How much does a custom AI assistant on Costpoint data cost?

The board asks for a number and the answer is "it depends," which satisfies nobody. Here is what the bill is actually made of, which lines come with a published price, what moves the lines that do not, and how to get a figure you can defend. Vendor prices are as of October 28, 2026.

The short answer: There is no list price, because a custom assistant is a scoped build, not a product. The bill has five lines: a one-time build priced by scope (how many questions, how many sources, and whether CUI pushes the work into a government cloud), Costpoint access your administrator grants (mostly time, with a few items Deltek sells separately), metered model usage at published per-token prices (as of October 28, 2026, $0.10 to $15 per million input tokens on Anthropic's and OpenAI's lists), hosting, and the people time to check answers. For comparison, Deltek's FAQ says Dela is included with the Costpoint license (on-premises customers bring their own Microsoft license) and Microsoft lists its Microsoft 365 Copilot enterprise add-on at $30 per user per month. The way to get a real number is to scope one read-only question and price that.

A CFO can look up what Copilot costs in a minute. A custom assistant on Costpoint data has no list price because nobody sells one off the shelf; the firms that build them quote by scope, which is the only honest way to price work that differs at every contractor. "It depends" is still not an answer for a board, so here is what it depends on.

The bill has five lines, and three of them have a published price

Take a custom assistant apart and the money goes to five places: the firm that scopes and builds it, your administrator and Deltek for access into Costpoint, the model vendor for every question answered, the cloud account that hosts it, and your own people, who check the answers. Three of those lines have a price list you can read today. Two do not, and those two are where the decision sits.

Source: Deltek, Microsoft, Anthropic, OpenAI, AWSWhat a custom assistant's bill is made of, and which lines have a published price
LineWhenWho sets itPublished reference point, as of Oct 28, 2026
Scoping and buildOne timeThe integratorNo list price. Quoted per scope: the questions, the sources, the environment.
Costpoint accessSetupDeltek, through your administratorWeb services: your administrator enables access per user; Deltek says to allow up to 5 business days for the cloud request. Read-only ODBC: a service request; fee not published. Costpoint BI Developer is sold separately from the SaaS bundle.
Model usageMeteredAnthropic, OpenAI, or a cloud providerClaude Sonnet 5.5: $2 in, $10 out, per million tokens. Claude Haiku 5.5: $0.10 in, $0.50 out, for prompts up to 100,000 tokens. OpenAI gpt-6.1-sol (standard tier): $2 in, $10 out. gpt-6-luna: $0.10 in, $0.50 out. Batch processing at 50 percent off on Anthropic's API and Amazon Bedrock.
HostingMonthlyYour cloud accountCompute, storage and a government region if CUI is in scope. AWS prices Bedrock by region, with GovCloud (US) on its own lines.
Your people's timeOngoingYouNo price list. Checking answers against a known number, reviewing the log, confirming each Costpoint release still fits.
What you already payMonthlyDeltek, MicrosoftDela is included with the Costpoint license per Deltek's FAQ; on-premises customers bring their own Microsoft license. Copilot Chat is included with eligible Microsoft 365 plans; the Microsoft 365 Copilot add-on lists at $30 per user per month on Microsoft's enterprise page ($21 on its business page), paid yearly.

Every figure is the vendor's own published list price or documentation on the date shown; none is a quote, a measurement or a client result. Government cloud pricing, Deltek's ODBC fee and Copilot pricing in GCC and GCC High are not published on the pages cited. The recommendation: price the two lines without a list, the build and your people's time, by scoping one question, because the three metered lines will follow from that scope.

The build is the line nobody publishes, and scope is what sets it

Four things move the build price, and you control all four. First, how many questions the assistant answers: one (where does year-to-date overhead sit against the provisional rate, and what moved it) is a small build; ten across finance and contracts is a different project. Second, how many sources each question needs. A question answered from Costpoint alone is simpler than one that joins Costpoint to the funding modification in SharePoint, and the joined questions are the ones that take your staff the most time, which is why they are worth more; we sorted them in Deltek Dela, Copilot or a custom assistant: which does a contractor actually need?

Third, the access path. Deltek documents several ways to read Costpoint without writing to it: web services an administrator enables per user, read-only ODBC and JDBC in Deltek's cloud through a service request, Costpoint Business Intelligence report subscriptions, or the database itself on premises. Each is granted by you, not by Deltek, and each shapes the build. Fourth, the environment. If CUI can reach the assistant, it must sit inside your NIST SP 800-171 scope, which in practice means Azure Government, AWS GovCloud, GCC High or on premises, with their own price lists and setup. Indirect rates, payroll and billing figures are sensitive business data but usually not CUI, so most finance questions avoid that cost; we drew that line in Is it safe to put CUI or contract documents into ChatGPT or Copilot?

Whatever the scope, the build should include a source under every figure, a log of who asked what and which reports answered, and read-only access that cannot post, adjust or approve. Those are not extras; they are why the assistant can be shown to an auditor, and why AI on Costpoint data need not become a DCAA problem.

The model is metered, and at list prices it is the small line

Model usage is the line CFOs worry about most and the easiest to pin down, because the vendors publish it. Anthropic's pricing page lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens, and the top of its list, Claude Opus 4.1 and Opus 4, at $15 and $75. OpenAI's pricing page lists gpt-6.1-sol at $2 and $10, gpt-6-luna at $0.10 and $0.50, and its top standard-tier model at $10 and $50. On both lists output costs five times input, and both document cached input at a small fraction of the base rate and batch processing at a steep discount.

