In 2025-2026, the real question is “why not build it with Claude in a weekend?” Fair question. We've watched enough teams try this to write the arc from memory. Some should build. Most shouldn't. Here's the honest math, and where Novaplan fits.
Two filters. If the first one fits, the math is on your side. If the second one describes your team, keep reading.
Build if you have spare engineering capacity that isn't tied to revenue features, your needs are unusual enough that no vendor can model them, or you have a structural reason to own the IP - acquisition optionality, vendor risk policy, a board mandate.
If two of those three are true, the math is on your side and the rest of this page isn't for you.
The question isn't really “can we build this?” Your team can. The question is what it costs to build and keep running, year after year, after the original engineers have moved on.
That second half is where the math breaks.
It breaks at three places, every time. Maintenance hours that nobody scoped. Connector drift that nobody owns. A hiring overlap the labor market doesn't supply.
Maintenance, fixes, the next request from finance, the integration that broke when the upstream API changed. Teams scope the build. They almost never scope the operate.
NetSuite ships breaking changes quarterly. So does Workday. So does HiBob. Each one is a small project. None of them is on the original PRD.
FP&A plus engineering is a rare combination. The person who can do both is also the person every consulting firm wants to hire. You won't replace them quickly when they leave.
Honest answer: yes, Claude builds the surface layer fast - a chart, a chatbot, a dashboard. That part is real. But an FP&A system a CFO can defend to an auditor has four more layers underneath the surface.
What you are actually buying is the head start - plus the team that ships your custom 10% in days through AI-first build, plus the FP&A and AI advisory included in the agreement, plus an MCP layer so your Claude can drive the system you already paid for.
The thing you would build with Claude is one of the five layers. The four underneath are the moat - and the part nobody scopes upfront.
Hand this list to whoever's going to build it. If they can't answer all five with specifics on a whiteboard, the architecture isn't there yet. Claude will not help them improvise their way out of any of these.
“When an April salary change applies retroactively to Q1 and Q1 has already been board-reported, how does the system handle the restatement without overwriting Q1’s prior history?”
Why it mattersThe bi-temporal data model question. Every value needs two timestamps: when it’s about and when it was first known. Claude generates CRUD scaffolding instantly. Reasoning about effective dates as a separate axis from knowledge dates is a 200-page spec your engineer will Google and improvise. Most improvised answers fail audit.
Cost when it failsRestated board numbers. Auditor questions you can’t answer. A month of forensic cleanup. Your credibility with the board.
“Show me on a whiteboard how the consolidation reconciles USD reporting to AUD source data: which exchange rate applies where, how every amount rolls up to one base currency, and how the variance cross-term is locked so two readers never get different answers.”
Why it mattersThe deterministic computation question. Multi-currency consolidation is the kind of thing engineers Google. They pick a rate convention; the auditor uses a different one; the numbers stop tying at year-end. Variance decomposition has the same trap: rate × volume × mix has a residual cross-term, and if two readers use different conventions for it, two readers get different answers.
Cost when it failsNumbers that don’t tie across reports. Audit findings. CFO career risk - this is the kind of thing that costs people their jobs.
“How will a junior analyst running a department-total spend report fail to back into my VP’s salary from the visible aggregate?”
Why it mattersThe field-level RLS plus aggregation-aware masking question. Role-based access is easy - Claude ships that in an afternoon. Preventing leakage through aggregates - so a "department total" view can’t be diffed against a "department total excluding one employee" view - is a research-grade problem. Engineers will not solve it. They’ll add a permission flag and move on.
Cost when it failsHR crisis. Legal exposure. A salary that leaks through a chart in a Slack channel and ends up in a lawsuit.
“It’s month 14. The engineer who built this just took the offer from Stripe. The next engineer opens the codebase and says "this is unmaintainable, we should rebuild." What do we do?”
Why it mattersThe compounding-maintenance question. Internal tools become legacy in 18 months. The architect is the only one who can debug it. When they leave, knowledge walks out the door. The next hire will quote a rebuild rather than learn an unfamiliar system - because that’s how engineers protect their careers.
Cost when it failsRe-fund the project at the worst possible time. Two engineers instead of one. Another year of "we’re almost there." Your FP&A team back in Excel for the duration.
