Cameron Stubbs • Sep 23, 2026 • 8 min read
AI Agents for Crypto Exchange Customer Support
Yes, an AI agent can handle a real share of crypto exchange support volume today. No, it should not touch anything that moves money or closes a dispute without a human signing off.
That split sounds obvious once you say it out loud, but most of what gets published on this topic skips straight past it. Search for this exact phrase and the results are vendor platforms selling their own tooling (Kommunicate, Lorikeet) or generic "what are AI agents" explainers aimed at token buyers rather than support teams. None of them are written by anyone who has actually had to defend an automated decision to a compliance officer, which is the part that matters if you run a UK-facing exchange.
Here is the version of this written from the build side, not the sales side.
What Actually Counts as a Support Ticket at a Crypto Exchange
Generic customer support advice does not map cleanly onto exchange support, because the ticket mix is different from a normal e-commerce or SaaS queue. A working exchange support inbox splits roughly into five categories:
Status lookups. "Where is my withdrawal" or "why hasn't my deposit confirmed" make up the bulk of ticket volume at most exchanges, often 40% or more. These are deterministic questions with a factual answer sitting in a database or on-chain.
KYC and verification chasing. A user submitted documents three days ago and wants to know what is happening. The agent's job here is retrieval and status, not judgement.
Fee, balance, and general account questions. "What's your withdrawal fee for this network" or "why was I charged this much." Answerable from a knowledge base or account record.
Fraud and account-takeover reports. A user believes their account was accessed without permission, or a withdrawal was made they did not authorise. High stakes, time-sensitive, and carries real legal weight.
Disputes and account recovery. Locked accounts, failed transactions that need investigation, and anything where the "right answer" requires judgement about what actually happened.
The first three categories are where automation earns its keep. The last two are where it needs to stop.
Which Tickets Are Safe to Automate, and Which Aren't
The framework Fracas uses when we build agent layers for clients, whether that's a Discord moderation queue or an exchange support inbox, is the same: agents handle triage and retrieval, humans handle anything with financial or legal consequence attached.
Automate fully:
Status lookups against transaction hashes and on-chain confirmation counts. This is retrieval, not decision-making. An agent that can query a block explorer or your own indexer and report back "your deposit has 8 of 12 required confirmations" removes a huge share of ticket volume with zero risk.
KYC document status checks and reminders. "Your documents are under review, expected turnaround is 24 to 48 hours" is a status report, not a decision about whether to approve someone.
Fee and product FAQ answers pulled from a maintained knowledge base.
Triage only, human approves the action:
Fraud reports. An agent can gather the details, pull the account's recent activity, flag the pattern (new device, new location, large withdrawal request in quick succession) and hand a fully-contextualised case to a human. What it should not do is freeze or unfreeze the account itself.
Withdrawal reversal or delay requests. The agent investigates and summarises. A person makes the call and the record shows who made it and why.
Keep entirely human:
Account recovery for locked or compromised accounts. Identity verification at this stage carries real consequences if it goes wrong in either direction: lock out a genuine user, or hand an account to an attacker.
Dispute resolution where the facts are contested. If a user says a trade executed at the wrong price and support logs do not immediately settle it, that needs a person weighing evidence, not a model generating a plausible-sounding resolution.
This is the same shape of decision Fracas applies when building community moderation agents for token projects: automate the triage, keep the judgement calls human. The stakes are higher here because tickets touch real balances, but the underlying architecture question is identical.
The UK Compliance Layer Most Guides Skip Entirely
Every vendor comparison piece on this topic mentions KYC, AML, and SOC 2 in the same three sentences and moves on. None of the ones we read mention the FCA angle at all, which matters if your exchange is UK-facing.
If your business is registered with the FCA's cryptoasset AML register, your complaint handling sits inside the same Consumer Duty expectations that apply to any other regulated financial activity. That means: fair treatment, clear communication, and a decision trail you can produce if the Financial Ombudsman Service or the FCA itself asks how a complaint was handled.
An agent that auto-closes a fraud report because it scored low on some internal confidence threshold is a Consumer Duty problem waiting to surface, even if the agent's call was correct, because there is no human decision behind it to point to. The fix is cheap: log every agent decision, log every escalation, and make sure a named person signed off on anything that resulted in an account action or a fund movement. This is the same principle Fracas applies to financial promotion content under FCA PS23/6: keep a named human accountable for anything with regulatory weight.
