A small shared chat-with-your-documents setup can start at $50 a month for two ChatGPT Business Standard or Claude Team Standard seats on monthly billing. An existing eligible subscription may avoid a new seat charge. Custom systems need a separate quote covering preparation, search, hosting, access controls and maintenance.
The subscription is only one line in the budget. Someone must choose the approved files, remove outdated versions, test answers and maintain access. An internal reference assistant and a customer-facing service need different levels of checking, so describe the intended users before comparing quotations.
Decide which system you are actually buying
“Chat with your documents” can describe three different arrangements. First, a person supplies a few files to an assistant and asks questions. Second, a team maintains a shared collection of reference documents. Third, a custom service finds relevant passages from a larger library and uses them to answer questions through a dedicated interface.
The third arrangement is often called retrieval-augmented generation, or RAG. In plain language, the system searches the approved material before composing an answer. Retrieval does not guarantee that the search found the right passage or that the answer interpreted it correctly. The owner still needs a way to inspect the supporting source.
Start by writing the question the system must answer. “What does our current cancellation policy say about package bookings?” is a manageable initial task. “Answer anything about every client, contract and financial record” is not the same purchase. It needs much more work on access, source selection and acceptable answers.
An illustrative barber shop has six procedure documents and two people who consult them occasionally. A maintained shared assistant may be enough to test the idea. A custom search service is hard to justify merely because a demonstration made document chat look impressive. First check whether ordinary folders and a clear index already solve the problem.
ChatGPT Projects can keep related files and instructions together, and Claude Projects support project knowledge and instructions. Both offer sharing on their business plans. Treat these as candidates for a small controlled library, not evidence that every document format, access arrangement or scale will work as required.
For a broader architectural choice, use the small-business knowledge-base comparison. Here the buying question is narrower: which costs belong in your estimate, and how can you test them before committing?
Separate the invoice from the work around it
A subscription buys access to a product. It does not automatically organise your files, decide which version is authoritative or teach staff when to reject an answer. Put those responsibilities beside the invoice so you can compare a simple setup with a more managed service fairly.
| Cost category | Usually incurred when | What to ask for |
|---|---|---|
| Document preparation | Initially, then as files change | Inventory, duplicate removal, readable text and named owners |
| Seats or service subscription | Monthly or annually | Actual users, minimum seats and relevant plan limits |
| Search and model usage | As the system is used | Measured cost per representative question and storage charges |
| Access and setup | Initially and when roles change | Who can read which sources, tested with ordinary accounts |
| Evaluation | Before launch and after important changes | Known-answer questions and failure checks |
| Maintenance and support | Ongoing | Updates, removed documents, alerts and named responsibility |
Use hours as well as money. If the owner prepares documents themselves, the cash payment may be zero but the work is not. Value that time using an internal planning rate you choose. Keep the rate labelled as an assumption rather than presenting it as an industry standard.
An illustrative nail salon has 80 PDF files. Sixty are clear text, ten are duplicated price lists and ten are photographs of printed instructions. Counting “80 documents” conceals the work. Remove the duplicates, identify the current price list and test whether the photographed pages become accurate readable text before estimating the rest.
OCR, or optical character recognition, turns an image of text into machine-readable text. It can misread a price, a date or a small footnote. A sample page that reads “$18” as “$78” makes a cheap import expensive later. Budget for checking difficult pages, and use the document-processing tutorial for preparation decisions.
Three starting budgets, with the assumptions exposed
A small internal collection using paid team seats
For an illustrative two-person trial, two ChatGPT Business Standard seats on monthly billing cost $50 a month. Claude Team Standard has the same monthly seat arithmetic. Annual billing is $20 per seat per month for these Standard plans, but an annual commitment should follow evidence that the workflow is useful.
Assume six hours of preparation and testing at an internal value of $30 an hour. That gives $180 of one-off effort. Allow one hour monthly to check and update the small collection: $30. The illustrative first-year total on monthly seats is $180 + 12 × $80 = $1,140, including internal time.
This is a budget scenario, not a guarantee that a particular library fits a particular plan. Confirm file limits, formats, sharing and account controls in the current product. An individual plan can be cheaper for one person, but do not compare its price without considering how business information is handled and who owns access.
A team using AI inside its existing office suite
Google Workspace Business Standard is about $14 per user per month on an annual plan and includes Gemini across its business apps. If the team already has that plan, using an included feature may add no seat charge. Preparation and checking still belong in the project budget.
