Why do quotes for AI knowledge search vary so much?
Two proposals can describe the same outcome and contain very different work. One may index a whole file server as it is. Another may include document clean-up, permission mapping, a test set of real questions and tuning against it. The second costs more to build and gives you evidence of how the system performs before staff rely on it, which matters because many retrieval faults start with document preparation and retrieval settings. Ask each supplier to state which documents are in scope, how many test questions they will measure against and what result ends the project.
What are the one-off costs?
The one-off work covers scoping, connecting to where documents live, extracting text, splitting documents into passages, handling permissions and testing against real questions. Scanned files, tables and documents with conflicting versions add preparation time, and so do permission structures with many exceptions. Much of this time is yours as well as the supplier's, because someone has to confirm which version of a procedure is current and check the test answers. Count that internal time when you compare a build with a licence.
- Number of documents, and the share that are scanned or image-only.
- Number of separate sources, such as SharePoint, Google Drive and a job system.
- Number of permission groups the system must respect.
- Number of test questions a person must check.
What does the search service cost to run?
The index has to live in a search service. On Microsoft's Azure AI Search, the dedicated pricing model bills a fixed hourly rate per search unit for as long as the service exists, whether or not anyone asks a question, and adding replicas or partitions increases the number of billed units. A serverless option, in preview, bills by compute used and storage instead. Optional features carry their own charges: semantic ranking by query, image extraction per 1,000 images, and vectorisation and AI enrichment at the model provider's rate. Ask suppliers on other platforms to show the same split between fixed capacity and usage.
How is model usage billed per question?
Each question sends the question, instructions and retrieved passages to a language model, and the model returns an answer. Providers bill by tokens, which are fragments of text, with separate rates for input and output. On Anthropic's published price list, for example, output tokens cost several times more per million than input tokens. Retrieved passages count as input, so the number and length of passages sent with each question drive this cost more than the question itself. Sending fewer, better-ranked passages lowers it.
Hypothetical example: 40 staff each ask 5 questions a working day, which is 200 questions a day or 4,400 over 22 working days. If each question sends 4,000 input tokens of passages and instructions and returns a 300-token answer, the month uses about 17.6 million input tokens and 1.3 million output tokens. Multiply each figure by your chosen provider's current per-million rate. These volumes are assumptions for illustration only.
How does a per-user licence compare with a build?
Per-user products such as Microsoft Copilot charge for each licensed person, whether they ask one question a month or fifty. Microsoft also offers metered pay-as-you-go billing for Copilot Chat and SharePoint agents through an Azure subscription, with budgets an administrator can set. A built system charges mostly by usage plus a fixed search service. The comparison turns on how many people need access and how often they ask. A large group of occasional users tends to suit usage billing, while a small group of heavy users working across email, files and meetings may suit a licence.
Which ongoing costs do people forget?
After launch, someone has to re-run the test questions on a schedule, review questions that returned nothing, re-index changed documents, remove superseded ones and check that permission changes have reached the index. Microsoft's documentation for Azure AI Search storage indexers notes that changes are detected automatically but deletions are not, unless a deletion detection policy was in place from the first run. Without one, deleted documents stay in the index. These tasks need an internal owner and time set aside each month, plus any supplier support you contract.