By Susan Beall, Founder & CEO, Decerio
A webinar landed in my inbox last week. The topic was Using Claude for Budgeting and Financial Planning. The pitch was that a general-purpose AI assistant can help finance teams organize information, evaluate assumptions, and explain outcomes. That pitch is partly right. It is also exactly why GovCon CFOs, controllers, and FP&A leads need to be careful about what they take from a session like that.
Generative AI is genuinely useful in finance. We use it ourselves. AI is a remarkable interface for narrating variances, summarizing board materials, drafting commentary, and pressure-testing assumptions in plain language. If your team is not using AI yet, they will be soon. And they should be.
There is a unstated leap inside the “AI for finance” pitch, though. It moves from AI as the interface a finance team uses, to AI as the engine a finance team relies on. That leap is where GovCon gets into trouble. The question that exposes it is sharper than asking whether the AI is any good.
What is your AI talking to?
Why That Is the Right Question
Every AI answer is the product of two things. The model that wrote it, and the system the model had access to. The model gets all the credit and most of the marketing. The system determines whether the answer is defensible.
Think about what must be true when a CFO commits to a forward-pricing rate, a provisional billing rate, or a multi-year financial plan. Six months later, through a DCAA audit, a reforecast, a board challenge, or a contracting officer’s request, three things must hold:
The number must be reproducible from the same inputs.
The number must be traceable through assumptions, allocation bases, and cost pools.
The number must be consistent with every other scenario the team has run, with the disclosed practice on file, and with what was committed last quarter.
A general-purpose LLM, on its own, does none of those three reliably. The AI is fine. Reliability of that kind was simply never the design intent.
Three Things a Chat Window Does Not Have
Reproducibility. Large language models are non-deterministic by design. Ask the same question twice and the answer may differ. That property is a feature when you are drafting a board narrative. It is disqualifying when you are committing a rate to a customer or a regulator. An FP&A model is reproducible by construction. Same inputs, same structure, same output. That property is the price of entry for any number that will face an audit.
Traceability. A defensible budget number traces backward through a chain. Source data, then assumption, then allocation base, then cost pool, then output. When DCAA asks how you arrived at a G&A rate, the answer is a structure, not a paragraph. LLMs are excellent at producing the paragraph. They were not designed to enforce, or even to consistently represent, the structure. A wrap rate that sounds reasonable can still be structurally non-compliant under FAR Part 31 or CAS. The auditor does not care that the math checked out in isolation. They care whether the bases, pools, and allocations were applied consistently with your disclosed practice.
State. A budget branches. What if we add five FTEs to the G&A pool? What if cost shifts from overhead to G&A? What if award timing slips a quarter? What if we win the recompete but are not awarded an option year? Each is a scenario. Scenarios have to agree with each other and with the base plan. Without a shared source of truth, what looks like scenario planning is really a set of plausible variations with no guarantee of internal consistency.
The fair counter to all of this is that modern AI workflows can call external systems through tool use and retrieval. That counter is correct, and that counter is exactly the point. The state, the structure, and the reproducibility all live in the system the AI talks to. Take the system away and you do not have a copilot. You have an assistant working from a blank page.
The Company That Builds Claude Says the Same Thing
This is not a competitive jab. Anthropic, the maker of Claude, treats financial use of its model as a high-risk category in its own usage policy. In the consumer-facing context where the policy is most explicit, Anthropic requires human-in-the-loop oversight and qualified professional review before any output is finalized or disseminated. Across all contexts, Anthropic markets Claude as an assistant rather than as a system of record.
If the vendor building the model is asking for a qualified human in the loop before finalization, the right follow-up question is not whether you should use AI in finance. It is what the qualified human is reviewing against. In GovCon, the honest answer has to be a governed model. A model that encodes pool integrity, FAR Part 31, disclosed practice, and indirect rate structure. Otherwise, the human is reviewing prose against prose.
Even “Real” FP&A Platforms Are Not Built for What GovCon Needs
There is a related trap worth pulling apart from the AI one. Most established FP&A and CPM platforms, the names you would put on an evaluation shortlist, are built horizontally. They support close, plan, consolidate, and report for general finance use across small, mid-market, and enterprise. They are excellent at what they do.
What they typically are not built to do is encode the specific structure DCAA expects from a government contractor. That structure includes indirect cost pool integrity, FAR Part 31 cost principles, disclosed-practice consistency, forward pricing rate logic, provisional billing rates, and scenario-driven indirect rate modeling that holds up to audit. We see this every quarter. The largest GovCon we onboarded recently, a top-20 IT contractor, did not choose us because horizontal platforms could not handle their scale. They chose us because horizontal platforms could not model their indirect rates fast enough across the variety of scenarios they had to defend. GovCon compliance had to be native rather than configured on top of a general model.
Audit defensibility in this industry is foundational. It belongs in the data model, not in a feature flag.
AI as Interface, Governed Model as Engine
The right architecture for AI in GovCon finance is layered. A governed calculation engine owns the rate structure, the pools and allocations, the disclosed practice, and the scenario lineage. AI sits on top as the interface that translates between human questions and model results.
That is what we are building at Decerio. The Decerio engine encodes DCAA-defensible indirect rate structure as a foundational data model, not a configuration layer. Decerio integrates directly with your ERP for actuals and, when you want it, with your CRM for pipeline data, and runs full pipeline pricing inside the platform. The Decerio API exposes that governed model to the tools your team already uses: Power BI dashboards, AI assistants like Claude, Copilot, and ChatGPT, and Excel exports for ad-hoc work. Decerio becomes the source. Your stack continues to operate. The numbers underneath it become reproducible, traceable, and defensible.
In practice, a Controller can ask what the G&A rate looks like if the team adds five FTEs to the pool starting in Q3, and the answer comes from the engine. It comes with a version, a lineage, and a path back to the disclosed practice it was built against. Same question tomorrow, same answer. Plus, a record of who asked, what changed, and what the auditor would see.
For the top 20 IT contractor mentioned above, indirect rate scenarios that used to require multiple spreadsheet versions and days of analyst time now run in minutes against a single source of truth. Every scenario carries a full audit trail.
What a GovCon Finance Leader Should Consider
Three things.
- AI is real, it is here, and it is going to reshape how finance teams work. Encourage your team to use AI for narrative, summarization, and exploration.
- A budget, especially one that drives indirect rates, pricing decisions, or audit-exposed numbers, is a model. It is not a paragraph. Treat the engine and the interface as separate problems.
- When AI gives your team an answer about a rate, an allocation, or a forecast, the right question is not whether the answer is good. The right question is what the AI is talking to, and whether you could defend that answer to DCAA tomorrow. If you cannot answer that second question with confidence, the answer to the first one does not matter much.
A Working Session, Not a Sales Pitch
If you are working through this question for your own organization, we would rather show you the difference than describe it. In a 30-minute demo, we will take one of your real indirect rate questions and walk through it as the Decerio engine would answer it. You will see the lineage, the scenario branching, and the audit path attached. The answer comes ready for DCAA submission.
About the author
Susan Beall is the Founder and CEO of Decerio, the GovCon FP&A platform purpose-built for indirect rate modeling, forward pricing, scenario planning, and the operational intelligence finance teams need to lead. Decerio integrates directly with your ERP and CRM, maintains DCAA compliance, undergoes annual SOC 2 Type II audits, and serves small, mid-market, and enterprise government contractors.
