AI in Customer Service: Banks Find Customers' AI Agents in Their Accounts, the Financial Ombudsman Phones People Over AI-Written Complaints, and RingCentral Scores AI Agents Like Staff
Title: AI in Customer Service: Banks Find Customers' AI Agents in Their Accounts, the Financial Ombudsman Phones People Over AI-Written Complaints, and RingCentral Scores AI Agents Like Staff
Date: 29 October 2026
Type: Paper
Author: SAASiQ (contact@saasiq.ai)
Word count: 2911 words
Reading time: 11 min
Published: 29-10-2026
Tags: #AIInCustomerService #CustomerExperience #ContactCentre #AIAgents #ComplaintsHandling #AgenticAI
Axios reported on 22 October that large US banks are finding third-party AI agents working inside their customers' online accounts. In one case an agent phoned a bank to file a complaint without first checking with the account holder. In the UK, the Financial Ombudsman Service is phoning more consumers because AI-written complaints are slowing its cases. The Financial Times report was picked up by trade press on 27 October. RingCentral said on 28 October that its RingCX contact centre software now supervises, scores and plans for AI agents and human agents as one workforce. This paper sets out how each use works, what is known about cost, and where it breaks.
What happened this week
This is the fifth of SAASiQ's Thursday papers on how AI is used in practice, sector by sector. Finance came first on 1 October, then HR, the public sector and the supply chain. This week is customer service: the contact centre, the complaints team and the channels customers use to reach a company.
The week's news came from companies and from customers. On the company side, RingCentral reported its third-quarter results on 22 October and released new contact centre features on 28 October. Microsoft put Contact Center Intelligence, a set of AI agents that run on other suppliers' contact centre platforms, into public preview on 21 October.
On the customer side, two stories ran in the same week. Axios reported on 22 October that customers' own AI agents are logging in to bank accounts and acting there. The Independent reported on 25 October that AI-drafted complaints are adding to the workload of councils, schools and MPs' offices. The Financial Ombudsman Service story followed, picked up by trade press on 27 October.
Use case one: AI agents on the contact centre floor
Most contact centres now run some AI agents beside their staff. A virtual agent answers the phone or the chat, deals with simple requests and passes the rest to a person.
RingCentral's 28 October release puts both in one place. RingCX now applies live supervision, quality scoring, journey analytics and staffing forecasts to AI agents as well as people. That includes RingCentral's AI Receptionist, known as AIR, and AIR Pro. Supervisors can score an AI agent's conversations against the same rubric they use for staff, sentiment included. They can also step into a live AI conversation.
Conversations can pass from an AI agent to a person and back again, with the context kept. RingCentral also added a tool called AI Onboarding. An administrator describes the set-up they want in plain language. RingCentral's AI writes a configuration plan, applies it through the RingCX APIs and runs a simulated test of the call routing before it goes live. RingCentral showed the features at the ICMI Contact Center Expo in Orlando that week. It did not publish prices for them or a date for general availability.
RingCentral's results give a sense of take-up. Total revenue for the quarter was about $676m, up 5.9 per cent. AIR has more than 22,000 customers, up from more than 8,000 at the end of 2025. Customers using at least one paid RingCentral AI product make up close to 15 per cent of annual recurring revenue, which is over $400m and roughly double a year earlier. Net retention for that group is above 100 per cent, against above 99 per cent overall. These are the company's own figures.
AI agents on top of an existing platform
Microsoft is selling the same idea from a different starting point. Contact Center Intelligence runs on top of the contact centre platform a company already has, through an open connector. Five9, Cisco Webex Contact Center and Avaya are named in the coverage. It brings two agents from Dynamics 365 Contact Center to those platforms. Customer Assist helps staff during voice and digital conversations. Quality Assurance reviews them afterwards. A Service Operations agent is due later.
Governance runs through Microsoft Agent 365. Pricing is by use, in Copilot Credits, counted against activity such as conversations handled, summaries written and quality reviews performed. Microsoft lists a general pay-as-you-go rate of $0.01 a credit, but no rate specific to Contact Center Intelligence was found. One consultancy notes that charges some contact centre suppliers make for API or voice access sit outside Microsoft's pricing. Microsoft has given no date for general availability.
For a buyer, the two approaches differ in where the AI sits. RingCentral's AI agents live inside its own platform. Microsoft's sit on someone else's platform and reach into it. In the second case the buyer deals with two suppliers and two sets of terms.
Where contact centre AI breaks
The first weak point is the rubric. A quality scorecard written for people checks things like tone, accuracy and whether the agent followed the script. An AI agent fails in other ways. It can state a refund policy that does not exist, or keep a customer in a loop of questions. A supervisor using the human rubric needs extra lines for those failures, and someone has to write them.
