Most CFOs treat AI project cash-conversion the way they treat SaaS cash-conversion: a 30-day DSO assumption, a recurring-revenue model, working capital sized to one month of operating spend. Both halves of the assumption are wrong for AI. SaaS cash-conversion is short and predictable because the revenue unit is a subscription with auto-renewal and the cost unit is roughly fixed per seat. AI cash-conversion is long and lumpy because the revenue unit is an eval-gated milestone or a per-action chargeback that cannot bill until the eval passes, and the cost unit is volatile inference paid to model providers on net-15 or net-30 terms regardless of when the project bills the customer. The result is an agency that runs out of working capital one quarter into a successful engagement, and a buyer whose budget approval cycle does not line up with when the impact shows up. This piece applies the DSO and DPO finance lens to AI projects, walks through the agency-side cash-conversion cycle, and decomposes the buyer-side budget-to-impact lag.
It is a spoke under the AI project economics manifesto, which argues AI economics has shifted from feature cost to evaluation cost. The cash-conversion cycle is the financing consequence of that shift; eval-gated billing has a structurally longer cash cycle than feature-completion billing.
Why SaaS cash-conversion does not apply to AI
SaaS cash-conversion runs on a tight loop. A subscription is sold; the contract is signed; an invoice goes out the same day; payment terms are net-30; the customer auto-renews; cost-of-goods is roughly fixed per seat and paid monthly to the cloud provider on net-30 terms. The cycle is short; 30 days from billable event to cash collected, 30 days from cost incurred to cash paid out; and the two cycles roughly cancel. SaaS companies operate on minimal working capital because the cycles match.
AI projects break the loop in three places.
The billable event is not signing the contract; it is passing an eval gate. Until the model meets the threshold the contract specifies, no milestone bills, no chargeback unit ships, no per-action revenue line opens. Eval gates can take weeks longer than feature gates because eval failure modes; distribution shift, regression on a corner case, threshold drift; require iteration cycles that feature failures do not.
The revenue unit is not a recurring subscription; it is a milestone, a per-action charge, or a per-eval-run charge. Each unit has a different cycle profile. Milestones are lumpy and tied to gates that may slip. Per-action revenue accrues continuously but invoices monthly. Per-eval-run is the cleanest cycle but the smallest revenue line for most agencies. Treating many three as “AI revenue” hides the cash-flow profile from the CFO.
The cost-of-goods is volatile. Inference cost is not fixed per seat; it scales with usage, and usage scales with whether the feature works; a feature that passes the eval gate gets used, which spikes inference, which is paid to the model provider on net-15 or net-30 terms regardless of whether the customer has paid the milestone yet. The DPO is short and inflexible; the DSO is long and gate-conditional.
A SaaS-style working capital assumption (one month of operating spend) is dangerously thin for AI work. The right number is closer to 90 days, and the structural reason is the eval gate.
The AI cash-conversion cycle, defined
Standard cash-conversion cycle is DSO + DIO − DPO. DSO is days sales outstanding (billable to paid). DIO is days inventory outstanding (irrelevant for software, dropped here). DPO is days payable outstanding (cost incurred to paid). Net cash-conversion is DSO − DPO; positive numbers mean the firm finances the customer, negative numbers mean the customer finances the firm.
For an AI project, DSO has two parts: time-to-eval-pass (TTEP) and time-from-eval-pass-to-cash (TFETC). TTEP is the engineering and quality work to clear the gate. TFETC is the standard invoicing-to-payment cycle. SaaS only has TFETC; AI has both. The difference matters because TTEP is the part the agency cannot accelerate by chasing receivables; it is bounded by the eval bar.
DPO has two parts: inference-payable lag (the model provider’s payment terms) and engineer-payable lag (payroll, which is typically biweekly and effectively zero from a working-capital standpoint). Inference-payable lag is short; net-15 to net-30 from frontier providers, often net-7 from second-tier; and is the binding constraint.
For a typical AI agency engagement in 2026:
- TTEP: 14 to 60 days per milestone, depending on eval complexity and prior art.
