AI Implementation Consulting: Why the First Six Months Are a Documentation Project, Not a Model Project
By Ed Hitchcock, Enterprise AI Systems Architect, SupplyTech Solutions
When a specialty hardware distributor asked us to lead their AI implementation, the leadership team came in with a list of vendor pitches. Copilots for sales. An AI agent for purchasing. A generative model bolted onto their ERP. Each pitch had a polished deck, a reference customer, and a six-figure annual price tag.
We declined to start with any of them.
The first phase of credible AI implementation consulting for a mid-market distributor is not a model rollout. It is a documented and costed view of how the company actually runs. Without that, every downstream investment, agents, ERP modules, copilots, prediction models, gets configured against guesses. The vendors will still sell you the software. The software still gets installed. The savings just never materialize, because no one ever quantified the work the AI was supposed to replace.
This post walks through the Phase 1 engagement we ran for a specialty hardware distributor, what we documented, what it cost to do, and what we found. The client mask is anonymized. The methodology, sequence, and outcomes are real.
The Client and the Starting Point
The company is a specialty hardware distributor running roughly $95M in annual revenue across three branches in the Midwest. They distribute fasteners, fluid power components, and specialty MRO supplies into manufacturing, agriculture, and construction end markets. Roughly 110 employees across sales, operations, finance, and warehousing.
Their stack is typical. A legacy ERP installed in 2009 and customized twice. A standalone CRM the sales team uses inconsistently. A warehouse system that exports flat files to the ERP overnight. A Power BI footprint with about six active dashboards. M365 Premium licenses purchased but underused. SharePoint deployed as a file dump, not a knowledge platform.
Leadership had spent nine months evaluating AI vendors. Three competing proposals ranged from $180K to $420K in year-one fees. None of them could answer a simple diligence question: how many hours per week does your team currently spend on the workflow this product replaces, and how much does that labor cost.
That gap is the reason most AI implementations stall after the pilot. The right first project is to close it.
What Phase 1 Actually Produces
In our methodology, Phase 1 is called Documented & Costed Processes. The deliverable is an operational blueprint that maps every recurring workflow in the company, identifies the owner, captures the step sequence, measures the time and labor required, and assigns a fixed and variable cost to each one.
It produces five things by the end of the engagement.
First, a process inventory. Every repeatable workflow across operations, finance, sales, and warehousing, captured in a structured taxonomy. For this client, the final count was 187 workflows.
Second, a Lean Value Stream Map for the workstreams that touch revenue or customer experience. We mapped 14 of the 187 in detail. Order-to-cash, quote-to-order, inbound receiving, cycle counting, returns processing, vendor onboarding, customer onboarding, and seven others.
Third, a costing model. Each workflow gets a labor cost, a tool cost, and where applicable a per-transaction variable cost. The model attributes those costs to the activities and groups them up to departments and to the P&L lines that already exist in their general ledger.
Fourth, a standardized template library. SOPs, work instructions, training documents, and process maps all conform to one format. That format is the input format that the eventual knowledge management system, AI agents, and ERP configuration work will consume.
Fifth, an opportunity register. A ranked list of workflows where automation, AI, or process redesign would produce a measurable return. Each entry includes the current state cost, the target state estimate, the implementation effort, and the dependency chain.
Nothing in that list is glamorous. Nothing in that list requires a model. All of it is required before you can credibly scope, build, or measure the impact of anything that does.
The Engagement Shape
We ran the engagement over 18 weeks with a small team. One SupplyTech principal on site one to two days per week, a SupplyTech process analyst on site three days per week, and a rotating cast of internal subject matter experts from the client. The internal SMEs were the source of truth. Our role was to extract, structure, validate, and cost the knowledge that already lived in their heads.
The work happened in three blocks.
Weeks one through four were a gemba walk. We sat with the people doing the work, watched them do it, and asked them to narrate. The output was a draft process inventory and a first cut at the workflows that warranted detailed mapping. The goal was to see the reality, not to confirm what the org chart said.
Weeks five through twelve were detailed mapping and costing. For each of the 14 priority workstreams, we built the value stream map, validated it with the workflow owner and a frontline operator, and ran the costing pass. Costing pulled time data from M365 audit logs and ERP transaction histories. Three workflows required full time studies because they had no digital trail. The rest we reconstructed from system data with sampling and validation.
Weeks thirteen through eighteen were synthesis. Building the template library, harmonizing the maps and SOPs, loading the structured outputs into SharePoint as the input layer for Phase 2, and validating workflows, costs, and opportunity rankings with each department head.
