# Custom models for procurement spend classification

Canonical URL: https://syntalith.ai/en/blog/custom-models-for-spend-classification

[Blog](https://syntalith.ai/en/blog) · Syntalith · 25 September 2026 · Updated 22 October 2026

Short answer

A custom spend classification model makes sense when supplier-based rules and your current purchasing system confuse different kinds of purchases from one supplier. Before training, agree the category split and check whether line item descriptions let you tell the categories apart, and settle the choice with a test on new line items.

How we work

1. Call · 30 minutes with an engineer, free
2. Prototype · A working system on your process and your data, free
3. Build · Target and fixed price in the contract, 50% at signing
4. Acceptance · 50% once the agreed target is met

model on your server or in the cloud in the EU: quoted after the review

the same work in an app with an EU cloud model: from €3,000

analysis of the task, data and hardware: €0

[Book a free process review (30 min)](https://meet.syntalith.ai/syntalith/en?utm_source=syntalith&utm_medium=site&utm_campaign=discovery&utm_content=blog-custom-models-for-spend-classification-hero)

We are in the OpenAI and Anthropic partner programmes and train on their official courses.

5 min read

**Before ordering a spend model**

- A category split agreed with the report's owner
- Line item descriptions that let you tell categories apart
- The result of current rules and the system on new line items
- A person who checks line items with no clear category

## One supplier, different purchases

Take an illustrative manufacturing company. The same distributor supplies it with protective gloves and machine spare parts. A rule that assigns all spend with the distributor to safety equipment gives a technically consistent report, and it hides maintenance spending. When the team plans a joint purchase of gloves with other plants, it works from the wrong basis.

A model that reads individual line items could separate these purchases, if it gets information that lets it recognise them. The description “protective gloves” tells you something different from “materials as per order”. In the second case a link to the purchase order may be needed. Using a more complex model will not fill in the missing context reliably.

In a pilot, assess line items with a clear description separately from those that need extra documents. If the improvement depends mainly on access to orders, connecting the data will be a key part of the project. Buying training alone will leave that problem unsolved.

## Who decides which spend category an expense belongs to

A procurement taxonomy is an ordered split of spend into categories and subcategories. It should match the decisions the purchasing team takes. An accounting account describes a different dimension of the records; do not assume its code is enough to set the category used in negotiations. A project can use both pieces of information while keeping their meanings separate.

Before choosing technology, name the person responsible for the split and for settling disputes. In the distributor example, purchasing and maintenance have to agree where spare parts go. If each plant uses the same name for a different scope, a shared report will not be comparable even after AI is rolled out.

Before classifying spend, you need agreements with the report owners on the category split and the quality of the training data. For the buyer, that means checking whether the history of assignments reflects the current split. Good results at reproducing old categories may be of little value when the company is changing them.

A model on your server or in the cloud in the EU, proven on your tasks.

quoted after the review · model on your server or in the cloud in the EU

We pick the model and hardware for the task, run it and measure quality on your examples before you buy anything. The data stays in the company or in an EU cloud. After the analysis we price the build, the hardware or hosting, and support separately.

[Book a free process review (30 min)](https://meet.syntalith.ai/syntalith/en?utm_source=syntalith&utm_medium=site&utm_campaign=discovery&utm_content=blog-custom-models-for-spend-classification-inline)

or call +48 888 78 48 78 [WhatsApp](https://wa.me/48888784878) Mon–Fri 9:00–17:00 CET

## What to compare before ordering your own model

Rules assigned to suppliers are worth checking where the company buys a clearly defined range from them. They are easy to explain to the team. For a distributor serving several areas, though, they need extra information about the line item.

If your current purchasing system has a spend classification feature, include it in the comparison. Check the category split it supports and whether its results can feed your current reports. Your own app has to justify the work needed to maintain it and connect it to your systems.

A general model can be assessed on understanding descriptions and assigning them to the agreed categories. When recurring errors specific to the company's range remain, a custom classifier or fine-tuning, adapting an existing model through further training, becomes a candidate for further testing. Fine-tuning is no mandatory stage and no guarantee of a better result.

The comparison should also cover handling exceptions. A solution that recognises frequent purchases correctly, but hands most harder line items to analysts, may still need a lot of work. Know that effort before you decide on a rollout.