“When the close breaks at 11 PM the night before the board meeting, who do we call?”
Why it mattersThe accountability question. There’s no on-call rotation for an internal tool. There’s no SLA. Your engineer is asleep, or out, or already past their retention bonus. The thing your finance team depends on most has no support contract.
Cost when it failsReputational - the slow-bleed kind. The board meeting where the numbers aren’t ready. The CFO who has to explain "we built this ourselves."
The honest framing
Claude is genuinely great at building a chatbot, a form, a dashboard. None of those are what fails when a finance system breaks. What fails is the bi-temporal data model, the multi-currency math, the audit trail, the permissions, and the on-call when your engineer is asleep. Those five things are what you pay us for - and they're what no engineer building with Claude in their first attempt is going to get right. We've already gotten them right, on production data, with the regrets baked in.
FP&A isn't one job. It's two. The platform handles the half that's undifferentiated drudgery. Your team and your favorite LLM handle the half that actually moves the business. We help you set up both.
The boring half (undifferentiated, recurring, error-prone)
None of this is what your FP&A team became finance professionals to do. None of it differentiates your business. All of it has to be right or the next conversation is with an auditor.
The insight half (differentiated, judgment-driven, where AI shines)
This is what your team became finance professionals to do. This is where AI multiplies the right person, instead of replacing them.
The reframe in one line
The right question isn't “should we buy Novaplan or use Claude?” It's “how do we get our FP&A team out of reconciliation work and into insight work, with AI doing both halves well?”
We do that two ways. First, the platform owns the boring half so your team stops spending 70% of their week on it. Second, we open the platform up via MCP so your team can point Claude, Cursor, ChatGPT, or whatever they prefer at the audited financials - and we coach them on the patterns that work. Most FP&A teams have never used AI well because nobody around them has. That part of the agreement is the part that compounds.
The build column is what every team ships when they actually try this. We've checked it against the teams we've watched fail and the ones we've watched succeed.
Before the line-by-line, the punchline
Build it internally (Claude-accelerated)
$1.2M-$2.5M
Even with LLM-assisted scaffolding cutting the v0 build by 30-40%, you still own the four hard layers, 5 years of maintenance, and the rehire when your engineer leaves
Anaplan / Adaptive / Pigment
$300K-$750K
License + implementation + the change orders the original SOW didn't cover
Novaplan Head Start path
$50K-$150K
60-day Head Start + Year-1 founding rate + 4 years at standard rate, no multi-year lock-in
LLM acceleration moves the build column number, not the gap. The 5-year maintenance, the GL connector drift, and the rehire when your engineer leaves are unchanged by Claude.
| Dimension | Build it yourselves | Novaplan |
|---|---|---|
| Time to first BvA cycle | 9-18 months | 14-28 days |
| How implementation actually happens | LLM-assisted, but your team configures and maintains | AI-first build via MCP write tools - our agents configure your stack at machine speed |
| Year-one engineering cost | 1-2 FTEs ongoing | Annual license, no seats |
| GL connector ownership | You. Forever. | Vendor maintains |
| AI inside the system (not on top) | Build smart insights, plan-from-chat, MCP access yourself | All shipped. Insights, planning, MCP for your favorite LLM, grounded in your data |
| When the close breaks at 11 PM | A Slack channel and a hope | A phone number |
| What your engineers stop shipping | Revenue features | Nothing |
| Custom features as needs evolve | Re-scope, re-fund, re-hire | You ship them with my guidance, or I ship them in days |
| Compounding upgrades | Re-fund the project | Renewal price |
| Walk-away if it does not fit | Sunk-cost paralysis | 60-day pilot, no obligation, fee credits to Year-1 |
Month zero is kickoff energy. By month six, the data model works for eighty percent of cases and the last twenty percent is where actual finance lives. By month twelve, engineers are quietly back on revenue features. By month eighteen, someone in finance pulls up a vendor in a meeting and says, this is what we were trying to build.
The full essay covers the 18-month arc month by month, the questions to ask before you commit a quarter, and where this gets personal. Written for the engaged reader, not the buyer in evaluation mode.
Read the full essay10-minute read - no email required