9 January 2026
Three questions to settle before any pilot
A pilot without three answers tends to become a demonstration that nobody can close. The first question is ownership. Who will live with this on a Tuesday, six weeks after the launch fuss, and what else will they put down in order to do that. If the owner is a committee, you do not yet have an owner.
The second question is the job. Is the model drafting, retrieving, classifying, or deciding. Those are different tasks with different costs of error. A drafting assistant that a person reviews is not the same system as an automatic message to a customer. Mixing the jobs in one sentence is how success measures later fall apart.
The third question is the material. Which files, which fields, which volume, and which of them are personal data under the Personal Data Protection Act (PDPA). If you cannot describe the material, you cannot describe the risk or the benefit. Settle the three questions in writing. Then choose a tool. Singaporeaitools will decline to cheer a pilot that still lacks them, because the quarter is too expensive to spend on theatre.
16 January 2026
Reading a vendor's security pack without a security team
Many mid-sized firms have no spare security analyst to read a sixty-page pack. You can still extract a useful picture if you ignore the badges on the cover and look for a few facts. Where is content stored. Is it used to train the vendor’s own models. Which subprocessors sit in the path. How deletion works, including backups. What logs you can export. How you leave.
If those facts are missing, ask for them in mail you can file. A sales call is not a record. If the answers change between the pack and the data-processing addendum, believe the addendum and ask for the pack to be updated. Region claims need a location, not a phrase about global clouds.
This reading does not replace your own security review on a regulated workload. It is a way to arrive at that review with fewer surprises. We keep a longer eight-point list on the governance page. Bring counsel in before you send personal data into a trial, even a short one.
23 January 2026
What staff may and may not paste into a public chatbot
Public chatbots are already open on many desks. Telling people to “be careful” is not a rule they can follow at speed. Write two lists. The allowed list is short: public wording, dummy examples, questions about grammar, and material you would be willing to print on a postcard. The forbidden list is also short: customer identifiers, unpublished numbers, health information, passwords, legal advice you have not issued, and anything covered by a confidentiality clause.
Then say what to do instead. If there is an approved tool with a tenancy you control, name it and how to get an account. If there is not, say that some tasks stay manual until there is. People paste into public tools because those tools are fast. A policy that only forbids, with no path, will lose by Friday.
Include a sentence on incidents. If something sensitive has already gone out, who to tell, and that telling quickly is expected. Shame is a poor control. Logging and a calm route to privacy or technology colleagues will do more. This is guidance, not legal advice. Your counsel may need a tighter rule for your sector.
6 February 2026
Why document-heavy pilots stall before the model is the problem
Retrieval-augmented generation means the system searches your documents and then asks a model to answer from what it found. When the answer is wrong, teams often request a larger model. The usual cause sits earlier. Titles differ across folders. Three versions of the same contract sit side by side. Scans have no text layer. Access rights mean the search cannot see the file a person would have opened by asking a colleague.
Spend the first weeks of such a pilot on the corpus. Agree a small set of questions the team already answers by hand. Call that your gold set. Clean the files that would be needed to answer those questions. Only then run the model. Success is whether the search returns the right passage, and whether a person still wants to use the draft.
If the documents cannot be made searchable in a month, say so and stop. Tuning a model on a heap will not create a records practice. Singaporeaitools would rather close a pilot early than watch a team lose confidence in a tool that never had a fair corpus.
13 February 2026
Measuring an assistant that drafts rather than decides
If a person must still send the email, file the note, or approve the figure, you are measuring a drafting assistant. Do not borrow accuracy numbers from a decision system. Useful measures look like time to a first draft, share of drafts a person actually uses, number of factual corrections a reviewer makes, and whether a wrong draft would have left the building without the review.
Agree the measures before the trial, on a volume you already understand. A week of anecdotes after launch will not settle an argument with finance. Sample enough cases to see a pattern, and write down what “good enough” would mean in advance. If reviewers still rewrite every sentence, the tool is not saving the process. If they accept drafts they have not read, you have a different problem, and logging who clicked accept becomes part of the design.
We help clients choose a small set of measures during scoping. We will not put a promised percentage of savings on a page before those measures have been watched on real work. The point of a pilot is to learn whether the process changed. The numbers should serve that question.