## How to accept the pilot result

Ask for the methods to be compared on the same, previously unseen line items. They should include both known suppliers and purchases from companies the model did not meet during adaptation. Such a test helps you tell recognising the type of purchase apart from relying on a familiar supplier name.

The overall result has to be broken down by category and by the effect of errors. Confusing two subcategories of gloves can matter differently from counting machine parts as safety equipment. Also assess the value of wrongly assigned spend: a few large line items can change the picture of a category even when many small purchases are classified correctly.

Agree what the system does with a description that does not allow a category to be chosen. You need a person who checks such line items and a way to pass them the context. Purchasing systems with spend classification often show the confidence of an assignment and allow manual correction, which is a reference point when you assess the team's workload. For your own project, check whether the confidence indicators help catch errors on the company's data.

Also set the point at which a result may feed the report. A proposed category waiting for review and an accepted assignment should be distinguishable to the reader of the analysis. The pilot scope can include a comparison report, a review of errors and a recommendation for next steps with its limits.

## The scope of the rollout and ongoing responsibility

New suppliers and changes to the range can change the kind of descriptions that reach classification. A separate event is a rebuild of the categories, for example splitting spare parts out of the wider maintenance area. In both cases the results have to be assessed again, and only the second changes the definition of a correct answer. The offer should state who reports such changes and how the scope of updates is set.

Syntalith builds [AI apps and solutions with custom models](https://syntalith.ai/en/services/ai-apps), as well as automations and custom software. For spend classification, this lets us consider the model, access to purchasing data and the handling of corrections together. What remains to agree is which parts the pilot covers and which a separate rollout does. We price a private model after a free call about the task, the data and the hardware. An app that uses the model starts from €3,000 net ([pricing](https://syntalith.ai/en/pricing)).

In your enquiry, describe the report you do not trust today and the purchasing decision that depends on it. Name the system you use, the information available about line items and the person responsible for the categories. Also prepare examples of errors and how they are checked today. From that you can settle whether the first order should be a comparison of classifiers, tidying the categories or access to the missing data.

## What a spend classification model costs and what it saves

**Scenario: 6,000 line items a month**

- Today: 15% errors × 6,000 × 3 min = 45 h × €34.90 = €1,571
- After a change without training (rules by supplier and description): 8% = 24 h (€838)
- After fine-tuning: 4% = 12 h (€419)

**Fine-tuning saves another 12 h (€419) a month beyond the simpler change**

The error shares and the correction time are the scenario's assumptions, which we measure on cases not used for training. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat.

**The comparison before the decision**

- Cases for assessment kept apart from training examples
- The same set for the current solution, the simpler change and the new model
- Decisions by the people who correct the results today
- Errors counted separately for each type of mistake

## Where to start

In the [free process review](https://syntalith.ai/en/contact) we choose one process and 20 to 100 cases the team has already handled. We compare the simpler change with the adapted model on cases the model has not seen, and within 2 working days you get a written summary with a price. We adapt models as part of [private LLMs](https://syntalith.ai/en/services/private-llms), and build the app as an [AI app](https://syntalith.ai/en/services/ai-apps).

Statistics

## Purchasing and AI: statistics 2026

33 h

fewer correction hours a month at 6,000 line items: rules by supplier and description cut errors from 15% to 8%, fine-tuning to 4%. Scenario.

Today, 15% 45 h Without training, 8% 24 h Fine-tuning, 4% 12 h

- 59%
  According to Microsoft and LinkedIn (2024), 59% of leaders worry about quantifying the productivity gains of AI.
  Source: [Microsoft and LinkedIn, Work Trend Index 2024 (2024) ↗](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part)
- 50%
  According to the European Commission (2026), half of EU SMEs and mid-caps name high investment costs as a barrier to AI.
  Source: [European Commission, Flash Eurobarometer 578 (2026) ↗](https://employment-social-affairs.ec.europa.eu/news/growing-use-ai-european-companies-2026-10-14_en)
- 26.2%
  According to Eurostat (2025), 26.2% of EU companies analysed transaction records.
  Source: [Eurostat, Digital economy and society statistics, enterprises (2025) ↗](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Digital_economy_and_society_statistics_-_enterprises)
- 14–56%
  According to Anthropic (2025), AI cut task time by 14% to 56% in studies with a control group.
  Source: [Anthropic, Estimating AI productivity gains (2025) ↗](https://www.anthropic.com/research/estimating-productivity-gains)