If you are not FCA-registered and operate purely offshore, this section matters less to you directly, but it is worth knowing it exists, because UK users will increasingly expect it and competitors who get it right will use it as a trust signal.
Build vs Buy: What This Actually Costs
Outcome-priced SaaS platforms for crypto support (Lorikeet, Fin by Intercom, and similar) charge in the region of $0.80 per resolved chat or email ticket as of their published 2026 rate card, rising to $1.20 to $1.50 for a resolved voice conversation. Compare that to a human-handled ticket, which typically costs $1.25 to $4 depending on your support team's location and setup.
That math favours automation quickly at volume, but the pricing only tells you what a resolved ticket costs, not whether the platform's guardrails match how your compliance team wants decisions made. Off-the-shelf platforms are built for a generic regulated-finance customer, not specifically for the FCA's cryptoasset regime or your custody provider's API.
A custom-built agent scoped narrowly, status lookups and KYC chasing only, typically runs £6,000 to £18,000 to build, with running costs (model API calls, hosting) usually under £200 a month once live. Multi-workflow builds covering fraud triage and dispute summarisation on top of status lookups start closer to £25,000. Maintained retainers for tuning and expansion run £1,500 to £5,000 a month, similar to the ranges we've quoted for Web3 project agent builds generally.
The decision usually comes down to two questions: is your ticket mix generic enough that a platform's default guardrails fit, and do you need the agent to reason over data (on-chain events, custody-provider webhooks) that a generic platform was not built to touch. If either answer points toward "no," build.
How to Start Without Creating a Regulatory Problem
Pick one workflow. Status lookups are almost always the right first choice: high volume, zero judgement required, and a wrong answer is embarrassing rather than dangerous.
Write down exactly what data the agent needs (transaction hash, confirmation count, deposit address) and where it comes from. If the answer touches your own indexer or a block explorer API, that's a well-defined build. If it requires guessing at intent, it is not ready to automate yet.
Set the escalation rule before you write a line of code: what confidence threshold sends a ticket to a human, and who reviews the agent's decisions in the first month. Every escalated output should get human review for at least four weeks before you trust the agent unsupervised on that workflow.
Only once the first workflow is running reliably, with a clean escalation record you would be comfortable showing a regulator, should you add fraud triage or a second workflow on top.
One thing you can do this week without touching any code: pull your last month of support tickets and sort them into the five categories above. That single exercise tells you what percentage of your volume is genuinely automatable versus what needs to stay human, and it becomes the specification for whichever build path you choose.
If you want to talk through what's realistic for your ticket volume and your regulatory position, book a call and we'll give you a straight read on where to start. You can also see how we approach AI agent builds for Web3 teams more broadly.
Frequently Asked Questions
Can AI agents actually resolve crypto exchange support tickets?
Some, yes. Status lookups, KYC document chasing, fee and balance questions, and FAQ-style queries are safe to automate end to end. Anything that moves money, unlocks an account, or resolves a dispute needs a human to approve the final action, even if the agent does the investigation.
What is the UK compliance risk with AI support agents at a crypto exchange?
If your exchange is registered with the FCA under the cryptoasset AML regime, your complaint handling has to meet Consumer Duty standards. An agent that auto-closes a fraud report or auto-approves a withdrawal reversal without a documented human decision creates a record you cannot defend if a regulator or the Financial Ombudsman asks how the decision was made.
How much does it cost to build a support agent for a crypto exchange?
Outcome-priced SaaS platforms charge roughly $0.80 per resolved chat or email ticket, rising to $1.20 to $1.50 for voice, as of 2026 rate cards. A custom-built agent scoped to triage and status lookups typically costs £6,000 to £18,000 to build, with running costs under £200 a month once it is live. The right choice depends on ticket volume and how much of your workflow is genuinely standard versus exchange-specific.
Should a crypto exchange build its own support agent or buy a platform?
Buy if your ticket mix is mostly generic (status checks, FAQs) and you want to move fast. Build if you need the agent to reason over on-chain data, integrate with a custody provider like Fireblocks, or if your escalation rules are specific enough that a generic platform's guardrails do not match how your compliance team actually wants decisions made.