AI can assist us. It cannot carry our accountability.

Notes from the CIPS Caribbean webinar on the ethical and practical use of AI in procurement, 15 September 2026.

One of the most rewarding things about speaking is discovering that an idea travelled further than the room.

Last Tuesday I joined the CIPS Caribbean webinar on AI in procurement, alongside Anesha Sadar, Dr Stuart Milligan and Olufemi Raheem. Professionals joined from Trinidad and Tobago, Barbados, Nigeria, South Africa and Canada. Stuart covered capability. Femi walked through live use cases inside a real supply chain. My part was the boundary: what AI must not be handed.

The message was fairly simple. AI can assist us. It cannot carry our accountability.

The law arrived before the guidance did

Three dates frame the position in Trinidad and Tobago. On 26 April 2023, Legal Notice 106 of 2023 brought the operative provisions of the Public Procurement and Disposal of Public Property Act into force, section 5(2) among them. In May 2025 the Government established the Ministry of Public Administration and Artificial Intelligence. In November 2025 that ministry launched a National AI Assessment with UNDP and UNESCO, and a national AI policy is being developed.

The Act says nothing about artificial intelligence. I have not found a line on AI in the OPR's handbooks or guidelines either. Meanwhile, someone in your organisation has already pasted a document into a chatbot. The gap between adoption and guidance is where the risk lives.

It is tempting to read that silence as freedom. It is the wrong reading. Section 5 sets out the objects of the Act: accountability, integrity, transparency and value for money; efficiency, fairness, equity and public confidence; local industry development, sustainable procurement and sustainable development. Section 5(2) requires a public body to carry out procurement in a manner consistent with those objects. The Act is technology neutral, and that cuts both ways. No rule stops you from using AI. Every rule you already follow still applies when you do.

So the question is never whether AI is allowed. It is whether this particular use serves the objects of the Act or puts them at risk.

Two lines I asked the room to hold

The first concerns bidder information. A public AI tool that trains on its inputs is a disclosure you cannot recall. Picture an ordinary afternoon: an evaluator has six bids to compare and a deadline, and pastes one into a free public tool for a summary. If that tool trains on what it is given, a bidder's pricing, methodology and commercial secrets have gone to a third party, and may resurface in fragments to someone else entirely. There is no remedy after the fact. Section 39 of the Act is headed Confidentiality, and it already requires submissions to be treated so their contents are not disclosed to competing suppliers. It was written for paper in envelopes. It does not care about the medium. The United Kingdom faced this squarely in its procurement policy note on AI, which tells contracting authorities to keep confidential tender information out of the training data of large language models.

The second concerns decisions. AI informs. Humans decide. Evaluation, scoring, award, disqualification and disposal stay with named officers, for two practical reasons. Under our Act an aggrieved supplier can challenge, and every decision must be explainable on the record: "the model scored it" is not a reason a review body will accept. And accounting officers answer for public money. Delegating the analysis to software does not delegate the liability. I have been the person who signs. The signature does not ask what tools you used.

There is a quieter reason too. A model trained on your historical award data will reproduce your history, including the parts you are trying to correct.

Five rules for this quarter

One. Classify the data before you touch the tool: public, internal or bidder-confidential.

Two. Public tools for public data only. Bidder information stays inside your security boundary, under contracts that exclude training.

Three. AI analyses, humans decide. Nothing a model generates enters an evaluation report unverified by an accountable person.

Four. Record AI use in the record of proceedings, exactly as you record any other material step.

Five. Ask bidders to disclose their AI use, and say in your tender documents how you will use it. The symmetry builds trust on both sides of the table.

None of this requires a new law. It requires a decision, and the work Stuart described underneath it: roles redesigned and people trained, so the hours AI frees up become higher-value work rather than an unmanaged gap.

Then the idea travelled

The day after the webinar, Ntando Mathaba, a procurement officer who joined from South Africa, published her own reflection on what she had heard. She wrote that the five-rule playbook became the seed for her article.

What I appreciated is that she did not repeat the presentation. She questioned it, researched around it, and worked out how the principles would land in her own organisation and in the South African procurement environment. She went and found the UK policy note herself, and noted something I had not said: that the disclosure it asks of suppliers is for information, not something scored against them at evaluation.

She also identified a practical problem that deserves more discussion than it usually gets. Rule one assumes you can see what the data is. Often you cannot. One bidder submits neatly separated documents that are simple to redact. The next scans an identity document, a company registration and a tax clearance certificate into a single attachment, bundled in with the response itself. She can redact what she knows to look for, but she has no control over how a bidder chooses to compile a submission.

My initial view is that this pushes classification upstream, into how you ask for submissions in the first place: separate uploads for separate document types, specified in the tender documents, so the structure of what arrives is something you designed rather than something you inherited. Until that is in place, the safer assumption is that a bidder submission is confidential in its entirety and redaction is not a defence. I do not think that fully answers her point, and it is the kind of gap that gets found by the people doing the work rather than the people writing the rules.

Her second observation is the one I would most like procurement leaders to sit with. The risk is not only exposure, it is retention. A bidder's pricing strategy fed into a public tool may reappear, reshaped, in someone else's result months later, and neither the bidder nor the buyer would ever know.

The questions are becoming universal

Our countries have different legislation, different institutions and different levels of digital maturity. The questions AI is forcing on us are converging anyway.

How much should we delegate to machines? How do we protect confidential information? Who remains accountable when AI contributes to a decision? And how do we take the benefits of the technology without surrendering human judgment?

These are the questions we are building around at DAG.

There was one more from the floor, from Peter Akpokodje: did I use AI to build the presentation? Yes. For research support and first drafts, under the same rules I had just given the room, with every statutory reference and figure checked against its source by me. Drafts, never decisions.

My thanks to CIPS Caribbean for creating the space, to Nasilee Smart, Naila Ramlogan and Savita Mace for the invitation, and to Anesha, Stuart and Femi for the discussion. And thank you, Ntando, for taking an idea from a Caribbean conversation and carrying it into another context.

The best ideas do not end when the webinar closes. They travel.