AI Funding Comes in Tranches

six months, the payback period increases to 3.45 years. If costs run 20% higher, it reaches 4.2 years. All figures are hypothetical and unrelated to the company’s actual projects. Funding for a project like this is released in tranches. Each tranche is tied to specific metrics that determine whether the project proceeds or is stopped.

I start with one question. What results would force the project to be stopped or redesigned? The team records the answer before launch and presents it with the budget calculations. The decision is made by the business head, CFO and CIO together. Finance is responsible for the acceptable payback period; the CIO is responsible for delivery and the pilot. At large companies and government agencies, AI budgets pass through several committees. Each committee examines its own area of risk.

Suppose customer inquiry processing costs 118 million rubles a year (about $1.46 million). The project reduces that cost to 79 million rubles (about $980,000). This includes 49 million rubles for licenses and computing (about $610,000), 20 million for the support team (about $250,000), and 10 million for human review (about $120,000). Launch costs 97 million rubles (about $1.2 million) and takes 17 months. Annual savings amount to 39 million rubles (about $480,000).

Licenses and equipment go into the budget estimate immediately. Other costs appear later. Data must be prepared and labeled. The system must be integrated with existing systems. Staff must be trained, and security must be checked. While the old and new systems run in parallel, the cost of this dual operation is accounted for separately. Delivery and support timelines, along with penalties for delays, are written into the contract. Contingency plans are prepared in advance, including whether a second supplier is available.

The pilot shows whether the system can handle the specific tasks required by the company. Two or three typical tasks are selected, along with real data. Acceptable error rates, response times and workload are agreed on before testing. The cost per accepted result is then calculated, including repeat attempts and human review. One example would be an inquiry processed correctly.

AI introduces risks that conventional systems do not have. Costs rise with usage, while quality can change when the model version or data changes. If the share of correct answers falls from 92% to 88%, human review is needed for 12% of inquiries instead of 8%, a 1.5-fold increase. The 10-million-ruble cost of human review (about $120,000) rises to 15 million rubles (about $185,000). The system now costs 84 million rubles a year (about $1.04 million), savings fall to 34 million rubles (about $420,000), and the payback period grows to 2.9 years. The project remains within the limit, but barely. Computing and human review costs rise with workload. If higher workload does not generate additional revenue, savings shrink. The cost per accepted result and quality on a benchmark set of tasks are therefore checked in every reporting period. The share of computing resources that leads to an accepted result is tracked as well. Everything else is treated as waste.

Operating costs are projected three to five years ahead, including energy, support, model updates, security and staffing. Errors are assessed separately. In this example, the supplier’s liability is capped at 25 million rubles (about $310,000), while potential damage is estimated at 310 million rubles (about $3.8 million). The company remains liable for the other 285 million rubles (about $3.5 million). That exposure is mitigated through human review, limits on the system’s authority, insurance and a rollback option. The rollback process is tested before launch.

Next, deviations are assessed. A six-month delay and an additional 18 million rubles in costs (about $220,000) push the payback period to 3.45 years. If costs rise by 20%, to about 95 million rubles (about $1.17 million), annual savings fall to 23 million rubles (about $280,000), and the payback period reaches 4.2 years. The acceptable payback period is approved during the budget review. In this example, the limit is three years. Meeting it requires annual savings of about 32 million rubles (about $400,000). The new system can cost up to roughly 86 million rubles a year (about $1.06 million); any cost increase above 8% pushes the project past the threshold. The baseline estimate meets the limit. Neither deviation does.

That is why a simple rule applies. Projects that stay below the approved threshold even under the worst-case deviation receive their full budgets. Other projects receive an initial tranche and a deadline by which their figures must be validated. Our project receives funding for the pilot and first phase. The second tranche depends on the cost per accepted result and quality on the benchmark task set.

When several projects are under consideration, they are compared by payback period under the worst-case deviation and by the cost of stopping them. A project with a longer payback period that can be stopped after a quarter without major losses is a safer bet than a faster-paying project that is expensive to exit.

Before approving the budget, ask the team how much the company would lose if the project were stopped after one quarter. If the team cannot name the amount immediately, it has not planned for termination.

Contributed by Stanislav Yezhov, Director of AI Development, PJSC Astra Group