Written by: Kacper Osiewalski, Lead Backend Engineer, Digital Colliers
DeepSeek just told the market it's raising prices after months of undercutting the frontier vendors. If you run AI workloads in iGaming and your unit economics quietly assumed that pricing would hold, your budget for 2026 is now wrong. The fix isn't to shop for the next cheap provider. It's to actually measure what a single automated decision costs you, and what it earns.
Most teams I talk to can tell you their monthly token spend down to the pound. Very few can tell you the fully loaded cost of one RG intervention decision or one fraud score. Those are different numbers, and only one of them matters when a vendor moves the price on you.
What a decision actually costs
A decision is the atomic unit of an AI workload in production. One affordability check. One deposit risk score. One responsible gambling nudge. Cost per decision should include every input that touched that outcome, not just the final model call.
A usable model has at least these terms:
- Inference cost (prompt tokens plus completion tokens, at current list price, not the price you signed at)
- Retrieval cost (vector DB reads, embeddings if you regenerate them, any enrichment API calls)
- Human-in-the-loop cost (analyst minutes when the decision escalates, weighted by escalation rate)
- Infra overhead (queueing, logging, the observability stack, allocated per decision)
- Rework cost (what you spend when the model is wrong and you have to re-decide)
If you can't compute these five per decision type, you can't rationally respond to a price hike from any vendor. You're guessing.
Instrumenting it without a six-month project
You don't need a data platform rebuild. You need a decision log. Every automated decision your platform makes should emit one row with a stable decision_id, the decision_type, the model and version, token counts in and out, latency, the downstream outcome, and whether a human touched it later.
A few practical notes from patterns I keep seeing:
- Log token counts at the call site, not from vendor invoices. Invoices arrive too late to steer anything.
- Tag every decision with the regulatory regime it fell under. UK affordability logic that triggers at £150 net deposits over 30 days is a different cost curve than a discretionary marketing prediction.
- Store the raw prompt and response, hashed if you must, so you can replay decisions when a cheaper model comes out. Replay is the only honest way to compare vendors.
- Track the escalation rate per decision type as its own metric. A 3% escalation rate to a compliance analyst can dominate your unit cost if analyst time is billed at compliance-team rates. Kindred publicly disclosed £14M in compliance-team costs for 2023. That's the denominator you're really competing with.
The cost of doing nothing
Here is the part operators keep underweighting. The downside of a bad decision in iGaming isn't a rerun cost. It's a regulatory cost.
UK AML penalties for the most serious breaches reach up to 15% of gross gaming yield. GDPR fines run up to €20M or 4% of global turnover. EU AI Act penalties for high-risk violations reach up to €15M or 3% of global turnover, and the high-risk obligations kick in from 2 December 2027. Around one in four UK-licensed operators fails to hit a satisfactory AML rating on first assessment. If your RG or fraud model is quietly cheap because it's quietly wrong, the price hike is not your biggest problem.
A cost-per-decision model lets you see this. A team that knows the true loaded cost of a false negative on affordability will happily pay more per inference for a better model. A team that only knows token spend will chase the cheapest vendor and eat the fine later.
What the operators who'll win in 2026 are doing
The pattern is consistent. They treat every AI decision as a logged, replayable event with a known cost, a known outcome, and a known regulatory tag. They can swap models in a week because their evaluation harness runs against real historical decisions, not synthetic prompts. When a vendor raises prices, they run the numbers the same afternoon and either absorb it, switch, or route by decision type.
That capability is boring. It's a decision log, an eval set, and a spreadsheet with honest inputs. It's also the thing that separates operators who compound from operators who panic every time a vendor updates a price page.