How to add an AI chatbot to Deltek Costpoint
Short answer
An AI chatbot for Costpoint answers staff questions from live project, labor and cost data plus your internal procedures, shows where each answer came from, and respects Costpoint security. It can run entirely on-premises with an open-weight model, and should start read-only.
Costpoint users ask the same questions again and again. How much funding is left on this project? Why did the overhead rate move? Is this timesheet code still open? Which charge code do I use for bid and proposal work? Today these go to a controller, a contracts analyst or a colleague who happens to know.
A chatbot can absorb the repeatable ones. It works only if it is grounded: answers come from tools that query Costpoint data and from a library of your own procedures, never from the model's memory of how GovCon accounting usually works. Your charge code policy is not the industry's.
The audience matters. A bot for employees (timesheets, charge codes, leave) is a different build from one for finance (variances, rates) or for program managers (funding, burn). Pick one and make it good.
What teams use it for
Employee timekeeping helper
Answers which charge code applies to a task, when the timesheet is due and how to correct a submitted entry, using your timekeeping policy and the list of open codes.
Program manager burn and funding questions
Returns incurred versus funded, monthly burn trend and projected exhaustion date for projects the user manages.
Finance variance questions
Explains movement in pool rates and flags the accounts behind it, with a link to the supporting report.
Contracts reference desk
Answers which clauses and billing terms apply to a project by combining Costpoint billing set-up with the contract document library.
Procurement status inquiries
Shows open purchase and subcontract commitments, receipts and invoice status for a project so program staff stop emailing the buyer.
Deltek Costpoint integration surface
- A read replica or extract of Costpoint project, labor, account, organization and commitment data, exposed to the bot through parameterized tools. Table names and structure vary by version.
- Costpoint web services and REST style endpoints where available, as an alternative to direct database reads.
- Single sign-on and role data used to identify the asker and apply project and organization scope.
- Document sources: timekeeping policy, charge code guide, disclosure statement, procedure manuals, indexed for retrieval with metadata such as owner and effective date.
- Costpoint screens and report links, which the bot can cite so users can verify in the system of record.
How it works
- 1
Choose the audience and top 30 questions
Collect real questions from help desk tickets and emails. Group them into data lookups and policy answers.
- 2
Build data tools for lookups
One tool per question family, parameterized and scoped to the user. Return numbers with as-of dates and period status.
- 3
Index policy documents
Chunk by section, keep effective dates, and retrieve the current version only. Retire superseded policies explicitly.
- 4
Add routing and refusal rules
Questions about legal interpretation, individual pay or anything outside scope go to a named person, not to a guess.
- 5
Embed it where people work
Deploy in your intranet, Teams or a web widget beside Costpoint. Pass the signed-in identity to the backend.
- 6
Review transcripts weekly
Look at wrong and unanswered questions, add tools or documents, and tighten instructions.
Example flows
Which charge code do I use?
Trigger: An engineer asks the bot how to charge time for a proposal support task
- Bot retrieves the current charge code policy section.
- It checks which codes are open and allowed for the user's organization.
- It answers with the code, the reason and the policy citation.
- If unclear, it routes to the timekeeping administrator.
Result: Fewer mischarges and fewer corrections after submission.
Burn rate question from a program manager
Trigger: A program manager asks when a project will run out of funding
- Bot calls the funding tool for the project they manage.
- It computes average monthly burn from the last three closed periods.
- It projects the exhaustion month and notes the assumption.
- It links the Costpoint report for verification.
Result: A stated, checkable answer, and an earlier funding conversation.
Keeping data on-prem
A Costpoint chatbot can run end to end on your own infrastructure. Serve an open-weight model (Llama, Qwen, Mistral, Gemma or gpt-oss class) with vLLM or Ollama on local GPUs, keep the vector index and the tool service in your data center, and authenticate through your identity provider. Nothing about timesheets, rates or contracts leaves the network, which is what many CMMC-minded contractors want. Smaller models are enough for lookups and policy answers when the tools do the heavy lifting; reserve a larger model for the explanation step if testing shows a need. Plan for GPU capacity during business hours and monitor response time.
Security checklist
- Pass the real user identity to every tool call and filter results by project and organization scope server-side.
- Keep personnel and pay data out of scope for general audiences; restrict those questions to authorized roles.
- Show sources and as-of dates so users can verify, and mark answers about policy with the document version.
- Retain transcripts per your records policy and restrict who can read them, since they may contain sensitive questions.
- Run adversarial tests for prompt injection from documents and form text before launch.
Ways to get it done
Deltek's built-in assistant features, if your version has them
Effort: Configuration only
Questions answerable inside the product; check licensing and data handling.
Generic enterprise chatbot platform plus a Costpoint connector
Effort: Weeks to months
You already license a platform and have connector skills; verify on-prem support.
Private chatbot built with Netray
Effort: Pilot in weeks for one audience
On-prem requirement, scoped tools, document grounding and transcript review.
Pitfalls
- !Letting the bot answer policy questions from memory. If the document is not retrievable, it should say so.
- !Mixing audiences. A single bot for everyone makes scoping and tone hard.
- !Stale policy documents. An outdated charge code guide is worse than none.
- !No escalation path, leaving users stuck when the bot declines.
Frequently asked questions
Can the chatbot see all of Costpoint?
It should not. Give it a defined set of tools and documents. Each tool filters by the asker's project and organization access. Anything outside those tools is simply not available to it, which keeps answers accurate and exposure small.
Will it replace our help desk?
It reduces repetitive questions, mainly timekeeping and status lookups. Judgment calls on cost accounting and interpretation still need your finance and contracts people. Design the bot to hand those off cleanly rather than attempt them.
How accurate will it be?
Accuracy comes from grounding. Numbers from tools are only as good as your queries, so validate against closed periods. Policy answers are only as good as your documents. We recommend a test set of real questions and a review cycle.
Can customers or subcontractors use it?
Technically possible, but start internal. External users raise authentication, data separation and contractual issues. If you go there, expose a very narrow tool set such as invoice status, on a separate deployment.
What hardware does an on-prem version need?
For a few hundred internal users and a mid-size model, one or two data center GPUs with adequate memory is typical, plus a server for the index and tools. Actual sizing depends on concurrency and model choice, which we test in the pilot.
Want AI chatbot working on your Deltek Costpoint?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment