AI-ready data for a manufacturer or distributor
AI-ready data means your ERP, spreadsheets and files agree on every part and price, and each source has an owner.
What AI-ready data means
AI-ready data is company data that an AI system can read, compare across systems and trace back to a source record. Gartner ties readiness to a specific use case, so the same ERP can be ready for one job and not for another.
For a plant or a distributor, test data readiness for AI one job at a time. Can the records behind the job answer the questions your best person answers today, with the same numbers?
The records AI reads
A quote or a margin report reads several records, often kept in different systems. Most of those records sit in the ERP (enterprise resource planning system), in spreadsheets, in drawing folders or in a shared mailbox.
- Item master. Each part or stock-keeping unit (SKU) has a description, a unit of measure, a cost and a status.
- Customers and price lists. Each account carries ship-to addresses, a price level, contract prices and rebates.
- Quotes and their outcomes. Each quote shows what was offered, at what price, and whether the customer ordered.
- Orders and shipments. Order lines hold quantities, promised dates and shipped dates.
- Bills of materials and routings. A bill of materials (BOM) lists what goes into a made part, and a routing lists the operations and hours.
- Purchase history. Receipts show the price actually paid and the date each order arrived.
- Drawings and documents. Drawings and spec sheets carry part numbers and revisions.
- The shared mailbox. Requests for quotes, orders and promises to customers arrive by email.
Definitions that make numbers comparable
Two reports can disagree when every record is correct. One counts a sale on the order date and the other on the invoice date, or one subtracts credit notes and the other does not.
Write each definition in plain words, such as what counts as a sale and how margin is calculated. Include the date that makes an order late and the cost each report uses. Agree on each definition with whoever closes the books, so every report and every AI answer reads the same list.
Gartner’s Rita Sallam made the same point about AI agents in May 2026. Without a clear understanding of the relationships and rules in a company’s data, she said, agents cannot operate accurately.
An owner for each source
Each source needs one named person who decides the right value when two systems disagree. Name the owners in writing, for example the controller for definitions and the purchasing lead for supplier prices.
Without an owner, nobody settles which number is right, and an AI answer has no agreed value to match.
What surveys found about data and AI results
The Business Development Bank of Canada (BDC) published a study of small and medium-sized enterprises (SMEs) in June 2026. The survey behind the study reached 1,500 Canadian business owners and decision makers.
In the BDC study, 94% of SMEs with fully integrated data were satisfied with the return on their AI investment. The figure was 63% for SMEs with partially integrated or unintegrated data. BDC describes integrated data as data grouped into a single system or environment.
Where AI drew on internal data, 85% of SMEs were satisfied with the return, against 64% where AI used only external data. Almost half of SMEs with digital data, 47%, had data that was not integrated at all or only very partially, according to BDC.
Gartner surveyed 1,203 data management leaders in July 2024. In that survey, 63% of organizations either did not have the right data management practices for AI or were unsure whether they did. Gartner also predicted, in February 2025, that through 2026, organizations would abandon 60% of AI projects that lack AI-ready data.
RSM’s 2026 survey of middle market companies asked manufacturers with moderate or limited AI pilot success why pilots failed. Data quality issues came first, at 51%, and integration challenges second, at 43%.
The data does not have to be perfect. PwC surveyed 767 operations and supply chain leaders at US companies for the 2026 Digital Trends in Operations Survey. Of those leaders, 87% said poor data quality had hampered progress toward value from digital initiatives. In the same survey, 73% agreed that data does not need to be perfect to drive value.
PwC reports that some companies succeed by improving data quality in a targeted way, step by step, alongside the wider change. For a smaller company, that means fixing the records the first AI job reads before touching the rest.
An AI-ready data checklist for a manufacturer or distributor
Answer each check yes or no for your company. Each check names a record that someone can open this week. A distributor can skip the two BOM and routing checks.