In words a finance team can use: OpenAI's own token guide says 100 tokens is roughly 75 English words, so a million input tokens is roughly 750,000 words of report rows, context and questions, and at the $2 list price that is $2. The assistant sends the model the few report rows a question needs, not the whole ledger, so a read-only finance assistant is a small consumer of tokens. What you cannot find on any price list is the build. Two cautions: prices change without notice, so read the page on the day you budget; and government regions are priced separately (AWS lists Bedrock prices by region with GovCloud (US) on its own lines, and Microsoft's Copilot pages state no GCC or GCC High price). Set a monthly spending ceiling in the cloud account before the first question is asked.

Costpoint access costs time more than money

Reading Costpoint is paid for mostly in your administrator's hours and Deltek's queue. Deltek's integration overview says users have no rights to integration services by default; an administrator must grant "Allow Application Access via Integration Services" per user. In Deltek's cloud, the SaaS Customer Administrator files a "Request Web Services" service request and Deltek says to "allow for up to 5 business days." Read-only database access runs through an "ODBC Read Only Access" request with an encrypted user list and up to five allow-listed IP addresses; whether it carries a fee is not published. Deltek's Costpoint SaaS terms say "Costpoint BI Developer and Costpoint Web Administrator are excluded from the bundle and sold separately." In GovCon Cloud Moderate, Deltek says only its Cloud Operations team touches the production database directly and ODBC access is provided instead.

None of this requires a Deltek partnership. Deltek's SaaS terms define an Authorized User to include a third party contracted to use the services on the customer's behalf: the integrator works under your license, with the access you grant. Rockford holds no Deltek partnership and does not need one.

The costs that are not on any invoice

Two costs never appear on a vendor's bill. The first is your people's time: checking the first month of answers against numbers a person already knows, reviewing the log, and looking again after each Costpoint release (Deltek's release notes index shows four quarterly releases a year, and a report that changes can change what the assistant reads). Budget that time, because an assistant nobody checks is one nobody trusts. In Deltek's 17th annual Clarity study (May 2026, 917 government contractors), 26 percent named difficulty validating AI-provided insights as a top concern and 45 percent were unclear on the return on investment. Sources under every answer settle the first; a scoped first task settles the second.

The second is the cost of being wrong. A figure that reaches a billing, a proposal or the incurred cost submission must be traceable to the books, and an AI analysis that supports a claimed cost is plausibly a record to retain under FAR 52.215-2. An assistant that drafts and flags, with people posting, approving and signing, keeps that cost near zero. Costpoint stays the system of record.

Is it an allowable cost?

Usually the question behind the question. FAR 31.205-33 says costs of professional and consultant services "are allowable" when "reasonable in relation to the services rendered and when not contingent upon recovery of the costs from the Government." Paragraph (f) says fees "are allowable only when supported by evidence of the nature and scope of the service furnished": details of the agreement, invoices "including sufficient detail as to the time expended and nature of the actual services provided," and the consultant's work products. The practical reading, research and not legal advice: buy a written scope, detailed invoices and deliverables you keep. Whether the cost lands in G&A or an overhead pool is your disclosed accounting practice and your auditor's question, within the general test in FAR 31.201-2.

How to get a number you can take to the board

Write down the one question your finance or contracts team answers by hand every month that costs the most hours. Fix its scope: read-only, from named Costpoint reports and, if needed, one document store, checked against a figure a person already knows. Ask for a fixed price for that task, with the model and hosting lines estimated from the published prices above and a usage ceiling set in your own cloud account. Compare the total with what you already pay for AI you may not be using: Dela, which Deltek's FAQ says is included with the Costpoint license for cloud customers, and whatever Copilot your Microsoft 365 plan includes. That is how Rockford prices work: one task, fixed scope, sources under every answer, then the next task once the first has earned it. Small proof first, then scale. We earn trust one task at a time.

Common questions

Why does nobody publish a price for a custom AI assistant on Costpoint?

Because the build is priced by scope, and scope is different at every contractor: how many questions the assistant answers, how many sources each question needs, which path your administrator grants into Costpoint, and whether CUI forces the work into a government cloud. The price ranges we found in a web search while researching this article were for generic chatbots and said nothing about a Costpoint build. The honest way to get a number is to scope one read-only question and price that.

Do we have to pay Deltek anything extra to let an outside firm read our Costpoint data?

Deltek's documentation describes the access paths as things the customer requests or enables, not products it sells to the integrator: an administrator enables web services access per user, the SaaS Customer Administrator files a Request Web Services or ODBC Read Only Access service request, and report subscriptions run from Costpoint Business Intelligence. Deltek's SaaS terms say Costpoint BI Developer and Costpoint Web Administrator are sold separately, and Deltek has not published whether read-only ODBC carries a fee, so ask your account team. No Deltek partnership is needed, and Rockford holds none.

Will model usage surprise us on the bill?

It should not, because it is metered against published list prices. As of October 28, 2026, Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and OpenAI lists gpt-6.1-sol at the same two figures; both vendors list a small model at $0.10 and $0.50 (Anthropic's for prompts up to 100,000 tokens). Both document cached-input and batch pricing at a fraction of those rates. A read-only assistant that answers a finance team's questions is a small consumer of tokens, and the usage log that shows who asked what also shows what each answer cost. Set a monthly ceiling in the cloud account before the first question.

Can we put the cost of an AI assistant in our G&A pool?

FAR 31.205-33 says costs of professional and consultant services are allowable when reasonable in relation to the services rendered and not contingent on recovery from the Government, and paragraph (f) says fees are allowable only when supported by evidence of the nature and scope of the service: details of the agreement, invoices with sufficient detail as to time expended and the nature of the services, and the consultant's work products. So insist on a written scope, detailed invoices and deliverables you keep. That is research, not legal advice; the allocation to G&A or overhead is your disclosed practice and your auditor's call.