Why do AI pilots at government contractors stall, and how do you avoid it?

Deltek's 17th annual Clarity study (May 2026, 917 government contractors) found that among IT and security respondents 37 percent described their AI as piloting, 38 percent as scaling and 5 percent as fully developed, and that 45 percent of respondents were unclear on the return on investment. Kevin Plexico, Deltek's SVP of Information Solutions, called the gap between adoption and execution the defining challenge. Unanet and CohnReznick's 2026 GAUGE report (June 2026, 1,204 GovCon professionals) found 70 percent of contractors using AI, up from 33 percent two years earlier. Contractors are adopting fast, and most are still short of fully developed. Where they stall is the same five places.

Stalling is the normal outcome, and the reasons are documented

In July 2024 Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, "due to poor data quality, inadequate risk controls, escalating costs or unclear business value." Gartner's undated article on the same subject, read October 30, 2026, says that "by the end of last year" at least 50 percent were, without saying which year, and lists five failure points: lack of business value, data that is not ready, escalating total cost of ownership, responsible AI as an afterthought, and poor change management. In February 2025 Gartner added that through 2026 organizations would abandon 60 percent of AI projects unsupported by AI-ready data.

RAND's August 2024 report, built on interviews with 65 experienced data scientists and engineers, found five leading root causes in industry: stakeholders misunderstand or miscommunicate what problem needs to be solved, the organization lacks the data to train an effective model, the project "focuses more on using the latest and greatest technology than on solving real problems," the infrastructure to manage data and deploy models is missing, and the technology is applied to problems too hard for it. Its first two recommendations: make sure the technical staff understand the purpose and domain of the project, and commit a team to a specific problem for at least a year.

McKinsey's State of AI in 2026 (1,719 respondents, surveyed May 4 to June 8, 2026) found 37 percent of organizations attributing any positive EBIT impact to AI, 6 percent qualifying as high performers, and one in five limiting AI use because of operating costs. Its reading of the stall: "The limiting factor is increasingly the organization's ability to absorb change." The high performers redesign the workflow rather than inserting AI into the existing one, are twice as likely to say their senior leaders demonstrate commitment, and have defined processes to measure impact.

The five places a pilot stalls at a Costpoint shop

It started with a tool. The board said "do something with AI," someone bought licenses or a demo, and the search for a problem began afterward. That is RAND's third root cause, and it is why the pilot never had a success condition: nobody wrote down which question it was supposed to answer, so nobody could say whether it did. The fix is cheap. Pick the question first. We sorted the candidates in What should a government contractor automate with AI first?: the recurring, read-only questions finance already answers by hand from Costpoint.

The security review started with nothing classified. The reviewer was handed a tool and asked whether it was safe, with no list of what it would read. With no classification, the reviewer has to assume CUI and the review stalls on FedRAMP and DFARS 252.204-7012 questions that may not apply. Indirect rates, payroll and billing figures are sensitive business data but usually not CUI; contract deliverables and technical data may be. Write down which Costpoint reports the pilot reads and whether any can hold CUI before the review begins, and most finance pilots turn out to be a question about access, logging and vendor terms rather than a government cloud. We drew the line in Is it safe to put CUI or contract documents into ChatGPT or Copilot?

Data access was requested last. Deltek's integration overview says Costpoint 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 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. None of that need be slow, but it is sequential, and a pilot that files it after the security review and the vendor demo has spent its calendar before the first real row of data arrives. File it on day one. Every path is granted by you, not by Deltek, and no Deltek partnership is involved.

The demo ran on sample data, and nobody checked a figure. The vendor's demo answered questions about a fictional company fluently. Nobody asked it where year-to-date overhead sat against the provisional billing rate for a pool the controller already knew. Deltek's Clarity respondents named difficulty validating AI-provided insights (26 percent) and the potential for inaccurate forecasts (29 percent) among their top concerns, and an unchecked demo is how those concerns stay unanswered. An answer without a source cannot be checked, and an unchecked answer cannot be trusted with a billing. Every answer shows where it came from. No black boxes.

Nobody owned the decision. The pilot had a sponsor, a project manager and a vendor, and no one whose monthly work would change. At the end there was no date on which someone had to say keep, fix or stop, so it stayed open. That is the change management failure Gartner lists fifth and the leadership commitment McKinsey's high performers report. The pilots that ship are owned by the person who answers the question by hand today.

Mo 1Mo 2Mo 3Mo 4Mo 5Mo 6

The pilot that stalls

1. Pick a tool, then look for a problem to give it4 weeks
2. Security review with nothing classified6 weeks
3. Costpoint access requested late, then waited on4 weeks
4. Demo on sample data; nobody checks a figure4 weeks
5. Waiting for a decision nobody owns6 weeks, open

Six months, no decision

The first task that ships

1. Name one question the controller answers by hand; classify its data1 week
2. Administrator files the read-only web services request2 weeks
3. Build; every answer shows its source4 weeks
4. Check a month of answers against known figures4 weeks
5. Owner decides: keep, fix or stop1 week

Three months to a decision

Illustrative steps and durations, not measurements from any client. The stalled pilot's five rows run end to end for six months without a decision; the first task's five rows run three months and end in one. The only real figure here is Deltek's "up to 5 business days" for a cloud web services request, which fits inside the second teal row.