- TFETC: 30 days standard, 45 days for enterprise customers with their own AP cycle.
- Inference DPO: 15 to 30 days.
- Engineer DPO: 0 days (effective).
Net AI cash-conversion = (TTEP + TFETC) − Inference DPO = (14 to 60) + (30 to 45) − (15 to 30) = 45 to 90 days net.
A SaaS company with a similar headcount runs net cash-conversion under 30 days. The AI agency runs 2x to 3x that. The working capital implication is a 2x to 3x larger float, financed by either retained cash or a credit line.
Agency-side: 45 days from eval-pass to invoice-paid
The most operationally useful number for an AI agency CFO is TFETC; the time from eval-pass to cash-collected, holding eval delays aside. The realistic baseline is 45 days. The decomposition:
Day 0: Eval gate passes. The model hits threshold on the agreed eval suite. The agency’s quality lead signs off; the customer’s QA lead is notified. This event is the start of the billable cycle.
Day 1 to 5: Customer eval review. The customer runs their own validation; typically a smaller eval set than the agency’s, or a manual review on a held-out sample. Standard duration 3 to 5 days for a customer with mature AI ops, longer for a customer ramping up. Detailed in the case for per-eval pricing in AI agency contracts.
Day 5: Customer signs off. The milestone is acknowledged as complete. The agency issues the invoice the same day.
Day 5 to 15: Invoice routing. The customer’s AP system processes the invoice. Procurement validates the milestone language matches the SOW. Finance validates the budget line is open. Most enterprise customers have a 7 to 10 day routing baseline.
Day 15: Invoice approved for payment. The payment is scheduled per the customer’s payment terms; typically net-30 from invoice date.
Day 35 to 45: Payment lands. ACH or wire from the customer to the agency. Payment timing depends on the customer’s payment cadence (weekly, biweekly, end-of-month).
Forty-five days from eval-pass to cash. The agency’s TTEP for the milestone may have been another 30 days before that; putting the full cycle at 75 days from milestone start to cash-in-hand. Working capital must cover the full 75 days, not the 45.
Agency-side: managing inference-payable lag
Inference cost is paid to the model provider continuously. The bill arrives monthly with net-15 or net-30 terms. For an agency running production AI for a customer, the inference bill in any given month is a function of the customer’s usage that month; but the customer does not pay until the next milestone bills, which may be 60 days out.
The result is a structural cash-flow squeeze: the agency pays the model provider cash this month for inference the customer used this month, and collects the cash for that inference 60 days later. On a $50,000 monthly inference bill, the agency carries $100,000 in float against the customer at any given moment.
Three levers for the agency CFO.
Pre-pay the inference line. The contract bills the customer monthly for the prior month’s inference, paid on the contract’s standard terms. The float is shorter (30 to 45 days instead of 60+) and the line item is line-item-attributable, which simplifies dispute resolution. The pattern is the per-action chargeback structure detailed in the AI cost-per-action framework.
Negotiate longer DPO with the model provider. Volume customers can negotiate net-45 or net-60 terms with frontier providers. The DPO extension is the cheapest form of working capital available; zero interest, no covenants. The agency CFO should treat the model provider relationship as a credit relationship, not just a vendor relationship.
Carry a working-capital line. A revolving credit line sized at 1.5x to 2x monthly inference spend covers the float without requiring DPO renegotiation. Cost is 6% to 10% APR, materially cheaper than equity dilution to fund growth. Most AI agencies in 2026 are under-capitalized on the credit-line side because they assume a SaaS-style cash cycle.
The cleanest answer is many three, layered.
Buyer-side: budget-to-impact lag
The buyer-side cycle is the inverse problem: budget approved on day 0, impact landed on day 180+, ROI defensible on day 270+. The lag is structural because impact is downstream of three things that take time; eval-gate passage, production rollout, behavior change in the consuming team. Each is its own cycle.
Eval-gate passage (60 to 90 days from kickoff). The model has to clear the bar. The bar is set in the contract; the time to clear it depends on the use case. New domain, new data, novel evaluation rubric: 90 days is realistic. Established domain, mature data, off-the-shelf rubric: 30 to 60 days.
