Why a Boutique AI Consulting Firm Beat Two Big Vendors at a $195M Specialty Hardware Distributor
By Ed Hitchcock, Enterprise AI Systems Architect, SupplyTech Solutions
A boutique AI consulting firm wins mid-market work the same way a good general contractor wins residential renovation work. Not by having more people. By showing up with the right people, refusing scope that does not fit, and staying on the job until the client can run the system without us. This case study walks through an engagement at a $195M specialty hardware distributor where two national systems integrators had already burned nine months and about $340K between them, and we finished the actual build in 16 weeks. The point is not that big vendors are bad. Some are excellent. The point is to describe when the boutique AI consulting firm model wins, and to be honest about when it does not.
The Distributor and the Two Prior Attempts
The client is a specialty hardware distributor in the Carolinas, about $195M annual revenue across nine branches and a central DC. They carry roughly 61,000 SKUs of decorative hardware, architectural fasteners, cabinet and door hardware, closet systems, and specialty finishes, serving custom homebuilders, cabinet shops, remodel contractors, and design showrooms. About 240 employees. ERP is a heavily configured Dynamics Business Central on Azure, CRM is HubSpot, and pricing intelligence lives across three spreadsheets one merchandising manager rebuilds every Thursday.
Two prior attempts had failed. Attempt one was a national systems integrator that scoped a $1.1M "AI transformation" starting with a discovery phase they insisted on running before naming any deliverables. Six months in, the CFO killed the engagement after the deliverables turned out to be a data catalog, a maturity model presentation, and a Power BI workspace nobody outside of IT logged into. Attempt two was a mid-tier platform vendor that sold them $180K in licenses for a "supply chain AI" tool that never got past pilot because the input data was in Business Central and nobody on the vendor's side knew how to extract it reliably. Between them: nine months, about $340K spent, zero workflows automated.
The CFO called us in April. Her framing was clear: "I need someone who will name three things they will automate, put a fixed price on it, and be gone by Thanksgiving. If you cannot do that, do not send a proposal."
What a Boutique AI Consulting Firm Actually Sells
The proposal we sent back was four pages. Three workflows named up front, chosen from a two-week diagnostic we ran on our own dime because we did not want a paid discovery phase to be the deliverable. Fixed hourly, 16 weeks. Handoff to an internal Operations Analyst by week 15. No platform licensing attached. No year-two managed services baked in. If they wanted us back for something new, that would be a new SOW.
That is the difference. A boutique AI consulting firm at this tier sells the outcome and the exit. National SIs sell platforms and multi-year relationships because their cost structure requires it. At a $195M distributor with 240 employees and no dedicated AI team, the exit is more valuable than the platform. If we do not leave a system the client can run and extend without us, we have built dependence, not durability.
The other thing a boutique AI consulting firm sells is the ability to say no to the wrong project. In our diagnostic, we identified six workflows the executive team had raised as candidates. We accepted three, rejected three, and told them why. Rejected: dynamic pricing across all 61,000 SKUs (data was too dirty and the merchandising team needed to fix product categorization first), AI-generated customer email responses (volume did not justify the build and the customer service manager was rightly worried about tone drift), and a chatbot for the design showrooms (their showroom staff explicitly did not want it and would have subverted it). Naming rejections out loud, in the readout, is part of the deliverable. That is a piece the two prior vendors had not done.
The Three Workflows We Accepted
The three workflows we accepted were purchase order acknowledgment reconciliation, quote turnaround on complex hardware assemblies, and slow-moving SKU triage across the branch network.
Purchase order acknowledgment reconciliation went first. Their buyers were sending POs to about 340 vendors and receiving acknowledgments back in a mix of PDF, email body, and vendor portal exports. The team was spending roughly 26 hours a week matching acknowledgments to POs to catch price changes, quantity substitutions, and lead time slips. We built an Azure Document Intelligence pipeline that extracts acknowledgment line items, a Power Automate flow that matches each line to the originating PO in Business Central via a read-only staging table, and a Power App queue where buyers see only exceptions. Deterministic matching handles about 78% of acknowledgments end to end. The exception queue takes buyers about 8 hours a week instead of 26.
Quote turnaround on complex hardware assemblies went second. A cabinet shop calling in for a full kitchen hardware package used to wait 4 to 5 business days for a quote because the inside sales rep had to look up 30 to 50 line items across multiple product lines, check current pricing, verify availability, and format the quote in a specific template. GPT-4o generates the assembly line list from a customer description, cross-references it against current Business Central pricing and stock, and produces a draft quote the rep reviews and adjusts. Median quote turnaround dropped from 4.5 days to about 6 hours. Not instant. We deliberately kept the sales rep in the loop for margin control on assemblies over $2,000, because the customers on those orders expect a human on the phone, not a bot.
