What AI Automation Consulting Actually Looks Like: An 18-Week Engagement Inside a $280M Industrial Parts Distributor
By Ed Hitchcock, Enterprise AI Systems Architect, SupplyTech Solutions
Most of the AI automation consulting proposals we see land on a $280M industrial parts distributor's desk fall into two categories. Either a national systems integrator quoting a two-year $1.4M program that starts with a six-month "discovery," or a small vendor promising to bolt an LLM onto whatever's already broken and call it done in eight weeks. Neither is what an operator actually needs. This is the case study of what fit-for-purpose AI automation consulting looks like at that revenue band: an 18-week engagement, one Ops Analyst embedded as the client-side owner, an architecture the internal team can run without us after handoff, and specific automation targets picked because they were bleeding hours, not because they photographed well in a deck.
The Distributor and the Original Ask
The client is a privately held industrial parts distributor doing about $280M in annual revenue across 11 branch locations in the Southeast and mid-Atlantic. They carry roughly 84,000 SKUs across bearings, power transmission, hydraulics, fluid conveyance, and industrial fasteners, serving MRO buyers at manufacturing plants, municipal utilities, food processors, and mining operations. Around 340 employees, ERP is an on-prem Epicor instance customized heavily over 14 years, CRM is a lightweight cloud tool the outside sales team barely uses, and inventory intelligence lives in an Excel workbook a purchasing manager rebuilds every Monday morning.
The COO called us in March. His original ask was blunt: "I have three vendor quotes for AI, they all want to sell me a platform, and I do not understand what any of them are actually going to do on Monday morning. Can you tell me what we should buy and what we should build, and if the answer is neither, tell me that." That framing is important, because it shapes what AI automation consulting is supposed to deliver at this revenue tier. Not a platform license. Not a Slack bot. A short, honest read on which workflows have enough volume and structure to justify automation, which do not, and a build plan that leaves the client capable of running and extending the system after we leave.
What Our Version of AI Automation Consulting Actually Includes
We wrote the SOW for this engagement to cover four things and nothing else. First, an assessment of the top 15 candidate workflows across purchasing, quoting, warehouse operations, and finance, scored on volume, structural stability, and dollar impact. Second, the design and build of the three workflows that scored highest, using the client's existing Microsoft and Azure footprint rather than introducing new platforms. Third, a governance framework covering data access, model outputs, human approval points, and an incident playbook. Fourth, a 60-day adoption period with a named internal owner trained to operate and extend the system.
What we did not include, and this matters: no reseller relationships, no platform commissions, no year-two managed services attached to the SOW as a soft requirement. Fixed hourly, 18 weeks, all deliverables in the client's tenant. If AI automation consulting only works when the consultant stays forever, it is not consulting, it is dependence.
The Workflow Selection Process
The most valuable week of the engagement was week two. We ran structured interviews with 14 people across purchasing, inside sales, quoting, warehouse ops, and finance. We asked each of them the same three questions: what do you spend the most time on that a robot could do, what would you refuse to let a robot do, and what have you tried to automate before that failed. Then we scored 15 candidate workflows against three axes. Volume, meaning does this happen enough per week to justify build cost. Structure, meaning is the input predictable enough that a model or rule engine can handle 80% of cases without a human. Dollar impact, meaning does automating this either save meaningful labor hours or protect margin.
The top three by combined score were not the ones the executive team expected. Vendor invoice matching came in first because the AP clerk was spending 22 hours a week on three-way matching against POs and receivers, and 18% of invoices were flowing through as exceptions that required chasing. Quote-to-order conversion for the top 200 accounts came in second because inside sales was retyping quote lines into the ERP for orders they had already quoted, at an average of 9 minutes per order across ~180 orders per week. Slow-moving inventory identification came in third because the purchasing manager's Monday workbook was catching about half of what it should catch, and the team was carrying an estimated $4.2M in inventory that had not moved in 18 months.
We deliberately dropped four workflows the executives had pre-ranked as priorities. Customer credit hold decisions came off the list because volume was low and each decision needed a human anyway. Sales commission calculation came off because the underlying commission plan changed every year and automation would have to be rebuilt annually. Two others came off because interviews revealed the "problem" was actually a training issue, not an automation opportunity. Naming those out loud in the readout is part of what AI automation consulting is supposed to do. Saying no to bad projects is worth more than saying yes to mediocre ones.
The Build Sequence
Weeks 3 through 15 were the build. We ran the three workflows in sequence rather than parallel because the client had one Ops Analyst who was going to own everything after handoff, and running three parallel builds would have meant she owned nothing well.