Count the yes answers in each group. Start with the groups that the first AI job reads, even when another group has more no answers. To score twelve related questions in your browser, take the AI readiness assessment for a plant.
Item master
- Every active part has one part number, one description and one stocking unit of measure.
- Every active part has a current cost, and the ERP records the date of the last cost update.
- Duplicate and inactive parts are flagged, so a search for a part returns one live record.
- Each part stores a conversion factor between purchase and sales units, such as one box equals 100 pieces.
Customers and prices
- Each customer has one account, and every department uses the same account number.
- Every price a customer pays comes from the ERP or one controlled price file, never from a personal spreadsheet.
- Contract prices and rebates carry a start date and an end date.
Quotes and orders
- Every quote has a status of open, won or lost, with the date of the decision.
- Each won quote links to the resulting order through the quote number.
- Order lines record the promised date and the shipped date.
Bills of materials and routings
- Each made part has one current BOM revision and one current routing, marked as the version in use.
- Labour hours land against a job and an operation, not only against a payroll week.
Purchasing
- Purchase receipts record the price actually paid, line by line, against the part number.
- Each supplier has one record, and each purchased part records a current lead time.
Drawings and the shared mailbox
- The file name of every current drawing or spec sheet carries the part number and revision.
- Customer requests and orders arrive in a shared mailbox, not only in personal ones.
Definitions, owners and exports
- A one-page list defines the terms that reports use, such as a sale and a late order, and the controller has agreed to the list.
- Each source has one named owner who decides the right value when two systems disagree.
- Someone inside the company can export a year of records from each source without calling the vendor.
Start with one report your team rebuilds by hand
Tell Derik which ERP you run and which report someone rebuilds in Excel every month. He will tell you what a four-week reporting sprint would cover, from the reporting database to the test set.
Name the reportCommon data problems and how to fix them
The problems below produce AI answers that look plausible and are wrong. Every fix starts in the system that already holds the record.
Duplicate parts
The same part can sit in the ERP under two numbers, for example one from a supplier catalogue and one typed by hand. Sales history then splits between the two records, so an AI answer about last year’s volume comes out low.
Pick the record that survives, and map every duplicate to that record in a cross-reference table. Mark the duplicates inactive instead of deleting them, and require a search before anyone creates a new part.
Units that differ between purchasing and sales
Purchasing buys a fastener by the box of 100, and the sales desk sells the fastener by the piece. Without a conversion factor in the ERP, a margin calculation can compare the cost of a box with the price of one piece.
Store the conversion on the part record, with the stocking unit as the base. Metal bought by weight and sold by length needs the same kind of factor for each size. Then check that every report converts to the stocking unit before the report adds or divides anything.
Prices kept in spreadsheets
Contract prices and rebates sometimes live in a sales rep’s spreadsheet instead of the ERP. An AI system that reads the ERP then quotes list price to a customer with a contract. The rep’s file may also hold deals that expired months ago.
Move each customer price into the price tables of the ERP, or into one controlled file with a named owner. Give every price a start date and an end date, so expired deals stop applying.
Quotes with no won or lost status
When quotes stay open forever, nobody can tell which prices won work. An AI system that drafts quotes from history then copies prices that lost as readily as prices that won.
Add a status of open, won or lost to every quote. Close a quote as lost after a set number of days without an order. Backfill the last year by matching quotes to orders on the quote number, or on the customer and the part.
Drawings not linked to part numbers
A drawing saved as “bracket final v3.pdf” in an engineer’s folder matches no part number in the ERP. An AI answer about that part then cites the wrong revision, or no drawing at all.
Rename each current drawing to the part number and revision, and keep the files in one vault or shared folder. Record the current revision on the part in the ERP, so the drawing and the record agree.
An older ERP on a local server
An older ERP version installed on a server in the building may have no API, the programming interface that other software uses to read data. The records are still usable.
A read-only database login or a scheduled export can copy the records out without changing anything in the ERP. Legacy ERP automation covers the other direction: entering orders into an old ERP through the same screens staff use, after a person approves each entry.