What We Found
Three findings drove the rest of the program design. None of them required AI to surface. They were visible the moment the work was documented and costed.
The first finding was that the company was spending approximately $1.6M to $1.9M per year on duplicative or working-around work caused by the gap between their CRM, ERP, and warehouse system. Roughly 22 to 28 hours per week, across the team, went into manually reconciling order data between the three systems. That is the workaround tax. It was invisible in the P&L because it was distributed across 30 employees in fragments of their day. Once we costed it, it was the single largest operational expense item not tied to direct labor or freight.
The second finding was that customer onboarding, a process leadership assumed took five to seven business days, actually took 14 to 21 business days from quote acceptance to first fulfilled order. The delay came almost entirely from manual credit setup, manual item master configuration, and back-and-forth between sales and finance that nobody owned. The customer experience problem the CEO had been hearing about for two years was a documented process problem.
The third finding was that 35 to 40 percent of the workflows in the inventory had no documented owner and no standard procedure. They worked because two or three long-tenured employees knew how to do them. Three of those employees were within five years of retirement. The knowledge transfer risk was concrete and quantifiable. We could put a dollar figure on what would break if any one of them left.
Those three findings reframed the entire AI conversation. Before Phase 1, the company was evaluating AI vendors to make existing work faster. After Phase 1, leadership could see that the bigger opportunity was to eliminate the workaround tax, redesign customer onboarding, and capture tribal knowledge before it walked out the door. Two of those three are not AI problems. One of them becomes an AI problem only after the documentation exists.
What Phase 1 Cost and What It Returned
The engagement was a fixed-fee project. Our piece priced in the low six figures, plus their internal time, which we tracked. The total all-in cost, internal and external, ran roughly 1.0 to 1.2 percent of revenue for the 18-week period.
The opportunity register that came out of Phase 1 carried a conservative estimated annual run-rate value of $2.4M to $3.1M, across automation candidates, process redesign candidates, and headcount avoidance from natural attrition. That value is not realized by completing Phase 1. It is enabled by it. The return is the ability to scope and execute Phases 2 through 5 against real numbers rather than vendor estimates.
Leadership chose to commit to two follow-on initiatives within 30 days of the Phase 1 closeout. A SharePoint and KMS build to capture the process content as the knowledge layer for everything that follows, and a focused redesign of customer onboarding using Power Automate, sequenced before any AI agent work. The AI vendor proposals all went back on the shelf, not because AI was the wrong direction, but because the company now had the data to know which AI investment was actually worth making.
Why This Sequence Matters for AI Implementation Consulting
The reason credible AI implementation consulting starts with documentation and costing is mechanical, not philosophical. Every layer above Phase 1 has a hard dependency on it.
A knowledge management system needs structured content with metadata to be useful. Phase 1 produces the structured content. Without it, the KMS becomes a search index over PDFs, which is what most companies already have.
Operational AI agents need documented workflows to act inside. An agent that drafts a vendor onboarding email is only useful if it follows the actual vendor onboarding process. If that process is not written down, the agent's behavior is invented, and the agent gets distrusted within a month.
A modular cloud ERP needs documented and costed processes to configure correctly. Otherwise the implementation team configures the system to match assumptions, the assumptions turn out to be wrong, and the company spends the post-go-live year fixing it through customizations and reports.
Enterprise AI for forecasting, planning, and strategic decisions needs clean, current operational and financial data flowing out of a coherent system. That flow starts with knowing what the processes are, what they cost, and where their data lives.
Skip Phase 1, and each of the four layers above it gets built on guesses. The guesses cost more to fix than the documentation work would have cost to do correctly. Every distributor we have worked with that started higher on the pyramid has come back to redo Phase 1 within 18 months. The ones that started with Phase 1 are 12 to 18 months ahead of the ones that did not.
What to Expect From an Honest Engagement
If you are evaluating AI implementation consulting partners and the conversation starts with a model demo or a copilot rollout plan, that is the signal that the partner is selling tools, not capability. A credible engagement starts with three questions. What workflows do you run. What do they cost you. Where are the gaps between what you think happens and what actually happens.
For the specialty hardware distributor in this case study, the answer to the third question was the most valuable output. They thought they had a tooling problem. They had a documentation and cost-visibility problem. AI was going to be part of the eventual solution, but it was nowhere close to the first move.
The first six months produce a blueprint. The blueprint determines what gets built. Everything else follows.
Comments