CFOs approving AI spend should demand a locked baseline, a full multi-year cost model, and a payback threshold before signing off, the same discipline applied to any other capital request. Most AI proposals fail that bar not because the technology doesn't work, but because the business case around it is vague: no baseline, no named workflow, no kill criteria if the numbers don't hold up.
This isn't a note about Zamp (we are an AI accountant company, so we're obviously biased toward AI adoption). It's a checklist for the finance leader who has to say yes or no to a request on their desk right now, and needs to know what questions separate a real ROI case from a demo dressed up as one.
Note on naming: this article is about the CFO's job of evaluating AI investment. It has nothing to do with "Zamp HR" or payroll products that share a similar name, and nothing to do with zamp.com, the US sales-tax compliance platform. Different companies, different products.
The pattern shows up across finance organizations the same way every time. A team pilots a tool, likes the demo, and brings a request to the CFO framed around a capability ("we want an AI agent for X") rather than a financial outcome ("this cuts invoice processing cost from $12 per invoice to $4, saving $180K a year at current volume"). Capability requests are hard to evaluate because they don't have a number attached. Outcome requests do.
The fix is procedural, not technical: require every AI spend request to name one workflow, one baseline metric, and one target delta before it reaches a spending decision.
Ask what this initiative maps to on the actual strategic plan, and what a named process, cost center, or revenue line looks like today without it. Then ask the harder question: what's the cost of doing nothing? Competitive pressure, margin erosion, and compliance exposure are real costs even when they don't show up on a line item, so they belong in the comparison.
One more filter worth applying: will this workflow still exist in three years, or is it a candidate for redesign or elimination? Funding AI to automate a process that's about to be restructured anyway is money spent on the wrong problem.
Reject any proposal that describes a capability instead of a problem. "We want AI for accounts payable" is not a business case. "AP processing costs $340K a year in headcount and rework, with a 4% exception rate that adds 6 days to close" is. Every request should arrive with a number, not a noun.
Before anything ships, lock the baseline: current cost per unit of work, current cycle time, current error rate, current volume. Without this, you cannot isolate what the AI actually changed from what would have happened anyway (seasonality, headcount changes, a parallel process improvement). Teams that skip this step end up arguing about impact after the fact with no way to settle it.
The sticker price on an AI tool is rarely the real cost. Build the full model:
A 36-month TCO view is the minimum. Anything shorter hides the maintenance and change-management cost that shows up in year two.
Run the same financial discipline you'd apply to a capital equipment purchase:
Payback expectations for AI initiatives generally land in the 18 to 24 month range for a first deployment, tightening to 9 to 14 months for teams with a mature deployment playbook already in place. If a proposal only shows the optimistic case, ask for the other two before approving anything.
An AI tool that nobody uses has an ROI of zero regardless of what the model can technically do. Require an adoption curve: expected usage at 30, 90, 180, and 365 days, with evidence from a comparable rollout, plus a named executive sponsor who owns the business outcome (cost, revenue, risk), not just the technical delivery. If nobody can say who is accountable for adoption actually happening, the ROI number in the proposal is a guess.
Every AI vendor relationship carries execution risk, data privacy exposure, and lock-in risk. Price these into the model rather than treating them as a footnote. Reasonable portfolio guardrails: cap total AI spend as a share of revenue, cap any single vendor's share of the AI budget, and cap variable/usage-based cost exposure so a spike in API calls can't blow through the budget unnoticed.
Fund in stages, not as one lump sum. Set milestones and a kill criterion, for example, re-approval is required if actual ROI misses the target by more than 30% at a gate. This turns AI spend into the same kind of disciplined capital allocation process as any other investment, instead of a one-time bet that's hard to unwind if it doesn't work.
Aggregate savings numbers hide problems. Track cost per transaction (per invoice, per ticket, per document) before and after, and track cost per accepted outcome, meaning output that ships without rework. A high rework rate quietly erodes ROI even when the top-line savings number looks fine, so it's worth asking for this metric specifically rather than assuming the aggregate number tells the full story.
The CFO's job here isn't to become the resident AI expert. It's to apply the same capital-allocation discipline that already exists for every other spend category: a named workflow, a locked baseline, a real cost model, a payback threshold, and a kill criterion if it misses. Teams that bring proposals in that shape get faster approvals because there's nothing left to argue about. Teams that bring capability pitches get sent back for more work, which is the correct outcome.
This budget-approval discipline sits inside a bigger shift: finance functions are increasingly running core processes through AI accountant systems that handle reconciliation, journal entries, and close tasks end to end. The CFO evaluating a specific AI spend request and the CFO evaluating whether to bring in an AI accountant more broadly are answering the same underlying question: does this workflow, at this cost, produce a return we can measure and defend. The checklist above works whether the request is a single-point tool or a broader finance automation initiative.
What should a CFO ask before approving an AI budget request? Ask for the specific workflow being automated, the current baseline cost and error rate for that workflow, the full multi-year cost of ownership (not just the license fee), and the payback period under a pessimistic scenario. If a proposal can't answer these with real numbers, it isn't ready for approval.
What's a reasonable payback period for enterprise AI spend? Most first-time AI deployments target an 18 to 24 month payback. Teams with a mature AI deployment process, meaning they've already done this once and have a repeatable rollout playbook, can often get that down to 9 to 14 months.
How much should a company spend on AI as a share of revenue? There's no universal number, but many finance teams cap total AI spend in the range of 1.5% to 4% of revenue depending on industry and maturity, with additional guardrails on how much of that goes to any single vendor to avoid lock-in.
Why do most AI ROI cases fail to hold up after deployment? Usually because there was no locked baseline before deployment, so nobody can isolate what the AI changed versus what would have happened anyway. The fix is procedural: lock the baseline metrics before the tool goes live, not after.
