AI is more than large language models: which technique for which problem
Since chat assistants arrived, many people have reduced AI to large language models. Yet AI also means machine learning and document reading, and many tasks can be automated without it. Here is how to choose the right technique at each step, at the lowest total cost for compliant work.
By Kantara Fofana
In short
- AI is more than large language models: according to the OECD, an AI system infers from its inputs how to generate predictions, content, recommendations or decisions, and machine learning and computer vision, which includes document reading, are part of it.
- Each technique has its place: rules for what must be exact and identical on every run, machine learning to classify and forecast, document reading to extract fields, a language model to draft and summarize.
- A good share of automation needs no AI at all: a formula, a flow or a software robot does the job.
- The right choice is made step by step, at the lowest total cost for compliant work, at the expected quality, and a single process often combines several techniques.
One stack of slips, four different problems
March, in a fifteen-person accounting firm. Slips and bank statements arrive by the hundreds, and someone suggests handing everything to a chat assistant: read the documents, classify the transactions, do the calculations, write to clients. Yet these are four problems, with neither the same solution nor the same cost.
Reading a slip is a job for document reading. Classifying transactions based on history is a job for machine learning. Calculating an amount is a job for a formula, which gives the same result on every run. Drafting the request to the client is work for a language model. Handing all four to the same tool means paying, on every run, for a plausible result where an exact one was needed, then having it checked again. Reducing AI to large language models and wanting “AI for everything” is a costly mistake.
What official definitions call AI
On 8 November 2023, the OECD Council adopted a revised definition: an AI system is a machine-based system that infers, from the input it receives, how to generate outputs such as “predictions, content, recommendations, or decisions”. The word “content” was added at that time to cover generative AI; predictions, recommendations and decisions were already in the 2019 definition.
The OECD’s explanatory memorandum, published in March 2024, names machine learning and knowledge-based approaches (systems that reason over explicitly described rules and relationships) among the techniques, and computer vision, natural language processing and speech recognition among the application areas. It lists classifying data, detecting anomalies and forecasting among the tasks AI performs, and states that the definition encompasses systems ranging from simple to complex.
At the federal level, the Treasury Board governs the automated decision systems of federal institutions. Its guide on the scope of the Directive on Automated Decision-Making lists rules, regression, advanced analytics, AI, generative AI, machine learning, neural networks, robotic process automation and natural language processing, then concludes that the directive covers more than AI systems.
The key point
A large language model is one AI technique among others, not AI itself. So the right question is not “do we need AI?” but “which technique for this step?”
The map of techniques: four families, four profiles
- Rules and conventional automation: a formula, a script, a flow triggered by an event, or a software robot that replays a person’s actions on screen (robotic process automation, or RPA). Everything is written in advance; nothing is learned.
- Machine learning: techniques that derive a model from past data rather than from written instructions, to classify a case, forecast a value or flag what departs from the norm.
- Document reading: optical character recognition (OCR) turns an image of text into text; extraction then returns each useful value, such as an amount or a box on a slip, with a confidence score, the estimated probability that the value read is correct.
- Language models: models trained on large sets of text that answer one token at a time (a token is a word or part of a word), choosing a likely continuation at each step.
| Technique | What it does well | Its limits | Cost per run | Data it needs | Reproducibility |
|---|---|---|---|---|---|
| Rules and conventional automation | Calculate, validate, route, enter data | Cannot handle the unexpected; a software robot must be readjusted when the application’s screen changes | Almost nothing once built | No training data: written, tested rules | Identical on every run |
| Machine learning | Classify, forecast, spot an anomaly in structured data | Needs a clean, representative history; loses accuracy when reality changes | Low per prediction | A history of examples, often labelled | Identical for a given model version; explainable by the importance of each factor |
| Document reading (OCR and extraction) | Read documents in volume, field by field, with a confidence score | Struggles with damaged, handwritten or new types of documents; reading a box is not understanding it | Low, billed per page analyzed | None for a prebuilt model; a few labelled examples for a custom model | Stable for a given version; uncertain fields go to a person |
| Language models | Draft, summarize, rephrase, extract from free text, converse | Can produce credible but inaccurate content; unreliable at arithmetic | Usually per token: it grows with the length of the text sent and produced | None to start: instructions and your reference documents | Variable: two identical requests can get two answers |
At Microsoft, document reading is billed according to the model type and the number of pages analyzed; its documentation also notes that generative AI services often bill each request by the number of input and output tokens. Total cost also includes building, maintenance and human review: a plausible result is paid for twice, once for the request and again for the review. And the families overlap: a language model is itself a product of machine learning, as are document-reading models, and Microsoft even offers generative document-reading models. When a document is of a new type, or a text must be understood rather than merely read, such as a clause or a free-form email, the language model often becomes the best reader: see Reading documents: when OCR is enough.