The closest thing to a ready-made document chat in Workspace is Gemini Notebook (formerly NotebookLM). You give it a set of sources and ask questions against them. Google says it is now included in all Workspace plans, with higher limits on qualifying plans, and that uploaded data is not used for model training. A single source can hold up to 500,000 words. Before pricing anything custom, a Workspace team should run the known-answer test described further down in a notebook, because a result that is good enough there costs nothing extra.
Do not multiply the entire existing Workspace bill and call it a new document-chat cost. Equally, if the team must upgrade to get the required features, include the difference for every affected account. Confirm the available features and current account price; a consumer Google AI subscription is a different purchasing route.
Microsoft 365 businesses should distinguish included Copilot Chat from a paid Copilot licence, particularly when the requirement involves reasoning across emails, meetings and files together. Buying a paid add-on to answer questions about one supplied document is a decision to test, not an automatic requirement.
A custom library with a maintained interface
For a custom service, request a scoped quotation. As an illustrative calculation only, 40 supplier hours at an assumed $80 an hour produce a $3,200 setup budget. Add ten internal hours at $30 for preparation and acceptance: $300. The opening cost is $3,500 before any paid software or recurring service.
Suppose the proposed monthly allowance is $40 for hosting and search-related services, $20 for model usage, two supplier maintenance hours at $80 and two internal checking hours at $30. That totals $280 monthly. The first-year planning total becomes $6,860. None of these service allowances is a vendor quotation or typical market price.
Ask suppliers to replace those placeholders with line items and limits. A system with several permission groups, frequent imports or customer-facing answers may need a different design entirely. Conversely, a small shared library may make this custom budget unnecessary. Pay for requirements that the simpler route cannot meet.
A yoga studio prices 500 questions a month
The main worked example is an illustrative yoga studio with four staff accounts and 60 approved procedure documents. Staff ask roughly 500 reference questions monthly, about class substitutions, membership rules and opening checks. Personal health information is excluded. The owner wants cited answers and a clear response when the files do not contain an answer.
The studio selects a four-seat ChatGPT Business Standard trial on monthly billing. Seats cost 4 × $25 = $100 monthly. It assigns eight hours to file preparation and four hours to writing and checking sample questions. At an assumed $30 internal hourly value, the one-off effort is $360.
Ongoing work is budgeted at two hours monthly, worth $60. The initial monthly total is therefore $160, and the first-year planning total is $360 + 12 × $160 = $2,280. This assumes usage and the file collection fit the selected plan. It is an illustration, not a reported client result.
Dividing $160 by 500 gives $0.32 per question before spreading the setup cost. Including setup across twelve months gives $190 a month, or $0.38 per question. That does not mean the vendor charges per question; it is the owner's way to compare total operating cost with the work being done.
Now test the time case. If finding an answer manually takes three minutes and checking an assisted answer takes one minute, 500 questions could release 1,000 minutes, or 16 hours 40 minutes, monthly. That assumes every question fits the use case and the checking minute is real. Measure both times on a representative sample.
Suppose the trial instead finds only 150 useful questions and an average one-minute saving after checking. The released capacity falls to two and a half hours, worth $75 at the assumed rate. That is below the $160 monthly operating budget. Better document organisation or fewer paid users may be the more sensible next experiment.
Why a tiny model bill can sit inside a large project
Tokens are the pieces of text an AI model reads and writes. API pricing commonly separates input tokens from output tokens. A custom service may send instructions, retrieved passages and the question as input, then receive the answer as output. ChatGPT and Claude subscriptions do not include separately billed API use.
At OpenAI's list prices for its low-cost gpt-6-luna model, input costs $0.10 per million tokens and output costs $0.50 per million. Consider 1,000 illustrative questions averaging 2,000 input tokens and 300 output tokens. That is two million input tokens and 300,000 output tokens: $0.20 + $0.15 = $0.35 for those model tokens.
This arithmetic is deliberately limited. It does not include retrieval calls, document processing, storage, hosting, retries, longer conversations or staff checks. Nor does it establish that the chosen model answers your questions accurately. A more capable model may be worth its cost if it reduces expensive correction work.
OpenAI's retrieval documentation prices vector-store storage, the searchable representation of your files, separately. The first 1 GB across stores is free; storage beyond that costs $0.10 per GB per day. The measured storage includes processed chunks and their search representations, so it is not simply the original folder size.
