To categorise spend when purchase order and invoice lines are vague, agree a short category list first, clean up supplier names, and apply simple rules for suppliers that only sell one thing. Then let AI suggest a category for the remaining lines with a confidence score. A category owner reviews the low-confidence lines, and each correction becomes a rule.
#Key takeaways
- The category list matters more than the tool; keep it short and owned.
- Supplier name clean-up and simple rules categorise much of the data before any AI.
- AI is good at the long tail: free-text lines, mixed Arabic and English descriptions, and new suppliers.
- A person reviews low-confidence suggestions, and corrections feed back as rules.
- Measure the share of lines left in 'other' and accuracy on a checked sample.
#Why is spend so hard to categorise?
The procurement analyst exports purchase order and invoice lines from SAP, Oracle or Odoo into Excel. Descriptions read 'items as per quotation', 'misc', 'services rendered' or a part number. The same supplier appears under four spellings, one in Arabic. Half the lines were raised after the invoice arrived, with whatever description the requester typed. The biggest category in the report is 'other'.
So the category manager cannot answer simple questions: how many suppliers do we use for cleaning, which business units buy IT hardware outside the framework agreement. The analyst spends days fixing the sheet each time someone asks.
Manual work like this is common: in the Procurement Tactics and Suplari AI readiness survey, checked november 2026, 72.1% of procurement teams described their manual workload as high or very high, with data still moved by hand between ERP, sourcing, contract and reporting systems.
#What should you fix before using AI?
- Agree a category list with two or three levels that matches how your category managers actually buy. Name an owner per category.
- Decide whether you need a standard code set. Public bodies often do: the UK Cabinet Office spend data definitions, checked november 2026 ask for spend to be coded against UNSPSC or CPV. Most mid-sized companies can map later.
- Clean supplier names into one record each, including Arabic and English spellings.
- Write rules for suppliers that sell one thing: the cleaning contractor is always facilities.
- List what is left. That is the long tail AI should work on.
The category list is where many teams go wrong. Coupa, checked november 2026 argues that organisations often do not put enough time and thought into their taxonomy, and that a large 'other' bucket undermines spend analytics.
#What does AI do with the long tail?
For each remaining line, AI reads the description, supplier, GL code, any linked PO or invoice text, and suggests a category with a confidence score and a short reason. It handles mixed Arabic and English descriptions, part numbers it can look up in your item master and new suppliers with no history.
High-confidence suggestions are applied and sampled. Low-confidence ones go to the category owner, who accepts or corrects them. Each correction becomes a rule or an example, so the same line is not reviewed twice. The person owns the category; AI suggests.
#When is a spend analytics tool or a better PO form enough?
If most spend already flows through catalogues and contracts with category codes, fix the purchase request form so requesters pick a category, and stop allowing free-text orders for catalogue items. If you already run a spend analytics module in your procurement suite, use its classification and spend the effort on the category list and supplier clean-up.
AI helps when free-text and after-the-fact orders are common, descriptions mix languages and the analyst repeats the same clean-up for every report.
#How does Stennir run this?
Stennir often finds spend categorisation during AI consulting: the AI Opportunity Audit ranks five workflows, each with a cost and a payback case, and recommends one build. If categorisation wins, AI implementation builds the rules, the AI suggestions and the review queue in your accounts, with category owners approving low-confidence lines and every change logged.
Arabic and English are fully supported, other languages on request. Managed AI Operations then tracks the share of lines in 'other' and accuracy on a checked sample, with a named owner. AI training shows category managers how to review and correct quickly.
We are
Stennir builds the categorisation rules and review queue on your own category list, with category owners approving what AI is unsure about.
We aren't
Stennir is not a spend analytics suite and does not sell a fixed taxonomy you must adopt.
#What should a procurement analyst do next?
Take one recent export, count the lines in 'other' and the distinct spellings of your top suppliers. If both are high, start with the category list and supplier clean-up. See AI for operations and AI for data entry, then take the free readiness assessment or book a 30-minute call.
#Common questions
What do you need from us?
An export of purchase order and invoice lines, your supplier master, your current category list or the one you want, and a category owner for each area who can review suggestions. Read access in your ERP is enough to start; nothing is written back until you approve it.
Does it work with Arabic descriptions?
Yes. Stennir builds categorisation for Arabic, English and mixed lines, and merges Arabic and English spellings of the same supplier using rules you approve. Other languages are supported on request and tested on your own data first. Unclear lines go to the category owner rather than being guessed.
Do we need UNSPSC codes?
Only if a regulator, a parent company or a customer requires them. Most mid-sized companies get more from a short category list their buyers recognise. If you need a standard code set later, each of your categories can be mapped to it once, without recategorising every line.
Where does our spend data stay?
Where your rules require. Hosting is agreed in writing before the build, in your cloud, your country or your existing systems. The data, the rules and the logs stay in your accounts, you keep the off switch, and you decide what any model provider may retain.