Waiting for Deltek is also a stall

There is a sixth place, particular to Deltek shops: the pilot is paused to see what the next release brings. Deltek ships four quarterly Costpoint releases a year, and the 2026.2 through 2026.4 release notes describe Dela doing more in-system work each quarter: AP voucher approvals, bill review in the Agentic Financial Close tab, financial and indirect rate anomaly cards, timesheet auto-posting. For in-system transactions that is the right place to look, and we said so in Deltek Dela, Copilot or a custom assistant: which does a contractor actually need?

But a question that spans Costpoint and the documents outside it (the funding modification in SharePoint, the subcontract, the proposal library) is not on the release schedule, and for GovCon Cloud Moderate customers Deltek's own documents disagree about Dela as of October 2026: the undated Dela FAQ in the Costpoint 8.2 help says Dela is not available in GCCM, while Deltek's August 2025 Costpoint Dela/GenAI overview says that for GCCM it runs on Microsoft's OpenAI service in Azure GCC High, so ask your Deltek account team which applies to your tenant. Waiting for a feature that is not announced is a stall with a better excuse. Read each release's notes, turn on what answers your question, and build the rest.

The pilot that reaches a decision

Reverse the five stalls and you have the plan. Name one question a controller or contracts manager answers by hand every month, and write down what a correct answer looks like. Classify the data that question touches before anyone reviews a tool. Have your administrator file the Costpoint access request on day one, read-only, under your license. Build so that every answer carries its source report and date, and keep a log of who asked what, because that log is what you show an auditor; we covered the records side in Can AI read our Costpoint data without creating a DCAA problem? Run real answers through at least one monthly close against figures a person already knows. Cap the model spend in your own cloud account, which answers Gartner's escalating-cost point and McKinsey's one in five. Then, on a date fixed at the start, the owner decides: keep, fix or stop.

Stop is a fine answer. A pilot that ends with a clear no after a close's worth of checked answers has done its job; the one that stalls has not answered anything. Small proof first, then scale. We earn trust one task at a time. What that first task costs, line by line, is in How much does a custom AI assistant on Costpoint data cost?

Common questions

If Gartner says 30 to 50 percent of generative AI projects get abandoned after proof of concept, is that an argument to wait?

It is an argument to pilot differently, not later. Gartner's stated reasons for abandonment (poor data quality, inadequate risk controls, escalating costs, unclear business value) are all things a contractor controls before the pilot starts. A read-only question your controller already answers by hand, with the data classified, the Costpoint access requested on day one, a source under every answer and a decision date, removes each of them. Waiting removes none of them, and McKinsey's 2026 survey found 44 percent of organizations now scaling AI across the enterprise, up from 38 percent a year earlier.

Our pilot stalled in security review. What went wrong?

Almost always, nobody classified the data before the review began, so the reviewer had to assume the worst. Indirect rates, payroll and billing data are sensitive business data but usually not CUI; contract deliverables and technical data may be, and a SaaS AI tool that can reach CUI is an external system under NIST SP 800-171 3.1.20 and a cloud service provider in scope under 32 CFR 170.19. Write down which Costpoint reports the pilot reads and whether any of them can hold CUI before the review starts, and the reviewer has a question that can be answered. For most finance questions the answer is no CUI, and the review becomes a check of access, logging and vendor terms.

Should the pilot belong to IT or to finance?

The person who answers the question by hand today should own the pilot, with IT and security as partners. RAND's interviews found that industry AI projects fail most often because stakeholders misunderstand or miscommunicate what problem needs to be solved, and McKinsey's high performers are twice as likely as others to say their senior leaders demonstrate commitment to AI initiatives. A controller who owns the question can tell whether an answer is right; a central team cannot. The owner also sets the decision date and makes the call to keep, fix or stop.

Does waiting for Deltek's next Dela release count as a stall?

It can. Deltek ships four quarterly Costpoint releases a year and the 2026.2 through 2026.4 notes describe Dela doing more in-system work each time (voucher approvals, bill review, financial and indirect rate anomaly cards). If your question is in-system, read the release notes and turn the feature on. If it spans Costpoint and documents outside it, or your tenant is GovCon Cloud Moderate, where as of October 2026 Deltek's undated Dela FAQ says Dela is not available and its August 2025 Dela/GenAI overview says it runs there on Azure GCC High (ask your Deltek account team which applies), nothing on the release schedule answers it, and waiting is a stall with a better excuse.