For example, 3 GB total leaves 2 GB billable. At that rate, a 30-day month costs $6 for storage alone. Ask the implementer to show measured storage after importing a representative sample. Do not scale a document count into a storage bill without checking what the processing produces.
Use awkward questions to expose hidden costs
Useful tests show more than whether the system can quote an easy paragraph. Set aside 20 questions whose answers a person already knows. Include absent information, contradictory files, restricted documents and a recently changed policy. Keep the expected source and answer beside each question.
Illustrative wedding planner example: an old package sheet says the coordination service costs $900; the current approved sheet says $1,100. Ask which price applies to a new enquiry. If the assistant blends them into “from $900”, the repair is a source-management rule, not a more persuasive prompt. Remove or clearly separate obsolete material and retest.
Illustrative tattoo studio example: reception staff should read booking procedures but not private consultation notes. Test the system while signed in as reception. A sentence telling the assistant “do not reveal private files” is not a substitute for restricting access to those files. Price actual permission enforcement into any custom proposal.
Illustrative personal trainer example: a customer asks whether a particular injury qualifies for a refund, but the approved policy covers only cancellation timing. A successful answer says that the files do not establish eligibility and routes the question to the owner. It must not invent a policy or offer medical advice.
Use only the approved cancellation policy supplied.
Question: Does a shoulder injury automatically qualify for a refund?
Give the relevant section and its date.
If the policy does not answer, say what is missing.
Do not infer a health or refund decision.An illustrative output is: “The cancellation policy dated 12 September describes notice periods, but does not state an injury exception. Ask the owner to review the request.” Keep that useful uncertainty. If an output adds “injury refunds are normally approved”, remove it: the supplied policy did not support the statement.
Illustrative nail salon example: the owner removes an obsolete treatment menu and asks the same question again. Check whether the old content can still be retrieved and how quickly removal takes effect. OpenAI documents that vector-store removal is eventually consistent, meaning old content can remain searchable briefly. Agree a removal procedure that fits the business's needs.
Illustrative barber shop example: ten occasional staff users ask only 40 questions monthly between them. At $25 per seat monthly, ten Standard seats cost $250 before maintenance. That is $6.25 per question in seat cost alone. A well-written procedure index may be a better first purchase; do not share one login to disguise the seat requirement.
Buy a small accepted result before a larger library
Ask for a pilot with one document owner, one approved collection and a written acceptance test. Specify what a correct answer looks like, when the assistant must decline and where staff should report a wrong answer. Keep the pilot internal until its failure patterns are understood.
For the yoga studio, an illustrative acceptance rule is: at least 18 of 20 questions answered correctly or correctly declined, every factual answer traceable to an approved source, and no restricted-source disclosure. The numerical threshold is the owner's chosen trial rule, not an industry benchmark. Any access failure blocks expansion regardless of the average score.
Run the same questions after replacing a policy, adding a document and changing a user's access. Measure the time needed to maintain the collection. If updates take longer than expected, reduce the initial scope or change the process before adding more users. The guide to wrong chatbot answers helps separate source problems from answer problems.
End the quotation with named responsibilities. Who approves a replacement file? Who removes old content? Who receives a failure report? Who can export the instructions and approved documents if the service ends? These tasks turn a promising demonstration into something the team can rely on, and they belong in the price from the beginning.
Further reads
- How to Summarise Long Documents With AI Without Missing Details — Check summaries against the details in long source files.
- How to Use ChatGPT Projects to Keep Client Work Separate — Organise a controlled collection before expanding access.
- How to Set Up Claude Projects as a Shared Team Assistant — Explore a shared reference collection using Claude.
- AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep approved documents and instructions portable.
- How Much Does a Custom AI Assistant Cost a Small Business? — Three ways to get an AI assistant that knows your business, what each costs up front and every month, and the upkeep that most quotes leave out.
- What Is RAG? A Plain-English Explainer for Business Owners — RAG explained without jargon: how AI looks up your own documents before it answers, traced through a real-looking policy question.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI, ChatGPT Business pricing, Projects and chats, and API pricing and Retrieval documentation; Anthropic, Claude Team pricing and How can I create and manage projects?; Google Workspace pricing and Gemini Notebook product page; Microsoft Learn, Overview of Microsoft Copilot Chat (all checked September 2026). Labour and custom-service budgets are illustrative assumptions.