An illustrative example: a distributor runs the checklist
This example is illustrative and describes no client. An industrial supply distributor wants AI to help the inside sales team, and runs the checklist before choosing the first job.
The two BOM and routing checks do not apply to a distributor, which leaves 17. The team answers yes to 10 of the 17.
| Group | Yes answers | What the team found |
|---|---|---|
| Item master | 3 of 4 | Two supplier catalogues created duplicate SKUs. |
| Customers and prices | 1 of 3 | Contract prices sit in sales reps’ spreadsheets, with no end dates. |
| Quotes and orders | 1 of 3 | Quotes carry no status and no link to orders. Order lines do record promised and shipped dates. |
| Purchasing | 2 of 2 | Receipts record the price paid, and each supplier has one record. |
| Drawings and the shared mailbox | 1 of 2 | Spec sheets keep the supplier’s file names. Requests arrive in a shared mailbox. |
| Definitions, owners and exports | 2 of 3 | Owners are named and exports work, but nobody has written down what counts as a sale. |
The team then counts the checks behind each candidate job.
| Job | Checks the job reads | Yes answers |
|---|---|---|
| Answer stock and order-status questions from the shared mailbox | 6 | 5 |
| Draft quotes at each customer’s price | 10 | 4 |
| Report margin by customer | 9 | 5 |
The stock and order-status job goes first, with 5 yes answers out of 6. The one gap is the duplicate SKUs, which would understate stock on hand, so the team merges the duplicates before go-live.
Before anyone relies on the system, the inside sales team collects 30 questions from last month’s mailbox and records the right answers. The system has to give the same answers, and a named person approves each reply before the reply reaches a customer.
Merging the duplicates also helps the other two jobs. The margin report then still needs prices in the ERP, start and end dates on contract prices, and a written definition of a sale. The quoting job needs the same fixes, plus a status on every quote and a link from each won quote to the order.
The margin report goes second, after the price work and the written definition, and quoting goes third.
The data warehouse approach
A data warehouse is one database built from the records in the systems a company already runs. Copies of the records in the ERP, the accounting system, the price spreadsheets, the drawing index and the shared mailbox arrive on a schedule. The ERP stays the system of record.
A data warehouse groups a company’s records in a single environment, which is how BDC defines integrated data.
- Read-only connections. The warehouse copies records and never writes back, so nothing in the live systems changes.
- Written definitions beside the data. Every report and every AI answer reads the same definition of a sale or a late order.
- The company’s own server. The warehouse sits on a server dedicated to the company, in Canada or the United States.
- Answers that name the source. A chat answers from the warehouse and shows the query and the rows behind each number. When the warehouse holds no answer, the chat says so.
- A test set. Questions with known answers rerun after every change, so a number that was right stays right.
AI reporting shows how the same database answers report requests in plain language. AI for ERP covers the ways to connect to the ERP, and legacy ERP automation covers older ERPs with no API. To score your own sources first, take the AI readiness assessment.
How ThriveAI helps
ThriveAI builds private AI systems on a company’s own data for manufacturers and distributors in Canada and the United States. Working sessions run by video call or on site.
The platform ThriveAI builds on is designed to keep each client’s data on a server dedicated to that client, in Canada or the United States. The client chooses where the model runs. One option is an open-weight model, a model whose files are published, on that same server. The other is a frontier model under a written zero-data-retention control, with the model tier named. Under that control, the provider keeps no copy of prompts or answers beyond the exceptions the provider lists. A hosted model may process requests in another country.
A named person approves every action before anything is sent or saved.
One way to start is the AI reporting sprint. In four weeks, the sprint copies the ERP and accounting records read-only into one reporting database. The sprint also writes the definitions with the controller and builds a test set of known questions. The database stays with the company, and later modules read from the same records.
To check your records on your own first, take the AI readiness assessment. The assessment scores twelve questions in your browser and shows the result without asking for an email.