Slow-moving SKU triage went third. The merchandising manager's Thursday workbook flagged SKUs with no movement in 12 months, but missed SKUs moving in one branch and dead in eight others, which is the real problem at a nine-branch distributor. An Azure SQL job now aggregates movement across branches nightly and surfaces three categories in a Power BI dashboard: dead everywhere, dead-in-region-but-alive-elsewhere (redistribution candidates), and dead-and-obsolete (return or liquidation candidates). GPT-4o writes a one-line explanation on each flagged SKU. Her Thursday routine went from 5 hours to about 40 minutes.
The Architecture Was Intentionally Small
Nothing in the stack was novel. Azure SQL for the analytical layer, Azure Document Intelligence for extraction, Azure AI Search for retrieval, Power Automate for orchestration, GPT-4o via Azure OpenAI for the two generative steps, three Power Apps for human review, Power BI for the SKU dashboard. Everything runs inside the client's existing Azure tenant. Incremental Azure consumption sits at roughly $1,650 per month at current volumes. Zero net-new platform licenses, zero new vendor relationships to maintain.
This is another place the boutique model earns its keep. A national SI would have proposed a proprietary orchestration layer, a vendor-branded vector database, and a "center of excellence" model with three-to-five FTEs on retainer. All of that is defensible on paper. None of it survives the first budget cycle at a $195M distributor with a single Operations Analyst as the AI owner. Small stacks that the client's team can extend beat large stacks that require external help forever.
What We Refused to Automate After the Build Started
Two things came up mid-engagement that we refused, and this matters for what a boutique AI consulting firm looks like when the client wants to expand scope on the fly.
The CFO asked, in week 9, whether we could add an AI-generated cash forecast to the SKU dashboard. We said no. Cash forecasting at a distributor mixes AR aging, seasonal buying patterns, vendor payment terms, and inventory turns. Building it inside a 16-week engagement would have short-changed one of the three original workflows or produced a forecast we could not defend under audit scrutiny. We offered to scope it as a separate 6-week engagement in Q1 next year. She agreed.
The COO asked, in week 12, whether we could plug the quote assembly generator into the HubSpot workflow so a rep could trigger it from a deal record. We said not yet. HubSpot integration was reasonable but the quote assembly model needed 60 more days of production use before we trusted it as an inline component of the sales workflow. We shipped the Power App standalone and told the COO we would revisit HubSpot integration in Q1 based on the usage data.
Refusing scope changes mid-engagement is a specific behavior of a boutique AI consulting firm that has ownership incentives aligned with the client. A vendor whose commercial model rewards scope creep will accept both of those requests, and the client will pay for it in delivery risk and adoption friction.
The 16-Week Sequence and the Handoff
The engagement ran Wk 1 to 2 diagnostic and scope contract, Wk 3 to 6 purchase order acknowledgment build, Wk 7 to 10 quote assembly build, Wk 11 to 14 slow-moving SKU build, Wk 15 governance documentation and adoption sessions, Wk 16 handoff to the Operations Analyst as the named owner. She had shadow-built the last of the three workflows alongside us so the handoff was more of a formality than a training exercise.
What Changed at 90 Days Post-Handoff
Purchase order acknowledgment matching time dropped from 26 hours a week to about 8. First-pass match accuracy on the exception queue after buyer review sits at 91%. Quote turnaround on complex assemblies dropped from 4.5 days to about 6 hours, and quote-to-order conversion on those assemblies moved up by 6 to 9 percentage points, which the sales director attributes to the faster response. The merchandising manager's Thursday routine dropped from 5 hours to 40 minutes, and the redistribution flagging surfaced roughly $310K in inventory the branches were able to rebalance in the first quarter. Total engagement cost landed at just under 0.11% of annual revenue.
The Operations Analyst has since built one new Power Automate flow on her own, a returns intake process, without our involvement. That is the actual test. Not the demo on handoff day. Whether the internal team can extend the system three, six, twelve months later without calling the consultant back.
When the Boutique Model Fits, and When It Does Not
The boutique AI consulting firm model fits mid-market distributors, manufacturers, and industrial services firms in the $100M to $500M revenue range, with an existing Microsoft footprint, and one internal person who can be trained as the ongoing owner. It does not fit companies below $75M because volume does not justify build cost. It does not fit companies above $750M because governance and platform breadth exceed what any boutique offers. It also does not fit any company where the sponsor wants AI as marketing rather than a workflow tool.
For distributors in the fit range, the tell is simple. Ask each vendor to name three workflows they will automate before you write a check. If they cannot, or want a paid discovery phase to answer that question, they are not the right partner at this revenue tier. Neither are we, if we cannot name three either.
Comments