What should a government contractor automate with AI first?
Not the proposal, not the close, and nothing that posts. The right first task is a question your finance team already answers by hand every week.
The short answer: start with the questions your finance and contracts staff already answer by hand from Costpoint every week or month: where year-to-date indirect rates sit against provisional billing rates, how far each project has burned against funded value, and which timesheets and unbilled items are outstanding. Those tasks repeat, the data sits in one system, every answer can be checked against a Costpoint report, and nothing posts to the books. Leave anything that writes to Costpoint, approves a transaction or touches CUI for a later phase, and use Deltek's own Dela features for in-system transaction work where your environment has them.
The question has changed. Deltek's 17th annual Clarity study (May 2026, 917 government contractors) found that 90 percent use AI in at least one business function or plan to in 2026. The same study found that 45 percent are unclear on the return, and that the top concerns are data privacy and security (37 percent), inaccurate forecasts (29 percent) and difficulty validating AI-provided insights (26 percent). Nobody is asking whether anymore. They are asking what first, and worrying that the first thing will be wrong in a way they cannot see.
So the first task should be chosen for two properties: it is cheap to check, and it cannot break the books. Everything below follows from those two.
Start where the hours already go
At a contractor of a few hundred people, finance and contracts run on a calendar. Every week someone chases timesheets and corrections, checks burn against funded value by project, and reviews unbilled and aged receivables. Every month the close brings labor distribution, indirect rate calculation against the provisional rates, status reports for program managers, invoices in whatever format each contract demands, and variance explanations for leadership. Quarterly it is estimates at completion and the provisional rate re-forecast. Once a year it is the incurred cost submission, due six months after fiscal year end under FAR 52.216-7, and whatever audits arrive.
Almost every item on that calendar is the same job: find the numbers in Costpoint, explain them, and be ready to prove where they came from. That is the work worth automating first, because it is already defined, already repeated, and already checked by a person who knows what a right answer looks like.
Five tests sort the candidates
Put each candidate task through five questions. It repeats on a weekly or monthly cadence. The data sits in one system you already run. A person can check the answer against a report in minutes. It reads from Costpoint and posts nothing. And the output stays inside the company rather than going into an invoice, a proposal or a government submission. A task that passes all five is a safe first project. A task that fails the fourth is not a first project at all.
| Candidate task | Repeats | Fit as a first task | Call |
|---|---|---|---|
| Year-to-date indirect rates against provisional billing rates | Monthly | 5 of 5 | Start here |
| Burn against funded value by project, with the 75 percent notice tripwire | Weekly | 5 of 5 | Start here |
| Missing and late timesheet follow-up | Weekly | 4 of 5 | Check what Costpoint already does |
| Project status report drafts for program managers | Monthly | 4 of 5 | Next |
| Proposal and pricing drafts from past performance | Per bid | 2 of 5 | Later |
| Posting vouchers, adjusting entries and approvals | Daily | 1 of 5 | Not first |
Fit is a judgment against the five tests, not a measurement. Failing the read-only test alone puts a task last. Your mix of contract types moves the order: a firm that is mostly fixed price cares more about estimates at completion than about provisional rates.
The two tasks at the top share a shape. Each is one question, asked on a schedule, answered from a small set of Costpoint reports plus one document the system does not hold: the provisional billing rate letter in the first case, the funding modifications and the program manager's forecast in the second. Each has a hard consequence when it is answered late. Rates billed above actuals become money owed at settlement. Work past funded value is work the government is not obligated to pay for, and FAR 52.232-22 separately requires written notice when costs are expected to pass 75 percent of the funds allotted.
Read-only comes first for a reason
DFARS 252.242-7006 describes an acceptable accounting system through 18 criteria. Two of them are where AI that writes to the books lands: "accumulation of costs under general ledger control" and "approval and documentation of adjusting entries." An assistant that only reads touches neither. One that posts, reclassifies or approves creates a second path into the ledger that your controls were not designed around, and that is a conversation to have with your auditor after you have a track record, not before.
Three more criteria explain why the answers still need discipline even when nothing is posted: the system must be able to "readily calculate indirect cost rates from the books of accounts," produce "billings that can be reconciled to the cost accounts," and hold "adequate, reliable data for use in pricing follow-on acquisitions." The moment an AI-produced rate or burn figure ends up in an invoice, a price or the incurred cost submission, it has to trace back to a Costpoint report. So the first task should produce answers that show their source every time: which report, which period, which filters, when the data was pulled. We walked through the clause text in Can AI read our Costpoint data without creating a DCAA problem?
One of our five principles says it plainly: "AI drafts and flags; people approve and sign." The first task should be one where that line is never in doubt.
