Draft — pending editorial review · not indexed
Field guide · updated 2026-08-06 · 7 min
Build, buy, or hire: three ways to get an AI assistant working for you
Subscribe to a product, commission a custom deployment, or have someone build it in-house — the real costs and failure modes of each path, and a sequence that avoids the expensive mistakes.
01The three paths
The three paths
Once you know what you want an assistant to do, there are exactly three ways to get it: buy a product (subscribe to something self-service, configure it yourself); commission a deployment (a vendor or partner builds around your workflows — the 'AI employee' market lives here); or build in-house (your own developers, or increasingly, one technical person with modern AI tooling). Each is right for a different shape of problem, and each has a characteristic way of wasting money.
02Buy — when your need is common
Buy — when your need is common
If your need recurs across thousands of businesses — answering enquiries from your facts, chasing leads, screening CVs, holding reminders — a product almost certainly exists at S$20–200/month. Products amortise engineering across customers: you get permissions, audit, and updates for a fraction of their cost. Expect real configuration effort (your facts and rules still need writing down; nothing exempts you from that) and accept the trade-off that you adapt your workflow to the product's shape.
The characteristic waste: subscribing, skipping the configuration, concluding 'AI doesn't work for us'. The product was fine; the facts were never loaded.
03Commission — when your need is specific and valuable
Commission — when your need is specific and valuable
When the workflow spans multiple systems, touches sensitive data with governance requirements, or is unusual enough that no product fits, you're commissioning: five-figure setup or four-figure monthly is the honest range in Singapore. The premium buys workflow mapping, integration, permission engineering, and someone accountable — the checklist from our AI-employee guide applies squarely.
The characteristic waste: commissioning before the workflow is stable, then paying change-orders as it shifts under the build. Commission workflows that have run repeatably for months, not aspirations.
04Build — when the assistant is close to your edge
Build — when the assistant is close to your edge
In-house building has become genuinely viable for smaller teams: modern models and tooling let one capable person assemble assistants that would have needed a team in 2023. It fits when the assistant touches proprietary data you won't send out, when the workflow is your competitive edge, or when you're iterating too fast for a vendor relationship. The real cost isn't the build — it's ownership: monitoring, model updates, permissions, and the bus factor of the one person who understands it.
The characteristic waste: a working prototype that becomes load-bearing without ever getting the boring parts — audit, fallbacks, documentation — and dies when its builder leaves.
05The sequence that avoids the expensive mistakes
The sequence that avoids the expensive mistakes
Write the brief first — role, inputs, actions, boundaries, budget (our finder produces exactly this document). Then try to buy: a week's trial against your brief costs almost nothing and either solves the problem or teaches you your real requirements. Commission only what a product demonstrably can't do, with your trial learnings as the spec. Build only what's genuinely yours to own.
The pattern behind every expensive AI mistake we've seen: the sequence run backwards — a commissioned build for a need a S$99 product covered, specified before anyone had operated anything. The cheap paths are also the fastest teachers; use them first.