Before you buy anything

A post went round this week saying the biggest barrier to AI in companies is messy knowledge, information and data. I half agree. Messy files do stop AI being useful after the first week. But in the SMEs I work with, the files are rarely the first problem, and a new system is almost never the fix.

This is what I would put in front of a 20- to 80-person company:

layer	what most SMEs need	cost
storage	The Google Workspace or Microsoft 365 you already pay for	already paid
housekeeping	One naming rule, one owner per folder, a clear-out every quarter	staff time
asking questions	NotebookLM (in Workspace) or Copilot (in Microsoft 365)	included, or a per-user add-on
capture	A phone recorder, meeting transcripts, one-page checklists	close to zero
privacy	A restricted folder and a written rule on what AI never sees	staff time
judgement	Your best people training juniors on real jobs	supervisor time

Four of the six rows are habits. Only one is a purchase most firms haven’t already made.

The Singapore numbers point in different directions. IMDA counted 14.5% of SMEs using AI in their business in 2024. A study commissioned by AWS found skills shortage was the top blocker, named by 43% of firms. EY put cost and not knowing which tools fit side by side.

So ask one question before you organise anything: has anyone here tried to get AI to answer a question from our own documents, and failed because the documents were the problem? If nobody has tried yet, the gap is skills and habits. A data clean-up won’t close it.

• storage

  Use the drive you already have

  Don’t add Notion, Confluence and a document system on top of it. Three places to look for a file is worse than one messy place.

• housekeeping

  Three rules, no software

  folders

  one shared drive per team, not per person.

  names

  a file name rule people will remember: date, client, type of document.

  owner

  one named person per folder who clears out old versions every quarter.

  A lot of paid “knowledge organisation” time ends up here. Try it in-house first.

• asking questions

  Comes bundled

  google

  NotebookLM is included in Workspace Business Standard and above. Make one notebook per team; answers point to the page they came from.

  microsoft

  Copilot does the same over SharePoint, as a paid add-on per user.

  sensitive files

  an open-source tool like Open Notebook can run on one office computer with a local model, so nothing leaves the building. Slower, and someone technical has to set it up.

  Test it with ten real questions your supervisors get asked every week. When an answer comes back wrong, the fix is usually the document, not the tool.

• capture

  The step people skip

  What matters most in an SME often isn’t in a folder. It’s with the technician who has been there 18 years, or the owner who knows which supplier is always late.

  Capturing it is cheap. Record the senior person talking through one job as they do it, get a transcript, and turn it into a one-page checklist. Put the checklist in the shared drive.

• privacy

  Before the first upload

  Customer IC numbers, staff medical certificates and salary files often sit in the same drives as the SOPs. Before you point any AI tool at a folder, decide what is off limits under the PDPA and move it out. A separate restricted folder is enough. Write the rule down before something leaks into a chat, not after.

• judgement

  Where files run out

  Writers on expertise from Michael Polanyi to Gary Klein make the same point: much of what an expert knows can be written down, but the judgement calls that set them apart from a newcomer mostly can’t. A notebook will hold the checklist. It won’t hold the feel for when the checklist is wrong. That still needs a junior working beside the expert and getting corrected, on real jobs.

• A data lake or data warehouse. That’s for firms whose product is data.
• A custom chatbot from a vendor, until the bundled tool has failed on your real questions.
• A “knowledge platform” subscription that copies the drive you already have.
• A long readiness study before anyone has tried anything.

IMDA’s CTO-as-a-Service gives SMEs a free first consultation on which tools fit, and pre-approved AI tools can be co-funded under the Productivity Solutions Grant. Use those before paying anyone for software advice.

Paid help makes sense on the people side: deciding whose job changes, getting your experts to pass on what they know, and training staff to check AI answers instead of trusting them. No tool does that part for you.

New to all of this? Start with a first month of AI for a small business →

Figures from IMDA’s Singapore Digital Economy Report 2025, AWS and Strand Partners’ Unlocking Singapore’s AI Potential (2025), and EY, “Helping SMEs take the next leap towards AI” ↗ (Jan 2025). The AWS study was paid for by a cloud vendor.

I use Anthropic’s Claude every day and shaped this page with it. Not affiliated with or endorsed by Google, Microsoft, IMDA, AWS or EY.