AI implementation in a smaller plant or distributor: from first use case to daily use
AI implementation is the work of taking one AI use case from an idea to a tool your team uses every day. At a smaller manufacturer or distributor, it runs in six steps: get access to the data, set up a private environment, put approvals in place, test the system on your own past work, train the team and measure the result. This guide covers each step, the points where AI projects fail according to published research, and how long the steps take where a published source says so.
What AI implementation covers
A use case is one specific job AI does, such as reading emailed purchase orders and drafting the orders in your ERP, the enterprise resource planning system that holds orders, stock and costs. Implementation starts once you have chosen that job, which AI strategy and roadmap covers, and it ends when the tool is part of the team’s normal day.
Two terms come up often. A pilot, also called a proof of concept, is a small test of whether the idea works on real data. Production means the tool runs on live work every day. Gartner measured the gap between the two in a survey of 644 respondents from organizations in the United States, Germany and the United Kingdom, conducted in the fourth quarter of 2023: “on average, only 48% of AI projects make it into production, and it takes 8 months to go from AI prototype to production.”
The table sets out the six steps.
| Step | What happens | Who takes part on your side | Done when |
|---|---|---|---|
| 1. Data access | Read-only connections to the systems the job uses | The owner of each system and whoever manages IT | The system reads a real sample of the records |
| 2. Private setup | Decide where data is stored, which model reads it and what the provider keeps | The owner and the IT lead | Storage, model and retention setting are written down |
| 3. Approvals | Drafts wait for a named person before anything is sent or saved | The people who own the job today | Every kind of output has a named approver |
| 4. Tests on your data | Real past jobs with known answers run through the system | The person who does the job today | The system meets the written pass rules |
| 5. Training | The team learns to run, check and correct the tool on its own work | Everyone who will use it | Each user has run real work through it |
| 6. Measurement | The business number is compared with the baseline | The owner | You decide to extend, change or stop |
Step 1: Get access to the data
The system needs to read the records the job uses, such as the ERP, the inbox where orders arrive, the shared drive and the drawings. Ask for read-only access first, so the system can look and cannot change anything. Give it its own service account, which is an account for a program with its own password and permissions, so you can see what it did and switch it off without touching anyone’s login.
How the connection works depends on the system. A current ERP may offer an API, a connection one program uses to call another, or a scheduled export. An older ERP installed on your own server may offer neither, and legacy ERP automation covers the options for that case. The Model Context Protocol (MCP) is “an open-source standard for connecting AI applications to external systems,” according to its maintainers, and the MCP server guide explains it. AI for ERP covers what works with the ERP you already run.
Before anything is built, pull a real sample of records and check them by hand. Part numbers, customer names and prices should match across systems.
Step 2: Set up a private environment
Decide where the data is stored, which model reads it and what the model’s provider keeps. A model is the AI program that reads a request and writes the answer. A hosted model runs on the provider’s servers, and under a zero data retention agreement the provider keeps no copy of a request or its answer. A model can also run on a server you control, which keeps each request on that machine.
Write the answers down: the storage location, the model by name, the region where it processes requests and its retention setting. The Canadian Centre for Cyber Security’s top 10 AI security actions, published in May 2026, include “Enforce data privacy, vendor and contractual controls.” Private AI for business compares the setups, data sovereignty in Canada covers what Canadian law requires, and AI security covers the rest.
Step 3: Put approvals in place
Set the system to draft, and have a named person approve each draft before anything is sent to a customer or saved in the ERP. This practice is called human in the loop, and the Cyber Centre’s list includes it: “Ensure that human-in-the-loop oversight and execution controls are in place.”
Decide who approves each kind of output. The estimator approves quotes, the buyer approves purchase orders, and whoever sends a customer email approves its text. Record each correction, because corrections become new test cases in step 4. Human in the loop shows how to design an approval step people actually read.
Start with one job
Tell Derik which job you want AI to take on and which systems it touches. He reads every enquiry himself and usually replies within one business day.
Start a conversationStep 4: Test the system on your own data
Before launch, run real past jobs with known answers through the system and grade each output against written pass rules. This test is called an eval, short for evaluation. Rerun it after every change to the prompt, the model or a connected system. The prompt is the written instruction the system gives the model. AI evals shows how to build the test set and the pass rules.
