- In-house and outsourced desks both charge by seats and hours, including quiet nights; an AI agent charges by contact, so nights, peaks and new languages barely move the cost.
- Compare every offer — human or AI — as cost per resolved contact on your own hourly volume curve, including the human team AI still needs.
- Manual QA covers around 2% of conversations, and with a BPO the provider often scores itself; scoring all conversations on one rubric lets you compare in-house, outsourced and AI fairly.
- Neither BPO nor AI lets you claim data never leaves; ask for tokenisation, EU endpoints, zero retention and an agent deployed in your infrastructure.
- For most operators the realistic answer is hybrid: AI for routine Tier-1 around the clock, a smaller human team for RG, VIPs, disputes and fraud — with the rules and test sets owned by you.
Sooner or later every operator asks the same question: should casino customer support be outsourced, kept in-house, or handed to an AI agent? The question is usually framed as a price comparison. It is really a comparison of three cost structures, three ways of controlling quality and three kinds of lock-in — and for most operators the honest answer is a combination of them, not a single winner.
This buying guide lays out the options side by side, explains where each one fails in iGaming specifically, and ends with a decision guide you can apply to your own desk.
The four options
- In-house desk. Your own agents, your own team leads, your own tools. Maximum control and product knowledge, highest fixed cost, slowest to scale into new languages.
- Outsourced desk (BPO). A specialist provider staffs your queues, often from a lower-cost location, under your brand. Fast to scale and flexible on seats, but your rules live in someone else's training materials.
- AI agent. Software that reads the player's account, answers in the player's language and executes a short list of validated actions, with humans handling everything sensitive.
- Hybrid. An AI agent takes the routine Tier-1 bulk around the clock; a smaller human team — in-house, outsourced or both — handles escalations, VIPs, responsible gambling, disputes and quality.
The first three are rarely pure in practice. Most in-house desks already outsource a night shift or a language; most BPO contracts already include some automation. The useful question is which part of the queue each model should own.
Cost structure: seats versus contacts
A 24/7 desk of 30–40 seats costs a mid-size casino €15–30K a month, and it is staffed for peak all year. Whether those seats are yours or a provider's changes the hourly rate and the contract terms, but not the unit of cost: you pay for seats and hours, including the quiet ones.
That unit is a poor match for casino traffic. Volume spikes around paydays, reload promotions, weekend afternoons and technical incidents, and drops sharply overnight. In one operator's history only a few percent of chats arrived outside staffed hours — yet a night shift is a full shift, and every new market adds native speakers for every hour you cover.
An AI agent changes the unit to cost per contact. The marginal cost of the 3 a.m. chat, the promotion spike and the extra language is close to the cost of the model call. What stays fixed is the human team for the contacts AI must never take, plus the effort of keeping the AI's quality measured. If you want to run the numbers on your own volumes, the ROI calculator does it in a couple of minutes.
A BPO quote per seat-hour and an AI quote per conversation are not comparable until you translate both into cost per resolved contact, at your real hourly volume curve, including the human team the AI still needs. Ask every vendor — human or AI — for that number on your own data.
Side-by-side comparison
| In-house | Outsourced (BPO) | AI agent | Hybrid | |
|---|---|---|---|---|
| Cost unit | Seats, salaries, management | Seats or hours | Contacts | Contacts + a smaller team |
| Night shift | Expensive, often cut | Cheaper, still a shift | Native, with a narrower mandate | AI overnight, humans on call |
| New language | Hire and train | Provider hires and trains | Configuration + calibration | Configuration + calibration |
| Peaks | Queues and concurrency overload | Contracted headroom | Elastic | Elastic for Tier-1 |
| Product and rules knowledge | Deepest | Depends on training and churn | Exactly as encoded — no more, no less | Rules in code, judgement in people |
| Quality coverage | Sampled (~2%) | Sampled, often by the provider | Can be 100% scored | 100% scored, humans included |
| RG, VIP, disputes | Handled | Handled, with handoff risk | Must escalate | Escalated to your best people |
| Lock-in | People and know-how | Contract terms, notice periods | Prompts, rules, test sets | Lowest, if you own the artefacts |
Quality control: who is scoring whom?
In most desks, in-house or outsourced, quality is measured the same way: a quality manager reads a sample of conversations against a scorecard. Realistically that is around 2% of conversations. With a BPO there is an extra layer: the sample is often read by the provider's own QA team, against criteria the provider helped write, and reported back to you monthly. Nobody is acting in bad faith, but the party being measured is also the party measuring.
An AI agent forces the question into the open, because nobody trusts an unmeasured bot. Done properly, every conversation — AI and human, in-house and outsourced — is scored on the same rubric, calibrated against your own QA lead's judgement. That is often the most valuable side effect of introducing AI: for the first time, you can compare your outsourced desk, your in-house team and your bot on one scale. The method is in iGaming support quality control software, and a ready-to-use rubric in the iGaming support QA scorecard template.
Languages and the night shift
These are the two lines where outsourcing traditionally wins against an in-house desk — providers can find native speakers and staff nights more cheaply — and where AI changes the comparison most.
