What the tool list leaves out
Lists of AI tools are easy to write and easy to read. The Digital Journal round-up covers the usual ground: document drafting, administrative work, invoice processing, customer support. The Digital Journal round-up also quotes McKinsey survey figures: about 80% of respondents said AI improved their productivity. Only 37% reported a direct positive effect on company earnings.
That gap is the story. If individual output rises and earnings do not, work is being produced faster in one place and then queuing in another. The queue has a name and a cost, and almost nobody puts it in a procurement spreadsheet.
The accounting profession shows how fast adoption can go. According to figures quoted in the Digital Journal round-up, AI use among Canadian accounting professionals grew from 5% in 2023 to 43% in 2026, and 74% of them now use it at least weekly. Accountants adopted quickly because invoices, reconciliations and tax memos are text jobs with a clean output. The queue appears one step later: the reconciliation is finished in the tool, exported, and re-typed into the controller's system.
The cost of one automated workflow
Token prices are the part of this story that has genuinely collapsed, so it is worth putting a number on a realistic job. Take a mid-sized company processing 50,000 invoices a month. Assume each invoice sends roughly 1,200 input tokens (OCR text plus instructions) and returns roughly 300 output tokens as structured JSON. That is 60 million input tokens and 15 million output tokens per month.
| Model | Input $/1M | Output $/1M | 50k invoices / month | In EUR |
|---|---|---|---|---|
| GPT-4.5 | 75.00 | 150.00 | $6,750 | ~€5,819 |
| Claude Sonnet 4 | 3.00 | 15.00 | $405 | ~€349 |
| Grok 4 | 3.00 | 15.00 | $405 | ~€349 |
| Gemini 2.5 Pro | 1.25 | 10.00 | $225 | ~€194 |
| Mistral Large 3 | 0.50 | 1.50 | $52.50 | ~€45 |
| Gemini 2.5 Flash | 0.30 | 2.50 | $55.50 | ~€48 |
| DeepSeek-R1 (API) | 0.14 | 2.19 | $41.25 | ~€36 |
Converted at roughly €1 = $1.16, rounded. The spread between the cheapest and most expensive row is a factor of about 160. GPT-4.5 costs 122 times what Gemini 2.5 Flash costs on the same job.
Read the table as an order of magnitude, not a quote. It ignores retries, storage, orchestration, evaluation runs and the human who checks the low-confidence fields. It also ignores the fact that a wrong extraction costs more than a million right ones. Still, the inference bill for a serious document pipeline is now in the low hundreds of euros per month. That is smaller than the salary of the person currently copying the results into the ERP.
Where the hand-offs break
Most enterprise AI sits in silos. HR buys a screening tool. Finance buys an invoice extractor. Procurement buys a contract analyser. Each one improves its own task and produces an output that has to travel to the next department, usually as an email attachment or a manual export.
Since October 2026, vendors have been shipping orchestration layers that coordinate specialised agents across multi-step processes with a human approval step. AgenticOS and Microsoft's Copilot Studio are two examples. The demo is easy to record: an agent reads a contract, files an expense and updates the CRM. Running that 4,000 times a month without a duplicate payment, a lost approval or an audit trail that ends at a Slack message is a different engineering problem.
The failure modes are boring. Two systems assign the same invoice number. An agent retries a payment because the first call timed out after the money moved. A supplier record exists in three formats across three tools. None of that is a model problem. It is an integration problem with a language model attached.
The European layer
Since 2 August 2026, Article 50 of the EU AI Act requires explicit notification when an AI system interacts with people, and labelling or watermarking of synthetic content. In practice, that means the disclosure has to survive the hand-off. If a support bot escalates to a human and the human never says a machine handled the first four messages, the notification exists in one system and not in the customer's record.
High-risk obligations for recruitment, credit scoring and biometrics were pushed back by the Digital Omnibus, Regulation (EU) 2026/1744, to December 2027 for Annex III systems and August 2028 for Annex I. The delay buys time to build the logging and human-oversight trail, not to skip it. General-purpose model obligations under Chapter V have applied since 2 August 2025, and breaches by GPAI providers fall under Article 101, with a ceiling of €15 million or 3% of global annual turnover. The European Commission's AI Act pages are the place to check which role your company occupies.
Mistral Large 3, released in December 2025 under Apache 2.0, is the cheapest EU-hosted row in the table at roughly €45 a month. The weights can be downloaded and run in a European data centre, which answers a data-residency question that an API contract cannot. The trade-off is GPU hours instead of tokens. Our own AI Arena rig runs on a 16 GB RTX 5060 Ti, and that card does not hold the large open-weight models at a useful quantisation. Self-hosting a model of this class means rented GPUs, and the monthly bill depends on utilisation in a way token pricing does not.
What to count before buying another tool
Count the hand-offs in one process, end to end. Count how many times a person copies data from one system into another. Time each copy. Count the rework that comes back as a correction. That number, multiplied by headcount and frequency, is the budget you are trying to protect, and it is usually larger than the entire inference bill.
The Digital Journal piece does not publish that number, and the McKinsey figures it quotes are self-reported survey answers. Eighty percent of respondents saying they feel more productive is not a measured output gain. It is a starting point for asking where the work goes after it leaves the tool.
Do we have to replace our existing AI tools to connect workflows?
Usually not. Most stacks already expose an API or at least an export. The work sits in the mapping between systems, the approval steps and the audit trail, not in swapping the model behind each tool.
Is self-hosting an open-weight model cheaper than paying per token?
Only at high, steady utilisation. At the volumes in the table above, API pricing is in the tens of euros per month. A rented GPU large enough for a model of that class typically costs more per day than the whole API bill.
Does Article 50 apply to an internal invoice bot?
The transparency duty is tied to interaction with people and to published synthetic content. A tool that only reads invoices and writes structured data inside the company does not talk to anyone outside it. The moment its output reaches a customer or is published, the labelling question changes.
