AI assistants are easy to demo and surprisingly hard to make useful. Before we build one for a client, we ask four questions. If the answers are weak, we say so — and often suggest something simpler.
1. Is there a real, repeated question to answer?
The best assistants answer the same kinds of questions many times a day: support agents looking up policies, engineers searching internal documentation, sales teams checking product details. If people rarely need the information, a good search page or a well-written FAQ is cheaper and more reliable.
2. Does the knowledge already exist somewhere?
An assistant is only as good as what it can read. We look for documents, tickets, wikis and databases that already hold the answers. If the knowledge only lives in people’s heads, the first project is writing it down — not adding a model.
3. Can we measure whether it helps?
We agree on a measurable goal before building: time saved per request, the share of questions answered without escalation, or fewer repeated support tickets. We then test the assistant against a set of real questions with known good answers, so quality is tracked rather than guessed.
4. What happens when it is wrong?
Every model makes mistakes. The question is how costly they are. For internal tools, showing sources and letting people check them is often enough. For customer-facing answers about money, health or legal matters, you need tighter guardrails, human review — or a different solution.
When we recommend something simpler
- A better search over the documents you already have
- Structured forms and workflow automation instead of free-text chat
- Cleaning up and connecting data before adding AI on top
Where to start
Start with a short use-case assessment: pick one workflow, collect fifty real questions and check whether the answers exist in your data. It takes days, not months, and tells you whether the bigger investment makes sense.
