Three quotes, three different worlds
You asked three suppliers to build an AI assistant for your team: something that answers questions from your documents and can look things up in your CRM. One quote is small and promises a quick turnaround. One is several times larger. The third is somewhere in between and full of terms you do not recognise. All three say they will use the latest AI models.
You cannot tell whether the cheap one has missed something or the expensive one is padding. The proposals describe different things in different language, and none of them quite matches what you had in mind. Choosing on price feels risky; choosing on the most confident sales call feels worse.
Why the quotes vary so much
"An AI assistant" covers a huge range. At one end is a chat window connected to a folder of documents using an off-the-shelf tool. At the other is an assistant that signs users in, respects who can see what, connects to several live systems, is tested against hundreds of real questions, logs everything, and is monitored and maintained. Both are honestly called a custom AI assistant.
The AI model itself is rarely the main cost. The work is in everything around it: getting your data into a usable state, connecting to your systems, making sure people only see what they should, proving the answers are good enough, and keeping it working as models and data change. Suppliers make different assumptions about each of these, often without saying so.
The drivers behind the numbers
| Cost driver | Why it moves the price |
|---|---|
| Data sources | Each system (SharePoint, CRM, ERP, database) needs its own connection and upkeep |
| Document condition | Scanned, duplicated or messy files need cleaning before AI can use them |
| Permissions | Making answers respect who can see what is significant work |
| Actions | Looking things up is simpler than updating records or sending messages |
| Evaluation | A proper test set of real questions takes effort, and is what makes quality measurable |
| Security and hosting | Private hosting or strict data rules add design and running costs |
| Support after launch | Monitoring, model updates and improvements are ongoing work |
A low quote is not necessarily wrong. It may be for a simpler product that fits your need. The problem is only when you cannot tell which product each quote is for.
How we help you get comparable quotes
Whether or not we end up building it, we help you pin down what you are buying.
- We work through the questions the assistant needs to answer and the actions it needs to take, using real examples from your team.
- We list the data sources involved and check their condition and access routes, so nobody is quoting blind.
- We define who can use it and what each person is allowed to see.
- We set out how quality will be judged: a test set of real questions with agreed good answers, and the standard the assistant needs to reach before launch.
- We state the hosting and data handling requirements, including anything your clients' contracts demand.
- We describe the support you expect after launch, so running costs and maintenance are priced rather than assumed.
The result is a written scope any competent supplier can quote against. It also tends to show whether a simpler, off-the-shelf option would do, in which case we say so.
What you get from this
Quotes you can put side by side, because they answer the same questions. You can see whether a price difference comes from a different approach to testing, a different hosting choice, or simply different rates.
You also go into the build with a shared definition of done, which is the thing that most often goes missing on AI projects and leads to arguments later.
And the test set outlives the buying decision. Whoever builds the assistant can use it during development, you can use it to accept the finished product, and it becomes the check that shows whether later changes made the assistant better or worse.
Is this your situation?
- You have quotes for a custom AI assistant that differ widely.
- The proposals describe the product in different terms.
- You are not sure which data sources and systems need to be included.
- Nobody has defined how the assistant's answers will be tested.
- You want to understand the cost drivers before committing.