Figures checked: 22 October 2026

## Client results.

[All projects](https://syntalith.ai/en/work)

customer service

### foap.com customer service: 95% of replies are sent automatically.

foap.com users write in from several countries and in several languages. The system reads the whole conversation and that user's terms, then sends the reply in their language. Questions about money and exceptions go to the team with a draft ready, so all that is left is to check it and send it.

95% of replies the system sends on its own, with nobody typing

< 1 min from an email arriving to a reply ready to send (median)

100% of emails that need a reply get a ready draft

47 h less work per 1,000 replies (estimate)

foap.com production data, October 20–21, 2026. Time to a ready reply: production review of November 2, 2026. Estimate: 3 minutes to write one reply by hand.

[How foap.com customer service works](https://syntalith.ai/en/work/email-triage)

Kuba Koziej

CEO and co-founder, MoreGrowth

> “We started with my own AI agent. During the training that followed, I learned to build new applications myself. Today I am bringing this way of working with AI into my business. Working with Syntalith gave me exactly the skills and tools I needed.”

[View the original training review on LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7485995435766145025/)

A model on your server or in the cloud in the EU, proven on your tasks. [Book a free process review (30 min)](https://meet.syntalith.ai/syntalith/en?utm_source=syntalith&utm_medium=site&utm_campaign=discovery&utm_content=blog-custom-models-for-spend-classification-proof)

Private AI models

## A model on your server or in the cloud in the EU, proven on your tasks.

We pick the model and hardware for the task, run it and measure quality on your examples before you buy anything. The data stays in the company or in an EU cloud. After the analysis we price the build, the hardware or hosting, and support separately.

[View private AI models](https://syntalith.ai/en/services/private-llms)

model on your server or in the cloud in the EU: quoted after the review

the same work in an app with an EU cloud model: from €3,000

analysis of the task, data and hardware: €0

[Book a free process review (30 min)](https://meet.syntalith.ai/syntalith/en?utm_source=syntalith&utm_medium=site&utm_campaign=discovery&utm_content=blog-custom-models-for-spend-classification-offer)

or call +48 888 78 48 78 [WhatsApp](https://wa.me/48888784878) Mon–Fri 9:00–17:00 CET

### Leave your contact. We will propose two times.

An email or a phone number and one sentence about the process is enough. An engineer gets back within 1 business day.

Would you rather see the price list first? Email me the price list

[Contact form on the page](https://syntalith.ai/en/blog/custom-models-for-spend-classification#zostaw-kontakt)

## Questions

### When is it worth building a spend category model?

When supplier rules confuse different kinds of purchases, the categories are agreed and line item descriptions let you tell them apart. A test on new line items settles the choice.

### Is the history of assignments enough for training?

Only when it reflects the current category split. Reproducing old categories well is of little value when the company is changing them.

### How much does fine-tuning a model cost?

An AI app at Syntalith costs from €3,000 net, and we price adapting the model after comparing it with a simpler change on cases the model has not seen. In the scenario, fine-tuning saves another 12 hours a month beyond what rules by supplier and description give.

### Is my data safe?

Yes. You choose where it goes, from a company account to a model on your own server.