Then run the system alongside the person who does the job today, which is sometimes called shadow mode. The system drafts, the person works as usual, and you compare the two. When the drafts meet your pass rules, the person switches to approving them.
Step 5: Train the team
Train the people who will use the tool on their own work: how to run it, how to check a draft, how to correct it and what to do when it is wrong. In Statistics Canada’s survey for the second quarter of 2026, 32.0% of businesses that had used AI in the previous 12 months reported AI-related training for existing employees. Among those with 100 or more employees the share was 68.1%, and among those with 1 to 4 employees it was 24.0%. AI training covers hands-on training on your own documents.
Step 6: Measure the result and move to daily use
Compare the business number with the baseline you recorded before the project started, such as minutes per order or days per quote. The same Gartner survey found: “The primary obstacle to AI adoption, as reported by 49% of survey participants, is the difficulty in estimating and demonstrating the value of AI projects.” A baseline recorded before the work starts gives you a number to show.
Keep watching after launch. The Cyber Centre’s list asks organizations to “Maintain operational resilience against model drift, hallucinations, bias and overreliance.” A hallucination is an answer the model makes up, and drift is a gradual change in results as the inputs or the model change. Review a sample of live work each week and add every corrected case to the test set. What the system costs to run belongs in the same review, and the cost of AI for a small business breaks that cost into parts.
How long the steps take
This guide found two published sources that give durations, and neither is specific to a smaller plant. Gartner’s survey found an average of 8 months from AI prototype to production. RAND researchers, in a 2024 report on why AI projects fail, advise that “leaders should be prepared to commit each product team to solving a specific problem for at least a year.” The rest depends on the job, the state of the data and how much of your team’s time the project gets.
Where AI projects fail
For that RAND report, researchers interviewed 65 experienced data scientists and engineers and identified five leading root causes of failure. In July 2024, Gartner predicted that at least 30% of generative AI projects, meaning AI that produces text or images, such as ChatGPT, would be abandoned after proof of concept by the end of 2025, “due to poor data quality, inadequate risk controls, escalating costs or unclear business value.” The table lists those failure points and the step in this guide that guards against each one.
| Failure point | Named by | The step that guards against it |
|---|---|---|
| People misunderstand or miscommunicate the problem AI should solve | RAND | Choosing the bottleneck in the strategy, and the baseline in step 6 |
| The organization lacks the data needed to train an effective AI model | RAND, and Gartner on data quality | The sample check in step 1 |
| The latest technology comes before the user’s real problem | RAND | Building with the person who does the job, in steps 3 and 4 |
| The infrastructure to manage data and deploy finished models is missing | RAND | Steps 1 and 2 |
| The problem is too difficult for AI | RAND | The tests on your own past work in step 4 |
| Inadequate risk controls | Gartner | The private setup and approvals in steps 2 and 3 |
| Escalating costs or unclear business value | Gartner | The measures and review in step 6 |
Make sure the first one is in daily use, has a named approver and has a number you can compare with the baseline. AI consulting covers what to ask if you bring in outside help for the next one.
Questions people ask
What is AI implementation?
How long does AI implementation take?
Why do AI projects fail?
What is a pilot in AI implementation?
Who should approve what the AI produces?
How should we train staff on a new AI tool?
How ThriveAI helps
ThriveAI is an AI engineering company in Ottawa that builds private AI systems for manufacturers and distributors in Ontario and Quebec, on their own data. It works on site with your team and builds each tool on your own data, and the people who will use the tool shape it while it gets built, so it fits how your team already works. Derik Lawlis, the founder, leads every project and stays close to the build.
ThriveAI’s systems read your ERP and your other software as they are, including older versions installed on your own server. Every connection only reads data, and nothing touches a live system until you approve what it will read. Each number links to the document or ERP record it came from, and nothing is sent or saved in your systems until the person responsible approves it.
The platform ThriveAI builds on is designed to keep each client’s data on its own server in Canada. You choose the AI model that reads it: one on that server, or a hosted model under a written agreement that the provider keeps nothing after answering. A hosted model may process requests outside Canada, so the contract names the model. Stage one is a working tool for the one job that costs you the most, at a fixed price, in weeks, and if you stop there, you keep it. About ThriveAI covers the company.