Vendor invoice matching went first, weeks 3 through 6. Azure Document Intelligence extracts invoice line items and totals from PDFs and email attachments. A Power Automate flow retrieves the matching PO and receiver from Epicor via a read-only staging table. Deterministic matching handles the 82% of invoices where quantities, prices, and terms align inside tolerance. Exceptions route to a Power App queue where the AP clerk sees a side-by-side view of what the model extracted, what the PO said, and what the receiver logged. Approval is one click. Rejection requires a reason code that feeds back into the training set.
Quote-to-order conversion went second, weeks 7 through 10. Not glamorous. An inside sales rep opens the quote in the CRM, hits "convert to order," and a Power Automate flow pulls the quote lines, resolves them against current Epicor pricing and stock, flags any line where quoted price differs from current price by more than 3%, and writes a draft sales order the rep approves in Epicor. Nine minutes of retyping per order becomes about 90 seconds of review. We considered adding a chatbot layer for order status queries. We rejected it because the volume did not justify the build and the outside sales team preferred texting the branch directly.
Slow-moving inventory identification went third, weeks 11 through 14. An Azure SQL job runs nightly against Epicor inventory tables, computes rolling 6-, 12-, and 18-month movement, cross-references open POs and forecast demand, and flags SKUs that meet dead-stock criteria. GPT-4o generates a plain-English one-line summary of why each flagged SKU landed on the list, referencing the movement pattern and any nearby substitutes carried in the same product family. The purchasing manager gets a Power BI dashboard on Monday morning instead of rebuilding the workbook by hand. What used to take her four hours takes 20 minutes of review.
Weeks 16 through 18 were governance documentation, adoption sessions with each team, and handoff to the Ops Analyst as the named owner. Not much drama in those weeks. That was the point.
The Architecture, Kept Boring on Purpose
Nothing in this stack was novel. Azure SQL for the data layer, Azure Document Intelligence and Azure AI Search for retrieval and extraction, Power Automate for orchestration, GPT-4o via Azure OpenAI for the two summarization tasks that needed it, three Power Apps for human-in-the-loop review, Power BI for the inventory dashboard. Everything runs inside the client's existing Microsoft tenant. Zero net-new licenses beyond incremental Azure consumption, which lands at roughly $1,900 per month at current volume.
We chose boring on purpose. A boutique AI consulting firm can be tempted to bring in a trendy vector database or a custom agent framework because it looks sophisticated in the architecture diagram. That temptation is where post-handoff failure lives. The client's Ops Analyst can extend Power Automate flows and modify Power App screens on her own. She cannot maintain a custom Python agent orchestrator, and she should not have to.
What Changed at 90 Days Post-Handoff
We checked in 90 days after handoff. The AP clerk's invoice matching time dropped from 22 hours a week to about 7. Exceptions still take real work, but the deterministic path handled 82% of volume as designed and first-pass match accuracy on the exception queue after her review sits around 94%. Quote-to-order conversion time dropped from 9 minutes per order to about 90 seconds of review, freeing roughly 22 hours a week across the inside sales team. Slow-moving inventory identification caught an additional $760K in dead stock the old workbook was missing, and the purchasing manager's Monday routine went from four hours to 20 minutes.
The Ops Analyst has since built two new Power Automate flows on her own, one for return-merchandise-authorization intake and one for customer part-number cross-reference cleanup. Neither was in scope for our engagement. Both took her under two weeks. That is the actual measure of whether AI automation consulting worked. Not the demo on handoff day, but whether the client can extend the system six months later without calling us.
When This Model Fits, and When It Does Not
This engagement model fits a specific kind of company. Mid-market, roughly $150M to $500M in revenue, privately held or private-equity-owned, with an existing Microsoft footprint and at least one internal person who can be trained as the ongoing owner. It does not fit a $50M business that lacks the volume to justify build cost, and it does not fit a $2B enterprise that needs governance depth and platform breadth we do not offer. It also does not fit any company where the executive sponsor wants AI as a marketing story rather than an operational tool. We turn down two of those calls a month.
If you are inside a mid-market distributor, manufacturer, or industrial services firm and you have vendor quotes that read like platform pitches instead of workflow plans, the check to write is smaller than you think and the timeline is shorter than you have been told. AI automation consulting at this tier is 18 weeks, fixed hourly, three workflows, one owner, and a stack your team can run without the consultant. Anything longer or more elaborate should have to justify itself before you sign.
Comments