Cloudflare has introduced an AI model called Clairic. It is designed to extract data from documents – and is intended to do so faster and more cost-effectively than large language models. At this point in time, no more is publicly and reliably known about the model; however, the general direction is clearly recognisable.
What is less interesting is the individual model than the pattern behind it. For around three years, the standard answer to almost every text task has been: a large language model. Clairic represents a countermovement that has long been taking place in practice – specialised, smaller models for narrowly defined tasks.
What actually happens when extracting data from documents
Document extraction refers to pulling structured fields out of unstructured documents: invoice number, amount and due date from a PDF invoice, line items from a delivery note, master data from a scanned form. The result should be machine-readable so that it can flow into an ERP system, a database or an approval workflow.
This task is narrowly limited. It does not require world knowledge, reasoning ability or creative formulation. Above all, it requires reliability: the same invoice should yield the same fields every time.
This is exactly where the mismatch arises. A general-purpose language model brings a great deal of capacity to such a task that is not needed at all – and this capacity is paid for on every call, in money and in waiting time.
Why cost and latency hurt particularly with documents
With a chat assistant, a response time of two seconds hardly matters. In batch processing of documents, the calculation looks different, because two factors come together:
- Volume: Document processes rarely run individually. We are talking about hundreds or thousands of records per month, often with several pages per transaction.
- Repetition: The same extraction logic is applied again and again to the same document type. Every call costs again.
What looks like a few cents per document becomes, over the course of a year, an item that becomes visible in budget planning. And what begins as a delay of seconds blocks approval processes, imports or nightly processing windows in aggregate.
There is also a point that is often overlooked in cost comparisons: the more strongly a process is tied to an external model provider, the more strongly throughput times depend on that provider's availability and load. Rate limits – that is, restrictions on requests per unit of time – are a real bottleneck in mass processing.
Specialised model or language model: a decision per task
The practicable stance is not "either or", but a division according to task type. A workable framework for projects:
- High volume, clearly defined target schema, little variance: Here much speaks in favour of a specialised model or a classic extraction service. Invoices, forms, delivery notes, master data sheets.
- Low volume, high variance, open-ended question: Here a large language model plays to its strengths. Summaries, free-text analyses, special cases, documents without a fixed format.
- Hybrid form: The specialised model handles the bulk, the language model call catches the outliers that the lean model cannot resolve reliably.
In practice, this hybrid form is often the most economical variant. However, it presupposes that the system recognises when it is uncertain – via confidence values, plausibility rules or a comparison with master data.
What should be clarified before choosing a model
Before a model is selected, it is worth working on the task definition itself. Experience shows that the following points decide the outcome more strongly than the question of which model is in use:
- Define the target schema: Which fields are needed, in which format, with which mandatory details? A fuzzy schema produces fuzzy results, regardless of the model.
- Build a test set: A collection of real documents together with verified expected results. Without this reference, no model comparison can be proven, only asserted.
- Determine the cost of errors: An incorrectly extracted amount in a payment approval weighs differently from a wrong title in a document repository. This determines how much checking is needed.
- Define the escalation path: What happens to documents that cannot be processed reliably? A manual review route is part of the concept, not an emergency solution.
Anyone who has clarified these four points can actually compare models – by hit rate, throughput time and cost per document. And can switch later when a cheaper or faster model becomes available.
Architecture: keeping models interchangeable
The strategically most important point about announcements like this is less the concrete product than the speed at which the offering changes. Anyone who today ties an extraction pipeline firmly to one provider pays integration effort again with every switch.
The countermeasure is unspectacular: a clearly defined internal interface between the business process and the model call. The workflow only knows "document in, structured data out" – which model works behind it is a configuration question.
Then a new specialised model can be tested in parallel operation against the existing setup without touching the running process.
This also applies to the reverse case: if it turns out that a lean model does not cleanly cover certain document types, it must be possible to return to the larger model without rebuilding.
Context for web and software projects
For projects, this above all means: the question is not "Which AI model do we use?" but "Which task are we solving, and what is the smallest sufficient tool for it?" Particularly for document pipelines in portals, customer areas or back-office workflows, this question has a noticeable impact on operating costs and response times.
In implementation, it is therefore advisable to plan the model call as a replaceable component and to measure quality using your own test set rather than product promises. This keeps a solution robust, even when the market – as it currently is – continues to move every six months.