One example per profession
Four typical cases, built for illustration: each combines several techniques.
Accounting firm: from the slip to the client request
Document reading extracts the amounts from T4 and RL-1 slips; fields with a low confidence score go to a technician. A machine learning model, trained on transactions the firm has already classified, suggests a category for each line of the bank statements, and the professional confirms. Calculations and consistency checks are rules. Finally, the language model drafts the request to the client for missing documents, reviewed before it goes out. See Tax season 2027.
Law firm: the deadline, the letter, the case law
A deadline is calculated with rules. Quebec’s Code of Civil Procedure provides that the starting day is not counted but the terminal day is, and that a time limit that would expire on a Saturday or a holiday is extended to the following working day (art. 83). Coded once, these rules always give the same date. The first draft of a demand letter goes to the language model, and the lawyer reviews it. For case law research, the model helps frame the question and summarize, but every decision cited is checked against its official source: see AI-fabricated case law.
Pharmaceutical company or CRO: the deviation, the form, the summary
At a sponsor or a contract research organization (CRO), a visit outside the window set by the protocol is detected by a rule. What cannot be described in advance calls for analysis: the ICH E6(R3) guideline, adopted in January 2025, provides for centralized monitoring in which the analysis of accumulated data helps identify protocol non-compliance and potentially unreliable data; machine learning can flag the site that stands out from the others. Document reading extracts the fields of scanned forms. The language model drafts the first version of the deviation summary, which the designated study lead corrects and approves. And since ICH expects computerized systems to be fit for purpose, every reproducible step makes their validation simpler.
Small business: the supplier invoice and the demand forecast
For a supplier invoice, document reading extracts the supplier, the amounts and the taxes; rules check the GST and QST; a software robot enters everything into the accounting software if it offers no way to import (see Quotes and invoices with Microsoft 365). To forecast demand, a machine learning model learns from sales history and seasons, and the importance of each factor can be measured. No language model is needed here.
Seven questions to ask before every project
Ask them at every step of the process, not about the project as a whole: the answer almost always points to the technique.
Must the result be exact, or is plausible enough?
An amount, a date, a tax: a rule or a formula. A first draft that a person corrects: a language model.
Must it be identical on every run?
If an auditor must find the same result tomorrow, prefer a rule, or a machine learning model whose version is frozen. A language model does not ensure this: Microsoft notes that generative AI models are designed to produce unique content, and won’t generate the same output for identical inputs.
Is the input structured?
Columns and codes: rules or machine learning. Scanned documents: document reading. Free text: a language model.
What volume?
A few cases a month are often handled by hand. Thousands of documents justify automation, with review targeted at doubtful cases.
What does one run cost, multiplied by the volume?
A low unit cost, multiplied by every page and every month, compares with the near-zero cost of a rule. Add the time spent checking.
Must the decision be explained?
In Quebec, a decision based exclusively on automated processing of personal information requires the business to give, on request, the reasons and the principal factors and parameters behind it (s. 12.1 of the private-sector privacy act; see Law 25 and AI). A written rule or an explainable model makes that possible.
Is the data sensitive?
Health information, social insurance numbers, financial statements: their sensitivity dictates where processing can happen, and with which vendor.
Four misconceptions
“AI means chat assistants”
Where businesses use AI, they use it first to analyze. In the second quarter of 2026, 19.2% of Canadian businesses had used AI to produce goods or deliver services over the 12 months preceding the survey, according to Statistics Canada. Among them, and businesses could cite more than one, data analytics (36.6%) and text analytics (34.5%) came ahead of virtual agents or chatbots (28.2%), natural language processing (27.0%) and large language models (24.8%). Interest is shifting, though: among businesses planning to use AI, 39.3% cited large language models in the third quarter of 2026, compared with 9.3% two years earlier (see Your competitors are moving to AI).
“A language model can do everything”
It excels at drafting, summarizing, rephrasing and conversing. But the Government of Canada’s guide on generative AI, written for federal institutions, notes that AI tools aren’t the best choice in every situation and advises against using generative AI when factual accuracy or data integrity is needed; another page of practical advice from the government lists arithmetic among the tasks it doesn’t do well. This is structural: according to Microsoft, a large language model generates its output one token at a time, each time predicting the token that should come next; it predicts how a result is written rather than performing the operation. See Don’t entrust a calculation to a language model.