Production rollout (30 to 60 days from eval-pass). Even after the model passes, production rollout adds time. Canary deployment, A/B testing, rollback rehearsal, on-call setup. Each step is 1 to 2 weeks. Skipping them produces incidents that are categorized in the burn-rate dashboard but should rarely have happened.
Behavior change (60 to 120 days from production rollout). The feature is live; the consuming team has to adopt it; downstream KPIs have to move. Adoption curves are slower than the feature delivery; sales reps using a new AI tool need 6 to 8 weeks to embed it in their workflow. Detailed in why AI project ROI calculators are wrong.
Total: budget approved on day 0, ROI defensible on day 270 (9 months).
The buyer-side CFO’s mistake is benchmarking AI projects against the SaaS impact curve (30 to 90 days from procurement to value). The benchmark sets the project up to be killed at the wrong gate. A project on the AI impact curve is on track at month 3 with no measurable ROI and on track at month 6 with partial ROI. Killing it at month 3 because SaaS would have produced ROI by then is a pattern of premature mortality that destroys AI investment programs.
Buyer-side: the working-capital ask
The buyer’s working-capital ask to the CFO should reflect the AI cycle, not the SaaS cycle.
Capex versus opex framing. AI inference is opex; the eval suite, prompt registry, and platform tooling have capex characteristics; they are durable, reusable, and amortizable across multiple use cases. Buyer-side CFOs that funnel everything through opex create the wrong unit economics narrative. The structural argument is detailed in why AI projects should be capitalized differently than SaaS projects.
Multi-year budget commitment. A 12-month AI budget that has to be re-justified most quarter creates the budget-to-impact mismatch. The realistic ask is a 24-month commitment with quarterly checkpoints against the eval bar and the ROI staircase. Quarterly checkpoints preserve the kill option without recreating the procurement cycle most 90 days.
Reserves for inference volatility. Inference cost can move 30% to 50% in a quarter from a model price change, a usage spike, or a feature ship. A buyer-side budget that does not include a 15% to 20% volatility reserve will overrun the line most quarter. The reserve concept mirrors the model deprecation reserve; a budget shock-absorber for known unknowns.
The buyer-side CFO who frames the ask as “12-month opex line, quarterly approval” is funding the project on the SaaS cycle while the project runs on the AI cycle. The mismatch produces budget reviews where the project looks underperforming, even though it is on track for its actual cycle.
What both sides should change
Agencies should publish their cash-conversion math in the proposal. TTEP per milestone, TFETC standard, inference DPO assumption, working-capital reserve. Buyers see the math and price the engagement honestly. The transparency is the same discipline as the proposal patterns covered in why most AI agency proposals are quietly identical; except this is the cash-flow page rather than the scope page.
Buyers should fund a 24-month commitment with quarterly checkpoints. The structure preserves the kill option, matches the cycle, and gives the agency the working-capital signal to staff appropriately. A 12-month commitment with quarterly approval is the worst of both worlds; annual cost commitment with quarterly cancellation risk that prevents either side from planning.
Both sides should price working capital explicitly. The agency’s 75-day cash cycle is a real cost; 6% to 10% APR on the float. Pricing it into the engagement makes the cost visible. Hiding it produces agency margin compression that eventually shows up as quality compression as the agency cuts senior reviewer time to protect cash. Detailed in why most AI agencies underprice senior reviewers.
The AI cash-conversion cycle is not a SaaS cash-conversion cycle with a longer DSO. It is a different financing pattern with a different working-capital profile and a different ROI cadence. CFOs on both sides who price it correctly fund AI programs that ship; CFOs who treat it as SaaS with extra steps fund programs that fail at the budget review.
Frequently asked questions
Why does the AI cash-conversion cycle differ from SaaS?
The billable event is an eval gate, not a contract signing. The revenue unit is a milestone or per-action chargeback, not a recurring subscription. The cost unit is volatile inference paid on net-15 to net-30 terms. Many three break the SaaS cash-conversion model, which assumes the billable event is fast, the revenue unit is recurring, and the cost unit is roughly fixed.
