How to use ChatGPT with Deltek Costpoint
Short answer
You can use ChatGPT with Costpoint by uploading sanitized exports for one-off analysis, or by building a custom GPT or app whose actions call a small API you control in front of Costpoint. For controlled unclassified information or cost and pricing data, a private model is usually easier to defend than a public service.
Most Costpoint users who try ChatGPT start the same way: export a report to Excel, paste it into a chat and ask for a summary. It works, and it is also where the risk starts. A labor distribution export with employee names and rates, or a project cost ledger with indirect rates, is exactly the kind of data a contractor is supposed to control.
The better framing is a spectrum. At one end is ad hoc analysis of data you have deliberately scrubbed. In the middle is a controlled connection: a custom GPT or an enterprise workspace app that calls an API you own, returns only what the user is allowed to see, and logs it. At the other end is not using ChatGPT at all for that data and running an open-weight model privately.
This page covers the first two honestly, and where the third makes more sense for a Costpoint shop.
What teams use it for
Explaining a variance in plain English
A controller uploads a scrubbed extract of pool and base by period and asks ChatGPT to rank the drivers and draft the narrative for the monthly review.
Drafting a proposal cost narrative
A capture team asks for a basis-of-estimate draft from labor categories and hours they paste in, then edits it by hand. No actual cost data from other contracts is included.
Spreadsheet formula and macro help
Finance staff paste column headers only and ask for the formulas that reconcile a Costpoint export to a billing schedule.
Custom GPT for status questions
A custom GPT with an action calling your internal project-status API answers questions like which projects are within ten percent of funding, using a narrow, permissioned endpoint.
Translating contract clauses to work instructions
Contract administrators ask for a plain-language checklist of cost-accounting obligations from a clause they paste in, without touching Costpoint data at all.
Deltek Costpoint integration surface
- Report and grid exports from Costpoint (Excel, CSV, PDF). These are the simplest input but are point-in-time and carry whatever columns the user chose.
- A reporting database or replica, if your deployment allows it, from which you can generate scrubbed, scheduled extracts for analysis.
- Costpoint web services and REST style interfaces, which can sit behind your own API gateway so a custom GPT action calls your service rather than Costpoint directly. Exact interface coverage varies by version and by on-premises versus Deltek cloud.
- OpenAPI specifications. A custom GPT action needs an OpenAPI schema for the endpoints you expose, so design a small facade API with typed responses.
- Costpoint user and role security, which a facade must respect by passing through or mapping the caller's identity rather than using one all-powerful account.
How it works
- 1
Classify the data first
Decide which Costpoint fields are CUI, export-controlled, personnel data or cost and pricing data. This decides whether any external service is allowed.
- 2
Start with scrubbed exports
Replace names with IDs, remove rates, round amounts if the question allows, and keep exports minimal.
- 3
Choose the right ChatGPT tier
Use a business or enterprise workspace with training on your data disabled, admin controls and retention settings, rather than personal accounts.
- 4
Build a facade API for live questions
Expose three or four read-only endpoints with parameter validation and row limits, deployed in your network, with authentication mapped to the user.
- 5
Publish an action with an OpenAPI schema
Wire the facade to a custom GPT or workspace app, restrict sharing to named users, and keep the instructions explicit about what it must refuse.
- 6
Review logs monthly
Check which endpoints were called and by whom, and prune any scope that nobody needs.
Example flows
Month-end variance narrative
Trigger: Controller finishes the month-end close checklist
- Run a saved scrubbed query that outputs pool, base and variance by account.
- Upload the file to an enterprise workspace chat.
- Ask for the three largest drivers and a first-draft paragraph.
- Controller checks every number against the ledger and edits.
Result: A draft memo in minutes, with the numbers verified by a person before it is used.
Program manager asks a status question
Trigger: A program manager asks the project-status custom GPT which of their projects are near ceiling
- The GPT calls the facade endpoint with the user's identity.
- The facade returns only that manager's projects.
- The GPT lists them with funded value, incurred cost and percent used.
- It links each project to the Costpoint screen for follow-up.
Result: A quick answer scoped to the person's permissions, with every call logged.
Keeping data on-prem
ChatGPT is a hosted service, so it does not run inside an on-premises or air-gapped Costpoint environment. If your contract requires data to stay within a defined boundary, treat ChatGPT as suitable only for non-sensitive tasks such as drafting text and formula help. For questions over live Costpoint data, run an open-weight model (Llama, Qwen, Mistral, Gemma or gpt-oss class) on your own GPUs with vLLM or Ollama, put the same facade API in front of Costpoint, and give users an internal chat interface. You keep the pattern ChatGPT users like, a natural-language question and a grounded answer, without the data leaving your network.
Security checklist
- Use business or enterprise accounts with training disabled and admin-managed retention; never personal accounts for Costpoint data.
- Do not paste CUI, export-controlled technical data or unscrubbed personnel and rate data into a public service.
- Put authentication and row limits in the facade, not in the GPT instructions. Instructions can be bypassed; code cannot.
- Log every action call with the caller identity and parameters, and review for unusual volume.
- Check with your contracting officer or security lead if a contract clause restricts third-party processing.
Ways to get it done
Scrubbed exports in an enterprise ChatGPT workspace
Effort: Hours, plus a data handling policy
Drafting, narratives and formula help where no sensitive fields are needed.
Custom GPT with actions over your facade API
Effort: A few weeks to design the facade and schema
Live status questions for non-controlled data and a modest user group.
Private open-weight model with a Netray-built interface
Effort: A scoped pilot of weeks; ongoing GPU hosting
Controlled data, CMMC-aligned environments, or any contract that limits external processors.
Pitfalls
- !Trusting arithmetic from a language model. Totals and rates should be computed by code or checked by a person against the ledger.
- !Uploading an export that still contains hidden columns, employee names or rates that were in the saved layout.
- !Giving a custom GPT action a single privileged service account, so every user sees everything.
- !Assuming a vendor's enterprise terms satisfy your contract flow-downs. They may not; verify with counsel.
- !Forgetting that exports are stale. A GPT answering from last week's file looks authoritative and is not.
Frequently asked questions
Is it safe to paste Costpoint data into ChatGPT?
Not by default. Safety depends on the data classification, your contract clauses and the account tier. Scrubbed, non-sensitive data in an enterprise workspace is often acceptable. CUI or cost and pricing data usually needs a private deployment or explicit approval from your security and contracts teams.
Can ChatGPT connect directly to Costpoint?
Not out of the box. You build or buy a connector: a custom GPT action calls an API you host, and that API talks to Costpoint through its supported interfaces or a read replica. The model never gets database credentials directly.
Does Deltek offer its own AI?
Deltek has been adding AI features across its products, and what is available depends on your version and plan. Check with your account manager. If their built-in capability answers your questions inside the product, that is the simplest route; a custom stack is for what it cannot cover.
Should we use a custom GPT or an MCP connector?
Custom GPT actions are specific to ChatGPT. An MCP server is a more portable way to expose the same read-only tools to several assistants. If you expect to use more than one client, building the facade as an MCP server is a reasonable investment.
What does it cost to run a private alternative?
Cost depends on user count and model size. A small team can run a mid-size open-weight model on one or two modern GPUs, often cheaper over a year than per-seat licensing for a larger group, though you take on operations. We size it against your actual usage.
Want ChatGPT 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