the scope hard. Agentic AI consulting that produces real adoption looks nothing like a horizontal copilot rollout. It looks like a small number of department-scoped agents, each grounded in documented processes, each capable of executing a specific set of multi-step actions inside the tools employees already use. This case study walks through how we built that digital workforce, what we shipped in the first 90 days, and where the operating leverage actually came from.
The Client and the Starting Point
The client runs four branches in the Southeast and Mid-Atlantic, distributing commercial restaurant, hospitality, and facilities equipment to operators, installers, and service contractors. About 165 employees. Annual revenue in the $135-145M band. Their tech stack at the time of engagement: an aging on-prem ERP for inventory and order processing, a separate CRM the sales team had stopped trusting two years earlier, SharePoint for documents, and Microsoft 365 across the board. Power Automate licenses were already purchased but barely used.
The two failed pilots had been built outside Microsoft 365. One was a chatbot wrapper on a hosted LLM. The other was a custom RAG application built by a regional consultancy. Both pulled data from the ERP through one-way exports. Neither could write back to a system of record. Employees treated them as expensive search tools and stopped opening them within six weeks of launch.
When the executive team brought us in, the ask was reframed: stop trying to build one AI for the whole company. Build agents that can actually do the work in three departments where the operational drag is most visible.
Why This Sequence Matters for Agentic AI Consulting
Most failed AI implementations we see share a single root cause. The agent has no defined work to perform. It exists as a general-purpose assistant grafted onto an undefined process. Employees discover within a week that asking the agent to "help with quotes" returns generic advice, not a completed quote sitting in the CRM ready for sales review.
The correction is not better prompts. The correction is process scope. An agent that can complete a defined sequence of steps inside an existing system is worth more than ten agents that can hold a conversation about the same domain. This is what shifts AI from a productivity demo to an actual digital coworker, and it is the principle that drives every department-scoped build we ship.
For this engagement, we defined three departments where the work was repetitive enough, documented enough, and contained enough to support an agentic build inside a 90-day window: order desk, customer service, and parts/aftermarket support.
The Three Agents We Shipped
We built three department-scoped agents in Microsoft Copilot Studio, each grounded in SharePoint content and connected to Power Automate flows that execute the actual write-back actions into the ERP, the CRM, and the ticketing system.
Order desk agent. This agent handles inbound order intake from the seven largest dealer customers, who collectively account for roughly 38% of the client's annual revenue. The agent reads incoming purchase order emails, extracts line items against the SKU master in SharePoint, flags pricing exceptions against the customer's contract sheet, and drafts a sales order in the ERP staging table for the order desk rep to review and post. When the customer's PO is clean, the rep clicks post. When something is off, the agent surfaces the exception in plain language with the contract clause it triggered.
Customer service agent. This one handles status and tracking inquiries that previously consumed roughly 22 hours a week across the four-person service team. The agent pulls order status from the ERP, carrier tracking from the shipping system, and outstanding service tickets from the helpdesk. It drafts a customer-ready response with all three data points stitched together, then either sends directly for inquiries it has high confidence in or queues for review when something is incomplete. The team approves or edits before send.
Parts and aftermarket agent. This agent supports the parts desk, where employees were spending substantial time looking up compatibility, cross-references, and supersession data across vendor catalogs and internal documentation. The agent retrieves the relevant catalog page, the supersession history, and the in-stock position from inventory, then drafts a parts quote with availability and lead time. It does not auto-send. The parts rep reviews and converts to a quote in the CRM.
All three agents are grounded against the same SharePoint knowledge base, but each has a scoped skill set and a defined set of Power Automate flows it is permitted to trigger. None of them can execute anything outside their defined scope. That constraint is the design choice that makes them trustworthy.
The Architecture
The stack we built sits inside the client's existing Microsoft 365 tenant. We did not introduce new platforms.
Knowledge layer. SharePoint Premium with AI-driven classification and metadata. We restructured the document libraries so each agent can retrieve exactly the SOPs, contracts, catalogs, and reference sheets relevant to its scope. Documents outside an agent's scope are filtered out at the retrieval layer.
Agent layer. Three Copilot Studio agents, each with custom topics, defined system prompts, and a curated list of plugin actions. Each agent has its own persona and instruction set tuned to the department it serves.
Workflow layer. A library of Power Automate flows, roughly 30 of them across the three agents, that execute the multi-step actions: pull data from the ERP, write to staging tables, draft emails, log activity in the CRM, create helpdesk tickets, post status updates.
System of record connectors. Custom connectors built against the on-prem ERP through an on-premises data gateway. Direct REST connectors to the CRM, the helpdesk, and the shipping system. Every write action is logged with an agent identity and a user approval flag so audit trails stay clean.
Approval and review patterns. Every agent action that touches a system of record routes through a Teams adaptive card for human approval, except for the small set of low-risk read-and-respond patterns the customer service agent handles directly. Nothing the agents do is one-way and unrecoverable.
The Engagement Shape
The work ran 14 weeks in three blocks.
Weeks 1-3. Process scoping. We sat with each of the three departments, documented the existing workflows, and identified the specific sequences where an agent could plausibly execute. We threw out twelve candidate use cases that looked attractive but lacked the documentation depth or the system access to support a clean build.
Weeks 4-10. Agent and workflow build. SharePoint restructuring, custom connectors against the ERP, Copilot Studio agent configuration, Power Automate flow construction, internal testing against historical data. We ran two end-to-end test rounds inside this block with department reps shadowing the agent.
Weeks 11-14. Department-by-department rollout with side-by-side operation for the first two weeks per department. Each rep ran the agent next to their existing workflow, compared outputs, and approved or rejected agent suggestions. We instrumented every approval, rejection, and edit to feed back into the agent's instruction set.
Total engagement cost landed at roughly 1.2-1.5% of annual revenue. The client's prior failed pilots had cost more in aggregate and produced nothing in production.
What We Measured in the First 90 Days Post-Launch
These numbers are early. They reflect the first 90 days after each agent reached steady-state usage in its department.
Order desk. Roughly 65-70% of inbound orders from the seven priority dealer customers now flow through the agent. Average time from email receipt to ERP staging dropped from 18-22 minutes per order to 4-6 minutes. The order desk rep is still in the loop on every order, but the manual extraction and contract-checking work is collapsed.
Customer service. The agent now drafts roughly 70% of inbound status and tracking responses. Direct-send rate on high-confidence responses runs around 30% of total volume. The team estimates 14-18 hours per week of administrative drag has been removed across the four reps.
Parts and aftermarket. Average quote turnaround for in-scope inquiries dropped from 35-45 minutes to 10-15 minutes. The agent has not changed quote conversion rates measurably yet, but the parts manager reports the team is handling roughly 25% more quote requests per week with the same headcount.
Cross-cutting. Agent activity is logged and reviewed weekly. Across the three agents we are seeing a sustained 70-80% approval rate on agent-drafted actions, which is the threshold we use to consider an agent stable enough to expand scope.
What We Did Not Do
We did not build a general-purpose company assistant. We did not give any agent the ability to issue refunds, write credit memos, or modify pricing without human approval. We did not deploy outside the three pilot departments.
Trust in an agentic system is earned at the scope where it executes reliably. Expanding scope before that trust exists is what produces the abandoned pilots we have walked into at other clients.
What This Says About Agentic AI Consulting in 2026
The vendor pitch deck era of AI consulting is ending. Operators have run enough failed pilots to know that a model demo is not an implementation, and a hosted chatbot is not an agent. What they want is a small number of agents that complete defined work inside their existing systems, with audit trails, with approval gates, and with measurable throughput improvements they can point at on a Monday morning.
That is what agentic AI consulting looks like when it actually moves the operation. Scope the work tightly. Ground the agent in real documented processes. Wire it to systems of record through proper write-back patterns. Keep humans in the approval loop for anything irreversible. Instrument every action so the agent's behavior improves on a known schedule.
The Commercial Equipment Distributor build we shipped for this client is now in scope expansion. We are adding a finance agent for AR follow-up and a service dispatch agent for the field tech scheduling team. Both will follow the same pattern: department scope, documented process, Power Automate write-back, human approval on anything irreversible. The architecture holds.
If an agentic AI consulting pitch starts with the model rather than the process, walk. The model is the easy part. Process scoping, system integration, and trust calibration are where the work lives.
SupplyTech Solutions builds operating systems for mid-market enterprises. We design and ship the documentation, knowledge architecture, agent layer, and ERP modernization that turns fragmented operations into a scalable platform. If you are running a distribution or services business and you want a credible read on where agentic AI fits in your stack, reach out.
Comments