Buying AI after the demo: a 2023–2026 retrospective

Purpose and scope

This retrospective was prepared on 3 October 2026. It covers selected external milestones, not DataXLR8's publication or delivery history. The dates below describe source events. They do not imply that this article, its recommendations or company experience existed at those dates.

Our reading of the sequence is practical: buyers have increasingly explicit material to draw on when specifying evidence, authority and operational responsibilities. These publications do not establish how quickly businesses adopted those practices, or prove that every AI product follows them.

2023 supplied a general risk-management reference

On 26 January 2023, the US National Institute of Standards and Technology published version 1.0 of its AI Risk Management Framework. It is voluntary guidance intended to work across use cases. Its publication supplies a dated reference point for discussing AI risks; it is not Australian law or a certification of any product.

For a buyer reviewing a proposal now, the useful question is how the supplier has translated general risk considerations into the proposed job. What outcome matters? Who can be affected by an error? What evidence will support the release decision? A framework citation in a slide deck does not answer those questions on its own.

That distinction also helps keep a historical article honest. The existence of a framework establishes that the guidance was published. It does not demonstrate that a particular business used it, that a supplier was compliant with it, or that a system's risks were resolved.

2024 added more specific privacy and architecture questions

NIST dates the release of its generative-AI profile to 26 July 2024. This is a more specific milestone within its risk-management work. Here it serves as historical context; this article does not claim to assess a system against every item in that profile.

In Australia, OAIC published guidance on commercially available AI products on 21 October 2024, with the current page recording an update on 17 January 2025. It explains privacy considerations around AI inputs and outputs, including the possibility that incorrect generated information about an identifiable person is still personal information. The publication did not create all underlying privacy duties on that date.

The buying implication is to examine the whole information flow. A description of where the main database is hosted is only part of that discussion. Ask about prompts, outputs, logs, support access and retained copies. Determine the applicable obligations for the actual organisation and use, rather than treating a general checklist as a legal conclusion.

On 25 November 2024, Anthropic announced the Model Context Protocol as an open standard for connecting AI assistants to data and business systems. The announcement supplied a common protocol vocabulary. Our practical inference is that connection and permission still need separate decisions: a connector's existence does not authorise an assistant to read every record or perform every available action.

Anthropic's 19 December 2024 article on effective agents distinguished predefined workflows from agents that dynamically direct their process and tool use. The live article now acknowledges later tooling changes. Its displayed publication date supports this milestone, while its present SDK and product references should not be treated as a frozen record of the 2024 ecosystem.

For today's proposal comparison, ask the supplier to explain why the job needs dynamic decisions. A fixed workflow can be a sensible choice when the route is known. The word “agent” does not settle that design question.

2025 made acceptance and operating authority easier to specify

Anthropic published its tool-writing guidance on 11 September 2025. It recommends evaluation tasks grounded in realistic uses and data, with outcomes that can be verified. This is vendor engineering guidance, not independent proof that any particular implementation will succeed.

Its practical value for procurement is the move from “show us an impressive example” to “show us the expected and observed outcomes on our agreed cases”. Include missing information, conflicting sources and unavailable tools. A trial should tell the buyer what happens outside the smooth path as well as inside it.

OWASP's 2025 edition of its excessive-agency guidance identifies excess functionality, permissions and autonomy as sources of risk. The exact publication day was not established in this research, so the edition year is the appropriate historical label. Its recommendations include limiting permissions and enforcing authorisation outside the model.

On 21 October 2025, Australia's Guidance for AI Adoption introduced six essential practices, according to the Department of Industry's history statement. The current implementation guidance discusses acceptance criteria, testing, accountable deployment decisions and monitoring. The historical launch date and today's living guidance are related evidence, but they are not proof that every sentence on the current page appeared at launch.

By 2 October 2026, the useful output is a better brief

This review's 2026 endpoint is the source-check date, not a claim that a new law or product suddenly resolved AI purchasing. The milestones provide different kinds of evidence: a US voluntary framework, Australian regulator and adoption guidance, vendor engineering recommendations and security guidance. Treating them as interchangeable would obscure their purpose.

Use the distinctions to complete this commissioning worksheet. The proposed questions below are our synthesis, rather than a copied regulator checklist.

These questions connect a historical reading exercise to a decision you can make now. If the next project concerns internal documents, use the source and answerability worksheet in Can your company knowledge answer the question? If it concerns taking action, start with Before an AI workflow acts, decide how it can fail.

Bring the resulting brief to a consultancy discussion with the unanswered questions still visible. They are useful inputs to scope, price and acceptance. A supplier's response to those gaps will tell you more about the proposed engagement than a retrospective timeline alone.