Check what Deltek already ships before you build
Deltek is building transaction automation inside Costpoint, and it would be a waste to rebuild it. Dela agents for time entry, expense reports and contracts arrived in July 2025. Costpoint Time & Expense lists "AI-assisted anomaly detection." The Costpoint 2026.3 release notes describe an Agentic Financial Close tab that shows "project product and customer product bills that Dela flagged for manual review," a Financial Anomaly Analysis card for income statement and balance sheet anomalies, and an automated Dela assessment of AP vouchers that exceed an approval threshold. The 2026.4 notes add an Indirect Rate Anomaly Analysis card and let administrators choose which timesheet applications run Dela analysis and auto-posting for FLSA-exempt employees. Deltek's own description of its agents in the 2026.1 release (January 2026), "capturing AI reasoning, timestamps, and confidence levels on each transaction," is the standard any in-system automation should meet.
That is why timesheet follow-up sits in the table with "check what Costpoint already does." If you run Costpoint in Deltek's commercial cloud with Dela enabled, some of the transaction work is arriving in the product you already pay for. Two things remain outside it. First, the questions at the top of the table span Costpoint and documents it does not hold, and often a second system such as GovWin IQ or a forecasting spreadsheet. Second, as of early October 2026 Deltek's Dela FAQ stated that Dela "currently is not available in Deltek's GCCM Cloud offering," and the 2026.3 and 2026.4 release notes do not say whether the newer features reach that environment. On that FAQ's statement, a contractor in GovCon Cloud Moderate has Costpoint and, for now, no vendor AI inside it. For that contractor the first task is the same, and it runs beside Costpoint on the read-only access the customer's own administrator grants.
One more line from the 2026.3 notes is worth knowing: Deltek says Dela "can now operate as a Model Context Protocol (MCP) server, enabling authorized external AI tools to connect to Costpoint through a standard interface," so that tools such as Copilot can "access Dela capabilities and interact with Costpoint data and workflows," and calls the release "the foundation for cross-platform AI integrations." The notes do not say how those tools authenticate or which data they can reach, so treat it as a path to evaluate, not a decision made.
Four things not to automate first
Anything that touches CUI. Indirect rates, payroll and billing data are sensitive business data but usually not CUI; contract deliverables and technical data may be, and they change where the AI is allowed to run. Sorting that out is a prerequisite, not a first project. See Is it safe to put CUI or contract documents into ChatGPT or Copilot?
Anything that posts. Covered above.
Proposals. They are the most tempting target and the worst first one: irregular cadence, data in three systems, output that goes to the government, and pricing that depends on rates you have not yet made trustworthy. Deltek has also announced its own proposal generation for Costpoint and GovWin IQ, so the ground is moving. Do rates and burn first; proposal work inherits the data access, the sourcing habit and the log.
"All our data in a chatbot." The project that starts with everything stalls on data and permissions before it answers one question well. We covered why in How do I build AI solutions with my company's own data?
What the first month should produce
One question, answered on its schedule, with the source under every figure and a log of who asked. A controller who has checked the answers against the reports for several cycles and stopped running the manual version. A count of the hours the question used to take. And a plain decision at the end: scale to the second question, or stop. "Small proof first, then scale. We earn trust one task at a time." If the first task cannot earn that trust in a month, the problem is the task, and a bigger project would not have fixed it.
Common questions
Shouldn't a government contractor start with proposals, since that is where the revenue is?
Proposals are where the pressure is, but they are a poor first task. The data spans GovWin IQ, a proposal library and pricing history, the output goes to the government, and the pricing side depends on indirect rates you should be able to trust first. Start with rates and burn; the proposal work inherits a data foundation and a habit of sourced answers.
If Deltek is adding AI agents to Costpoint, why build anything separate?
Deltek's Dela features work inside Costpoint on in-system transactions such as voucher approval, bill review and anomaly flags. The first questions here span Costpoint and documents outside it, such as the provisional billing rate letter, funding modifications and the program manager's forecast. As of early October 2026, Deltek's Dela FAQ also stated that Dela is not available in its GovCon Cloud Moderate environment.
Does a first AI project on Costpoint data need DCAA approval?
No rule requires it. As of October 2026, DFARS 252.242-7006 and the FAR records clauses do not mention AI. Keep the first task read-only, make every figure traceable to a Costpoint report, and keep a log of who asked what, and you can explain it to an auditor in one page when asked.
How do we know whether the first AI task worked?
Measure the hours the recurring question used to take and what it takes now, count how often a person caught a wrong or unsupported answer, and ask whether the controller trusts it enough to stop running the manual version alongside it. If the manual version is still running after a quarter, the task has not been automated yet.