When we started building Toby, the first design decision was one acronym: PII. Before anything goes to a model, names, tax file numbers, bank details and salaries are taken out, and each identifier is replaced with a code that means nothing outside your account. We could only make that decision once we knew what every other acronym on a vendor's security page was promising, and what it was not.
Our earlier piece, Twelve AI words you will hear this year, covered the words. This one covers the letters: the acronyms on product pages, in tender documents and in the security questionnaire a vendor sends back. Each entry gives what the letters stand for, what the thing is, and what to ask when you see it.
GPT
Generative pre-trained transformer. The model design behind ChatGPT and most of its competitors: trained in advance on a large amount of text, and generating its answer one piece at a time. "GPT" in a product name tells you which family of model is underneath, and nothing about how the product checks what the model writes.
ML
Machine learning. Software that learns patterns from examples instead of following rules someone wrote. Bank-feed categorisation that suggests an account code from past coding is ML. It predicts from what it has seen before, so a new supplier or an unusual transaction is where it is weakest. Ask what it does when it is unsure.
NLP
Natural language processing. The older umbrella term for software that reads or writes human language. Most of what vendors now call AI is NLP done by a large language model. The term on its own tells you nothing about which method is used.
OCR
Optical character recognition. Turning an image of text, such as a scanned receipt or a photographed invoice, into text a program can read. It is where errors enter a document pipeline: a 7 read as a 1 becomes a wrong figure that everything downstream treats as fact. Ask whether low-confidence reads are flagged for a person.
API
Application programming interface. The published way one program talks to another. When a tool "connects to Xero", it does so through Xero's API, with the permissions you grant. Ask which permissions (Xero calls them scopes) it requests, and whether it asks to write to your file or only to read it.
SLM
Small language model. A language model small enough to run on modest hardware, sometimes on a single machine in the vendor's own data centre. It is cheaper and easier to keep in one country than a large model, and usually less capable on hard questions. The trade is a fair one to ask a vendor to explain.
RLHF
Reinforcement learning from human feedback. A training step in which people rate a model's answers and the model is adjusted towards the answers they preferred. It is a large part of why assistants sound helpful and polite. It also rewards answers that sound satisfying, which is one reason a model rarely says "I don't know". The next piece in this series covers that.
AGI
Artificial general intelligence. A hypothetical system that matches people across most intellectual work. Nothing sold to a practice today is AGI. When the term appears in a sales conversation, it is describing the vendor's ambitions, not the product.
PII
Personally identifiable information. Names, tax file numbers, bank details, salaries and anything else that identifies a person. The question for a vendor is whether PII is sent to the model, and if it is, whether it is removed or replaced before it goes.
APP
Australian Privacy Principles. The thirteen principles in the Privacy Act 1988 that govern how covered businesses handle personal information. APP 8 deals with sending it overseas, and APP 11 with keeping it secure. They apply to your use of AI products as much as to your filing cabinet.
PIA
Privacy impact assessment. A written review of how a new system will handle personal information and what could go wrong. The Office of the Australian Information Commissioner suggests one when choosing a product that uses AI on personal information. It is also a useful document to ask a vendor for.
SOC 2
System and Organization Controls 2. A report by an independent auditor, under an American Institute of CPAs standard, on a service provider's security controls. A Type II report covers a period of time, not a single day. Ask for the report, not just the badge, and check that the systems you would use are in its scope.
ISO/IEC 42001
The international standard, published in 2023, for an organisation's AI management system: who is accountable, how risks are assessed, and how a system is monitored after release. It certifies the vendor's process. It says nothing about whether a particular answer is right.
How we use these letters
At Toby we let the model explain and route, and nothing else. It is not given names, tax file numbers, bank details or salary figures, and it does not produce the figures in an answer. Those come from tables a named person has signed off against the published source, and the working is done in ordinary code. When a vendor's page uses one of the acronyms above, the question after it is usually the one that matters.
Next
How a model answers, and why it sounds sure.
Sources
- Privacy Act 1988 (Cth), Schedule 1, Australian Privacy Principles 8 and 11.
- Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products (October 2024).
- ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system.
- AICPA, SOC 2 reporting on controls at a service organisation.
- Ouyang et al., Training language models to follow instructions with human feedback (2022).