“Machine learning is outdated”
It remains the basic tool for classifying, forecasting and detecting. Its share among businesses that use AI is stable: 18.2% in the second quarter of 2026, compared with 18.6% a year earlier, while the share of businesses using AI rose from 12.2% to 19.2%, according to Statistics Canada. On structured data, it costs little per prediction, can be measured against history and can be explained: Azure Machine Learning computes the importance of each feature, meaning each data field used, for the whole model or for a single case.
“Automation requires AI”
Microsoft describes desktop flows, the robotic process automation capability of Power Automate, as designed for simple or complex rule-based tasks. A reconciliation, a deadline calculation, data entry between two applications: written and tested once, they give the same result on every run. AI becomes useful where the input escapes the rules: a badly scanned document, a free-form email, a case never seen before.
Where to start on Monday
- Pick a costly process and break it down into steps.
- Ask the seven questions above at each step.
- Give formulas, rules and flows the work they can do first.
- Keep machine learning for steps with a reliable history, and the language model for text, always reviewed.
- Estimate the cost per run multiplied by your annual volume, review time included.
- Ask every vendor which technique it uses at each step, and why.
Frequently asked questions
- What is the difference between AI, machine learning and a language model?
- AI is the whole field. Machine learning is one of its families of techniques: it derives a model from past data to classify, forecast or detect an anomaly. A large language model is a particular machine learning model, trained on large sets of text to produce text.
- Is optical character recognition (OCR) artificial intelligence?
- Under the OECD definition, yes: the OECD’s explanatory memorandum lists computer vision and the recognition of data among the application areas and tasks of AI. In practice, a document-reading model is trained on examples and signals its doubts with a confidence score, which is used to send uncertain cases to a person.
- Is a software robot (RPA) artificial intelligence?
- Not in itself. A software robot carries out predefined actions on screen: it follows a script and infers nothing from what it receives. AI is not limited to what learns, though: the OECD also includes knowledge-based approaches, which reason over explicitly described rules. Software robots are often associated with AI because a step that learns or interprets is added to them, such as reading a document or classifying an email.
- Does machine learning need a lot of data?
- It needs a history that represents the problem, often labelled: transactions already classified, past sales, annotated documents. For a custom document-reading model, Microsoft advises, for example, at least five samples of each variant of a document, scanned or digital, and more to improve accuracy.
Sources
Every fact in this article links to its official source, consulted on October 29, 2026.
- OECDRecommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449), definition of an AI system (opens in a new tab)
- OECDExplanatory memorandum on the updated OECD definition of an AI system, OECD Artificial Intelligence Papers No. 8, March 2024 (opens in a new tab)
- Treasury Board of Canada SecretariatGuide on the Scope of the Directive on Automated Decision-Making (opens in a new tab)
- Government of CanadaGuide on the use of generative artificial intelligence (opens in a new tab)
- Government of CanadaGenerative AI in your daily work (opens in a new tab)
- Statistics CanadaAnalysis on artificial intelligence use by businesses in Canada, second quarter of 2026 (opens in a new tab)
- Statistics CanadaAnalysis on planned use of artificial intelligence by businesses in Canada, third quarter of 2026 (opens in a new tab)
- LégisQuébecCode of Civil Procedure (CQLR, c. C-25.01), art. 83 (opens in a new tab)
- LégisQuébecAct respecting the protection of personal information in the private sector (CQLR, c. P-39.1), s. 12.1 (opens in a new tab)
- ICH (International Council for Harmonisation)Guideline for Good Clinical Practice E6(R3), adopted on 6 January 2025 (opens in a new tab)
- Microsoft LearnInterpret and improve accuracy and confidence scores (Azure Document Intelligence) (opens in a new tab)
- Microsoft LearnService quotas and limits, Azure Document Intelligence (billing) (opens in a new tab)
- Microsoft LearnModel interpretability, Azure Machine Learning (opens in a new tab)
- Microsoft LearnIntroduction to desktop flows, Power Automate (opens in a new tab)
- Microsoft LearnHow generative AI and LLMs work (opens in a new tab)
- Microsoft LearnUnderstand tokens (opens in a new tab)
This article provides general information. It is not legal, tax or accounting advice for your situation.
Your case deserves thirty minutes.
A conversation to understand your context and tell you, frankly, whether we are the right choice. The diagnostic that follows is free of charge.
Book a call