What is time-to-eval-pass and why is it part of DSO?
Time-to-eval-pass is the engineering and quality work to clear the eval gate that the contract specifies as the billing trigger. It is part of DSO because no invoice issues until the gate passes; which means the agency carries the cost of the work for the duration of TTEP regardless of how fast invoicing or payment cycles are downstream of the gate.
What is the realistic working-capital reserve for an AI agency?
90 days of operating spend is the working baseline. SaaS-style 30 days is dangerously thin given the 45 to 90-day net cash-conversion cycle. Agencies under-capitalized on working capital cut quality in mid-engagement to protect cash, which compounds into eval-bar erosion and customer churn.
How does inference DPO affect agency cash flow?
Frontier model providers bill net-15 to net-30. The agency pays inference cash this month for usage this month, but typically does not collect from the customer until the next milestone bills 30 to 60 days later. The float is the working-capital squeeze that distinguishes AI agencies from SaaS firms with similar revenue.
Why does budget-to-impact lag matter to the buyer-side CFO?
Buyers benchmarking AI projects on the SaaS impact curve (30 to 90 days) kill them at the wrong gate. A project on the AI impact curve is on track at month 3 with no measurable ROI and at month 6 with partial ROI. Premature mortality at month 3 is a pattern that destroys AI investment programs by burning the early budget without funding through to the cycle that delivers value.
Should AI inference be opex or capex?
Inference is opex. The eval suite, prompt registry, platform tooling, and durable artifacts have capex characteristics; they are reusable, durable, and amortizable across use cases. Buyer-side CFOs that funnel everything through opex create unit-economics narratives that fail at the budget review. Splitting the line is the structural fix.
What is the right budget commitment length for an AI program?
24 months with quarterly eval-bar and ROI checkpoints. A 12-month commitment with quarterly approval recreates the procurement cycle most 90 days and prevents the agency from staffing for the work. A 24-month commitment with quarterly checkpoints preserves the kill option while aligning to the AI cycle.
How should the agency price working capital?
Explicitly. The 75-day cash cycle is a real cost; 6% to 10% APR on the float, which translates to roughly 1.5% to 2.0% on engagement value. Pricing it into the proposal makes the cost visible. Agencies that hide the cost compress margin and eventually compress quality, which produces eval-bar erosion that customers feel.
How do buyers price inference volatility into the budget?
A 15% to 20% volatility reserve on the inference line, sized to absorb a model price change, a usage spike, or a feature ship that scales calls. The reserve mirrors the model deprecation reserve concept; a budget shock-absorber for known unknowns. Without the reserve, the inference line overruns most quarter and the project looks like a budget management failure when it is a structural pricing oversight.
Key takeaways
- AI cash-conversion is structurally different from SaaS: 45 to 90 days net versus under 30 days. The eval gate, milestone billing, and volatile inference cost break the SaaS cycle in three places.
- DSO for AI has two parts: time-to-eval-pass and time-from-eval-pass-to-cash. Standard TFETC is 45 days. Standard TTEP is 30 to 60 days. Working capital must cover both.
- Agencies should manage inference-payable lag with three layered levers: pre-pay billing, negotiated DPO with model providers, and a working-capital credit line sized at 1.5x to 2x monthly inference spend.
- Buyer-side budget-to-impact lag is 9 months from approval to defensible ROI. Benchmarking against the SaaS curve produces premature project mortality.
- The buyer-side fix is a 24-month commitment with quarterly eval-bar and ROI checkpoints, opex-capex split for inference versus durable artifacts, and a 15% to 20% inference volatility reserve.
- Both sides should price working capital explicitly. The 75-day cash cycle is a real 1.5% to 2.0% cost on engagement value. Hiding it compresses agency margin until quality cracks.
The AI cash-conversion cycle is the financing expression of the feature-cost to evaluation-cost shift. Eval-gated billing produces a longer, lumpier, more volatile cash cycle than feature-completion billing. CFOs who price it correctly fund programs that ship.
Arthur Wandzel