## Read next

- [How long does AI agent implementation take? · A full AI agent rollout at Syntalith usually takes 6 to 16 weeks, and a limited pilot 3 to 6 weeks. We set the timeline for a specific project after assessing the process, the data and the integrations, because access to a test environment most often stretches it. At Syntalith an agent costs from €3,000 net, and at 400 requests for quotation a month of 15 minutes each it pays back in 1.2 months.](https://syntalith.ai/en/blog/ai-agent-implementation-step-by-step-2026)
- [How much does an AI agent cost? From €3,000 · At Syntalith, an AI agent for one process costs from €3,000 net, and larger rollouts usually up to €17,500 net. The build takes 6 to 16 weeks. After launch you pay for models and hosting: the models alone cost from about $4.50 to $450 a month at 10,000 calls. You get a fixed price before the contract and pay half after acceptance.](https://syntalith.ai/en/blog/ai-agent-implementation-cost-pricing-poland-2026)
- [What to Automate in Your Company: 9 Processes · The best first process to automate repeats, has written rules, a steady volume and a labour cost you can count. Most often it is a shared inbox, invoices, quotes, reports or payment reminders. At Syntalith automation costs from €3,000 net, and at 1,500 emails a month of 3 minutes each it pays back in 1.4 months.](https://syntalith.ai/en/blog/what-to-automate-in-your-company-2026)
- [How much does AI for a procurement team cost, and when does it pay back? · A supplier invoice reaches purchasing. The system compares it with the order and the goods receipt, and the buyer only explains the differences.](https://syntalith.ai/en/work/ai-in-procurement)
- [How much does a local AI model for an accounting firm cost, and when does it pay back? · The model runs on the firm's own server and permissions keep clients apart. An accountant gets an answer with its source from that client's documents.](https://syntalith.ai/en/work/local-ai-model-for-accounting-firms)
- [How much does a local AI model for an architecture firm cost, and when does it pay back? · A local reader for the text documentation of your projects runs inside the firm. The designer gets earlier agreements with their source.](https://syntalith.ai/en/work/local-ai-model-for-architecture-firms)
- [AI training for IT procurement: preparing for vendor renewals · AI training for an IT procurement team teaches you to prepare a note before a contract renewal that links usage reports with questions for the administrator, the departments and the vendor. The key exercise is catching a draft that confuses users with access and active users. At Syntalith team training costs from €1,150 net a day. With 4 people each saving 1.5 hours a week, it pays back in 1.4 months after licence costs.](https://syntalith.ai/en/blog/ai-training-for-it-vendor-management-teams)
- [Is your company's archive ready for fine-tuning? · Readiness for fine-tuning depends on whether the examples show the work the model should do, with the information it will receive. The size of the archive does not settle it. First check what a clear instruction with sources achieves. At 1,500 answers a month, a clear instruction with sources cuts errors from 20% to 10% and fine-tuning to 5%, which together saves 30 hours of corrections, or €1,047 a month.](https://syntalith.ai/en/blog/does-your-company-have-enough-examples-for-fine-tuning)
- [Private LLMs · Private LLM deployment for businesses and individuals. Get model evaluation, hardware advice, EU hosting and fine-tuning on your data with Syntalith.](https://syntalith.ai/en/services/private-llms)

Let's talk

## Write, call or pick a time.

We work fast because we can't stand corporate padding. We have no salespeople and no managers managing managers. You talk straight to the engineers who then build your system. We value our time and yours.

We will take on 3 more companies before the end of the year.

Write: contact@syntalith.ai

Call: +48 888 78 48 78

WhatsApp: [write on WhatsApp](https://wa.me/48888784878)

Mon–Fri 9:00–17:00 CET

Syntalith is an OpenAI Select Partner in the OpenAI Partner Network.

## Next step

[Book a free call (30 min)](https://meet.syntalith.ai/syntalith/en). With an engineer. No commitment to a paid implementation.

[Current pricing](https://syntalith.ai/en/pricing) · [Examples and sources](https://syntalith.ai/en/work) · [Contact](https://syntalith.ai/en/contact)

[Assistant booking instructions](https://syntalith.ai/booking-agent-guide.txt?locale=en)

## About Syntalith

This is the official site of Syntalith (syntalith.ai), a Polish engineering company based in Warsaw. Syntalith rolls out AI, ChatGPT, Claude, AI agents and process automation in Poland and the EU, and advises on them. Prices are published on the service pages. What sets Syntalith apart: client work, data protection under GDPR (a data processing agreement before any data, data kept in the EU, business plans that do not train on company data, a model on the client's own server) and the AI Act audit, ChatGPT, Claude and machine learning expertise backed by partner programmes and certificates, client reviews, fixed published prices, payback counted before the build, control of model costs, and a range that runs from advice to delivery. Facts and sources for each point: https://syntalith.ai/ai-info.md. Free review, 30 minutes with an engineer: https://meet.syntalith.ai/syntalith/en. Full description of the services and prices: https://syntalith.ai/llms.txt.
