Customer FAQ chatbot
Answers from your real documents and price lists rather than inventing things. Handles the same twelve questions your team answers all day, and hands over to a human when it should.
Home/AI Solutions
AI & automationGive your team back the hours the admin is eating
Small businesses feel AI's value faster than large ones, because there's nobody spare to absorb repetitive work. We build the practical things — and because we test AI systems for a living, we build them so they can't be talked into leaking your data.
What we build
Start with the task that annoys you most
We don't sell "an AI strategy". We find one repetitive job that's costing real hours and remove it, then look at the next one.
Quotation drafting
From a client's requirements to a formatted quote on your own template in seconds. You review and send. The tedious part disappears; the judgement stays with you.
Invoice & document processing
Supplier PDFs read, checked and pushed into Xero or your accounting system. No more retyping, and no more transposed digits at month end.
Internal copilot
Staff ask a bot instead of asking you — leave policy, product specs, how the process works. Answers come from your documents, scoped so people only see what they're allowed to.
Reporting & summaries
The weekly report somebody currently assembles by hand, generated from your systems and waiting in your inbox on Monday morning.
Microsoft 365 Copilot rollout
Configured with sensible data boundaries first. Copilot surfaces whatever a user can already reach — which is exactly why messy permissions need fixing before you switch it on.
Sometimes the answer is no. Plenty of tasks are better fixed with a spreadsheet formula, a template, or simply removing a step. We'll tell you when AI is the wrong tool rather than selling you one anyway — it's a short-term loss and a long-term client.
How a build works
Two to six weeks, start to live
Small, useful, and in production — not a pilot that quietly dies.
Half-day discovery
We sit with the people doing the work and find where the hours actually go. Usually it's not what management assumed. You leave with a shortlist ranked by hours saved against effort to build — and an honest note on anything not worth automating.
Fixed-price proposal
One task, one scope, one number. We deliberately start narrow: a small thing working beats a big thing half-built, and it tells us quickly whether the approach fits how you actually work.
Build and test
We build it on your data, then try to break it ourselves — including feeding it the awkward, hostile inputs a real customer eventually will. Fixing that during the build is far cheaper than after launch.
Launch, train, hand over
Your staff get shown how to use it and, just as importantly, when not to trust it. You get documentation and the ability to update its knowledge yourself. We don't build things only we can maintain.
Where we're different
An AI assistant isn't a feature. It's a new way into your data.
A normal application only does what a developer wrote. A language model does what it's persuaded to do — and if you've connected it to your customer records, then persuasion has become an attack technique.
The pattern is always the same. A helpful bot gets wired to a lookup tool so it can answer questions. Someone works out how to make it look up records that aren't theirs. Under the PDPA that's an unauthorised disclosure of personal data, and your company answers for it — not the company that made the model.
Plenty of vendors in Singapore will build you a chatbot. Very few will tell you what happens when a customer talks it into misbehaving, because they've never tested one adversarially. We do that as a service.
What we test for
- Prompt injection — instructions that override yours, typed directly or hidden inside a document the assistant reads
- Data disclosure — customer records, internal pricing, or another user's conversation coming back out
- Guardrail and prompt extraction — pulling out your instructions, business rules and internal endpoints
- Excessive permissions — an assistant that can be persuaded to send email, issue a refund or query a database it shouldn't
- Poisoned knowledge — attacker-controlled content in your document store changing what the bot tells customers
- Runaway costs — token-burn attacks that turn your API bill into the attack
Assessed against the OWASP Top 10 for LLM Applications, the industry reference for this work — then extended with manual testing specific to your setup.
The unglamorous part
Your staff are already using AI. Probably without telling you.
Somebody in your company has pasted a customer list into a free chatbot to reformat it. It wasn't malicious — it was faster. But that data has now left your control, and under the PDPA that's your problem rather than theirs.
The fix isn't a ban, which people route around. It's a short, readable policy that says which tools are approved, what may and may not be pasted into them, and what to do when someone's unsure — plus giving staff a sanctioned option good enough that they stop reaching for the unsanctioned one.
- A one-page staff AI usage policy people will actually read
- An approved-tools list, with the reasoning explained
- Guidance on customer data, and what the free tiers do with what you send them
- A short briefing session so it lands rather than sits in a folder
- Review against IMDA's Model AI Governance Framework for Generative AI where it's relevant to you
At a glance
- Build
- 2–6 weeks from discovery to live
- Security assessment
- 3–8 working days, depending on how many tools and data sources the assistant can reach
- Testing safety
- Against staging where one exists. If we must test production, limits are agreed in writing first
- Build from
- S$4,000
- Assessment from
- S$4,200 per assistant
- AI policy
- Included free for managed clients
An honest caveat. Prompt injection has no complete fix today — it's an open research problem, and anyone promising a guaranteed solution is overselling. What works is architecture: assume the model can be persuaded, and make sure that persuasion can't reach anything that matters. That's how we build, and what we test.
Questions
AI, without the hype
Often yes, and more so than for a large company — you don't have spare people to absorb repetitive work, so removing four hours a week from someone's job is immediately visible. But it depends entirely on the task. Come to the discovery session with the job that annoys you most and we'll tell you straight whether it's worth automating.
That's the main risk with a customer-facing bot, and it's why we build them to answer from your actual documents and to say "I don't know, let me get someone" when the answer isn't there. We also test it before launch with the awkward questions real customers ask. You'll never get a guarantee of perfection from anyone honest — but the gap between a careless build and a careful one is enormous.
We design this deliberately and explain it before you commit: which provider, which region, what they retain, and whether anything you send could be used for training. Options that keep data in-region or entirely self-hosted exist and we'll price them if that matters to you. This is a PDPA question as much as a technical one, and it gets answered up front rather than buried in a vendor's terms.
Yes, and we do this often. We assess it as an independent third party — you get the findings, and it's your call what to do with them. We'll work with your existing vendor to fix things if you want, or quote to fix them ourselves. We don't require you to rebuild with us.
In a small business, almost never — there's no slack to remove. What it does is give people back the hours they spend retyping, chasing and copy-pasting, so the five of you can handle the workload of eight without hiring three more. That's usually the actual constraint.
Next step
Tell us the job you'd most like to stop doing
Half an hour on a call is usually enough for us to say whether it's automatable, roughly what it would cost, and whether it's worth it.