AI Gains Lost to Manual Workflows
Release date:
Short answer
A director at consultancy YCP says Singapore firms are layering AI onto manual workflows and multi-stage approval queues, so the speed gain never reaches the customer or the ledger. He puts the miss rate on digital transformation initiatives at up to 70%, and points to undefined success metrics, no accountable senior owner, and legacy KPIs as the causes. For an SME, that is an argument for fixing system integration in Singapore before buying more AI tools, not after.
What happened
In a Singapore Business Review video, Jason Ong, Director at YCP, said many companies still run on spreadsheets, documents and several approval stages even after introducing AI. His example is blunt: the AI produces an answer almost instantly, but that answer then sits in an approval queue with several handoffs and sign-offs. He added that firms often launch AI without naming the financial or operational metric it should improve, and without a senior leader accountable for that outcome. Adoption suffers when staff have to open yet another application. He also flagged legacy KPIs that reward old ways of working, and suggested tracking how often people override AI decisions, since a high override rate tells you either the model is wrong or people do not trust it yet. He speaks at the Singapore Technology Enterprise Summit 2026 on 12 November.
Why it matters for an SME in Singapore or Southeast Asia
Most businesses of 10 to 200 people here did not start with a system. They started with a spreadsheet, a WhatsApp group and one person who knows how it all fits together. AI gets bolted on top of that, usually as a chat window or a drafting tool, and the output goes straight back into the same manual chain: someone copies it into the ERP, someone else checks it, a manager signs off, and the finance team re-keys it a third time.
The AI saved four minutes. The queue around it costs two days. That is the whole story.
The practical version of Ong's point is this: the constraint is rarely the model. It is the number of places a piece of data has to be typed, copied, approved or reconciled before it does anything useful. If your quotation takes three days because it moves from email to spreadsheet to a PDF to a signature, an AI that drafts quotes in ten seconds has not helped you. You have made the first step faster and left the other nine alone.
We see the same pattern in why Singapore AI pilots stall before scale: the pilot looks fine in a demo, then meets the real workflow and loses. The fix is unglamorous. Connect the CRM, the ERP and the internal tools so the same record is written once and read everywhere. Then the approval step either disappears or becomes a single click on a screen that already has the context. That is the ground AI-powered digital transformation actually stands on.
What to do about it
- Name the number before the tool. Pick one metric the project must move: days from enquiry to quotation, invoice cycle time, hours of data entry per week, error rate on order entry. Write it down. If nobody can name it, you are buying a demo, not a result.
- Assign one owner. A named senior person accountable for that number. Not a committee. In a 40-person firm that is usually the operations lead or the finance lead, with the owner signing off on scope.
- Map the handoffs, not the software. Take one process end to end, say quote to cash, and count every place a human retypes, copies or approves something. That count is your real baseline. It is also where the time goes.
- Connect the systems underneath. Make the CRM the single source for customer and opportunity data, the ERP the single source for orders and stock, and build the integration between them so neither side re-keys. Our notes on CRM vs ERP cover which to do first if you can only do one this year.
- Automate the re-keying before the thinking. The copy from form to ERP, the stock check before a promise date, the invoice raised from a signed order. These are boring and they are where the hours are. Workflow automation examples has a list you can score against your own process.
- Then add AI, on connected data. An assistant that drafts the quote from the CRM record, or an agent that checks stock and lead time before replying, only works if it can read the real data. See AI chatbot vs AI agent for where the line sits.
- Track overrides. Log every time a person changes what the AI or the automation produced. A high rate is diagnostic, as Ong says. Either the logic is wrong or the team does not trust it, and those need different fixes.
- Retire the old KPI. If your sales team is still measured on call volume, they will keep making calls instead of working the qualified pipeline the CRM now shows them. Change the metric with the system, or the system loses.
Where it does not apply, or what to be careful of
- Not every approval is waste. Anything touching money above a threshold, contracts, credit terms, safety or regulated data should keep a human in the loop. The goal is fewer pointless sign-offs, not zero oversight.
- Do not strip controls to hit a speed target. If you remove an approval step, make sure the check it performed now happens in the system, automatically and visibly. Removing a control without replacing it is how errors get through.
- AI does not remove errors. It moves where errors happen. Clean, connected data reduces them; scattered data just lets the model produce wrong answers faster and with more confidence. Results depend on the state of your underlying records.
- Small firms may not need custom work. If you are 12 people running one process, an off-the-shelf CRM with a decent integration to your accounting tool is often the right answer and cheaper than anything we would build. Custom internal tools earn their place when your process is genuinely different or the volume justifies it.
- The 70% figure is a consultancy's estimate, not your diagnosis. Treat it as a prompt to check your own project, not as a prediction about yours.
If you want a second opinion on where your workflow actually loses the time, book a discovery call. We look at how the work runs today before proposing anything, and scope comes after that conversation. You can also see what connected systems look like in practice in our case study on a custom ERP for retail operations and the AI support and lead system.
Takeaway
The news is not that AI disappoints. It is that AI inherits whatever workflow you already have, including the approval queue and the spreadsheet. Speed at the front of a slow process is still a slow process. Fix the connections between your CRM, ERP and internal tools first, name the metric and the owner, then put AI on data that is already joined up. That order is the difference between a pilot that demos well and a system your team actually uses.
Frequently asked questions
Does this mean we should stop our AI project?
No. It means check what the AI output feeds into. If the answer still lands in an approval queue, a spreadsheet or a WhatsApp thread, fix that path first. The AI can stay; the handoffs around it need to change.
How do we know if our workflows are the bottleneck?
Count the handoffs between an AI or system output and the action it is meant to trigger. If a person re-keys the same data into a second tool, or several sign-offs sit between the answer and the decision, that is the bottleneck, not the model.
What metric should we pick before starting?
One operational or financial number the project must move, such as days to quote, invoice cycle time, or hours spent on data entry per week. Name the senior person who owns that number. Without both, you cannot tell whether the project worked.
Do we need to remove approval steps?
Only where the risk is genuinely low and the AI has been tested. Keep a person in the loop for anything with money, contracts or compliance attached. The point is to stop approving things nobody actually reads.
How does Kelin Studio help with this?
We connect the CRM, ERP and internal tools so data flows without re-keying, then add AI on top of that connected data. It starts with a discovery conversation about how your operations actually run today. Scope follows that, not a price list.
