AI Budget Allocation

In many budget meetings, the AI line gets a number before anyone has chosen a specific problem to solve. The figure is set by comparison with a competitor or by a vendor's estimate, and a year later nobody can say what exactly that money made better.

An AI budget differs from a software budget in one fundamental way. With conventional software, most of the cost sits in purchase and implementation and then subsides. Here the cost continues after go-live, because data ages, the model needs retraining, and the process around it requires continual adjustment.

Where the budget usually leaks

The recurring pattern of waste is simple: a large share of the money goes to licences and infrastructure, while data preparation is handed to the existing team as a side task. The result is an expensive tool running on incomplete and inconsistent data, producing output nobody can rely on; the cost was paid but no value was created. The second pattern is buying capacity that is only needed at peak and sits idle the rest of the year.

Four real cost headings

• Data preparation: collection, cleaning, standardisation and labelling; in most projects the heaviest heading and the most underestimated.
• Integration: connecting the solution to existing systems such as finance, inventory or customer management, which rarely gets its own line.
• Compute and licences: servers, GPUs or model services; part of the picture, not the whole of it.
• Maintenance and retraining: the running cost after go-live, for monitoring output quality and refreshing the model with new data.

Staged funding instead of an annual figure

The mechanism that genuinely contains risk is staged allocation: a limited amount to prove value, then a larger amount only if a pre-agreed metric is met. Its importance lies in making a stop decision respectable. When the whole budget is committed at once, continuing a fruitless project looks cheaper to the organisation than admitting a mistake, and that is precisely where sunk cost accumulates.

Why this matters more for Iranian businesses

Two factors make budgeting errors costlier in Iran. First, currency volatility invalidates hardware and external-service estimates within a few months, so long-term foreign-currency commitments carry double risk. Second, access to certain cloud services is constrained and may change mid-project. Both point to the same conclusion: budgets should move in short steps, stay open to revision, and lean as far as possible on infrastructure that is actually available.

A three-month allocation path

• Month one: pick one high-frequency process with measurable cost and estimate what it costs today, broken down into labour, error and rework. Without that baseline number, no return calculation means anything.
• Month two: set a firm ceiling for the pilot and direct most of it to data preparation. Write the stop criterion at the same time, before work begins.
• Month three: compare the result against the baseline. If break-even sits within a reasonable horizon, fund the next stage; if not, stop the project or change its scope.

Costly budgeting mistakes

• Leaving maintenance out of the initial calculation; a project that justifies itself in year one can turn loss-making in year two.
• Computing return on the basis of hypothetical labour savings when nothing in the cost structure actually changes.
• Spreading the budget across several small projects at once; resources thin out and none reaches the threshold where results appear.
• Ignoring the time cost of staff who must prepare data and review output; that cost is real even when it never appears on an invoice.

How to measure the return

Choose a cost metric that was measurable before the project too, such as cost per case, process cycle time or rework rate. Then place break-even on the calendar: from which month does cumulative saving overtake cumulative cost. If that question has no clear answer, the project is not yet ready to receive a budget.

Frequently asked questions

• What share of the budget should go to data?
  There is no fixed figure, but if data preparation is budgeted below licences, the estimate is very likely optimistic.
• Buy a ready service or build in-house?
  For generic use cases a ready service is faster and cheaper; where your competitive advantage lives inside the process itself, building in-house becomes justified.
• What if the budget is very tight?
  Narrow the scope, not the data quality; one narrow process with clean data returns more than a broad project on messy data.

Takeaway

An AI budget starts working when it stops being an annual line and becomes a chain of small decisions each with a clear stop criterion. Establish the baseline number, take the data share seriously, and tie funding for each stage to the result of the one before it.

Glossary

• Total cost of ownership: The combined cost of buying, implementing and maintaining a solution over its useful life.
• Break-even point: The moment cumulative saving or revenue equals cumulative cost.
• Sunk cost: Money already spent that cannot be recovered and should not influence a continue-or-stop decision.
• Staged funding: Step-by-step financing in which each stage is conditional on the result of the previous one.
• Data preparation: Cleaning, standardising and labelling data before it is used by a model.
• Model retraining: Periodically refreshing a model with new data so its accuracy does not decay as conditions change.
• Pilot: A small, low-cost version built to prove value before the main investment.