For a human desk, each new market means native speakers for every covered hour, a training cycle on your bonus and payment rules, and cover for holidays and churn. For an AI agent, a new language is a configuration change, a native-speaker review of tone, a test set and a shadow-mode period before release. The night shift becomes a policy rather than a payroll line: answers and account lookups stay automatic, account writes queue for the morning, and responsible-gambling signals trigger your on-call rule. Details are in 24/7 multilingual casino customer support automation and, for market-specific tone, localising casino support by market.
Humans still matter here. Native speakers review tone per language, and somebody must own the 3 a.m. responsible-gambling escalation. But those are a handful of reviewers and an on-call rota, not thirty seats.
Data, compliance and lock-in
Every option sends player data somewhere. With a BPO, real people at a third party see full player records, usually through your helpdesk and back office, under a data processing agreement. With an AI agent, text is sent to a model provider — and the honest questions are whether player identifiers are tokenised before they leave, whether the model runs on EU endpoints with zero data retention, and whether the agent itself runs inside your infrastructure so account reads and audit logs stay with you. Neither model lets you claim that data "never leaves"; both can be done responsibly. See how PlayerDesk handles data and EU AI Act and GDPR for AI player support.
Lock-in looks different too. With a BPO it sits in contract terms, notice periods and the trained team you would have to rebuild. With an AI vendor it sits in the artefacts: the intent map, the escalation rules, the knowledge base, the validators and the hand-labelled test set. Insist that these belong to you and are delivered in a usable format at contract end. If they do, switching vendors is a project, not a restart.
Whoever runs your desk, the rules, the rubric and the test set should belong to you.
See which part of your queue each model should own.
Send an anonymised chat export; get topics, languages, hours and the share of the queue that is automatable on your book.
A decision guide
| If your situation is… | The realistic answer |
|---|---|
| Small volume, one or two languages, a team that knows the product well | Stay in-house; add AI for nights and peaks first |
| Fast expansion into new markets | AI for Tier-1 in each new language; outsource or hire only for escalations once the market proves itself |
| An existing BPO contract you are broadly happy with | Keep it for Tier-2 and escalations; move routine Tier-1 to AI at renewal; score both on one rubric |
| Heavy VIP or high-stakes player base | Invest in the human VIP and escalation team; let AI clear the routine queue so they have time for it |
| Strict data constraints from a regulator or group policy | AI agent deployed in your infrastructure with tokenisation and EU endpoints; keep human access to full records minimal |
In almost every row the answer is hybrid. Not because hybrid is fashionable, but because the casino support queue has two very different halves: a large, repetitive, rules-driven half that should cost per contact, and a small, high-stakes half — responsible gambling, VIPs, disputes, fraud — that deserves your best people. Paying human seat rates for the first half is what makes the second half underfunded.
If you go hybrid, roll the AI out the safe way: shadow mode first, then supervised, then autonomous topic by topic, as described in how to automate iGaming customer support.
The short version
In-house gives you control, outsourcing gives you flexible seats, and an AI agent gives you a cost per contact that doesn't care about night shifts, peaks or new languages. None of them alone fits a casino support queue. Let AI take the routine Tier-1 bulk around the clock, keep a smaller human team — yours, a provider's or both — for everything sensitive, score all of them on one rubric, and make sure the rules and test sets belong to you. For the bigger picture, start with AI player support for iGaming operators, or read what separates a real agent from a FAQ bot in AI chatbot for online casinos.
Frequently asked questions
Is it cheaper to outsource casino customer support or use AI?
It depends on volume shape. Outsourcing lowers the hourly rate but you still pay for seats and hours, including quiet nights and every language you cover. An AI agent costs per contact, so nights, peaks and new languages add little. Compare both as cost per resolved contact on your own hourly volume, including the human team AI still needs for escalations.
Can AI replace an outsourced casino support team completely?
No, and it should not try. Responsible gambling, self-exclusion, VIP-sensitive matters, disputes, complaints, fraud and AML must stay with people. AI can take the routine Tier-1 bulk such as bonus, withdrawal and KYC status questions, which lets a smaller human team focus on the cases that need judgement.
How do you control the quality of an outsourced support desk?
Score conversations yourself, against your own rubric, rather than relying only on the provider's internal QA sample. Automated quality control can score every conversation on the same scale for in-house agents, outsourced agents and AI, once it has been calibrated against your own QA lead's judgement.
What is a hybrid AI and human casino support model?
An AI agent handles routine Tier-1 contacts in every language around the clock, with a narrower mandate at night, while a human team, in-house, outsourced or both, handles escalations, VIPs, responsible gambling, disputes and quality review. All conversations are scored on one rubric.
How do you avoid vendor lock-in with an AI support provider?
Make sure the intent map, escalation rules, knowledge base, action validators and hand-labelled test set belong to you and are delivered in a usable format at contract end. If you own those artefacts, switching vendors is a project rather than a restart.
Last reviewed 9 October 2026. Figures marked as industry ranges are ranges, not Arctura claims; operator-specific numbers are only quoted after validation on the operator's own baseline.