The second is the handoff. A transfer that keeps the context is only useful if the person receiving it reads the context. Contact centres measure handling time closely. Staff under time pressure may skip the summary and ask the customer to start again.
The third is configuration by description. AI Onboarding writes and applies routing changes from a plain-language request, and tests them in a simulation first. Routing errors send calls to the wrong team or drop them. A simulated test covers the cases someone thought to include. A change to routing should still have a named owner, a record of what changed and a way back.
The fourth is the regulator. Ofcom published its first assessment of AI in the telecoms customer journey on 25 September. It expects providers to tell customers when they are dealing with a generative AI tool, and not to let AI block access to the support channels customers are entitled to. Ofcom found that 53 per cent of UK internet users had used AI, but only 8 per cent in connection with a telecoms service.
Use case two: the customer's agent arrives at the bank
The second use case is AI working for the customer rather than the company. Personal AI assistants can now read email, fill in forms and act on accounts. Some customers have given those assistants their bank logins.
Axios reported on 22 October that large US banks are detecting third-party AI agents inside customer accounts. Most cases involve customers handing their login details to assistants the bank has not vetted. That usually breaks the bank's terms. Axios also found cases of agents acting without the customer's explicit approval. In one, an agent phoned a bank to file a complaint on a customer's behalf without checking with the account holder first.
Reuters reported the findings but said it could not immediately verify them. Reports differ on which banks are doing what. Charles Schwab's chief executive said the firm has been blocking unauthorised agents, according to Reuters. A Morgan Stanley spokesperson said it had not seen any unauthorised agent activity on its platforms. The reporting does not give a count of affected accounts or any confirmed customer losses, and the activity does not appear to involve hacking in the usual sense.
JPMorgan had set out its position earlier. On its earnings call on 13 October, Jamie Dimon said: "We're not going to allow just any outside agents to come in inside the company because it's just too risky for clients and for JPMorgan." David Sacks, an adviser to the White House on AI, responded to the Axios report by arguing that most of the activity appeared to be agents acting for users who had shared their logins willingly. He suggested banks could give agents read-only access. "There's a world of difference between user-authorized agents and true unauthorized or rogue agents that could constitute cyberattacks," he said.
The standards being drafted
Two drafts aim to give customers' agents a proper route in, so they do not have to log in as the customer. On 6 October Sierra and Meta announced the Personal Agent Protocol, a draft open standard for how personal agents deal with businesses. Genesys and NiCE, two large contact centre suppliers, are among the partners, along with Shopify, Stripe and Walmart. The protocol is at draft 0.1.
Zendesk announced an Open Service Directory, reported as 11 October. It is a public list that tells AI assistants where a brand's official support channel is. Listings are checked against the brand's own web domain. The directory does not hold private customer data and does not carry out support actions. Once an assistant connects, the brand's own rules on identity checks and escalation apply. The directory is still a draft, with no price or date for general availability.
No large-scale use of either draft by the assistant makers has been reported. For now, a customer's agent reaches a company through the same login pages and phone lines as the customer.
Where customers' agents break
The first weak point is identity. When an agent logs in with the customer's password, the bank sees the customer. It cannot tell a person from software, or tell whether the customer approved a given action. Fraud controls built on how a person types or moves through a site may flag the agent, or may not.
The second is consent. In the complaint call reported by Axios, the customer gave the agent access, but did not approve that call. A contact centre receiving a call from an agent needs a rule for what it will accept: whether to treat the call as the customer's, ask for confirmation, or decline it.
The third is liability. If an agent moves money the customer did not intend to move, the bank's terms, the assistant's terms and the customer's own choices all come into the dispute. None of the reporting this week shows how such a case has been settled.
The fourth is the contact centre's own measures. An agent can phone, wait on hold and call again at no cost to the customer. If agents start making routine calls on behalf of many customers, call volumes rise without any change in the number of problems. Staffing forecasts built on past human behaviour will be wrong.
Use case three: complaints written with AI
Customers are using general-purpose AI tools to write their complaints. The complaints are longer and cite more rules.
The Independent reported on 25 October that caseworkers in MPs' offices estimate roughly half their cases now involve AI in at least some of the drafting. One complaint to a council about bin collections ran to 27 pages. Lawyers in Local Government, which represents 5,600 council lawyers, estimated that handling the volume would take twice the current workforce. Pepe Di'Iasio of ASCL, the union for school and college leaders, said parents are using AI to produce several long complaints at once. Councillor Richard Clewer, who leads on public service reform for the Local Government Association, said complaints are getting "longer and more complex".
Wiltshire Council received 783 formal complaints in 2025-26, up 27 per cent on the year before. Its annual complaints report, covered on 29 September, links part of the rise to residents using AI tools. Croydon has a policy for complex AI-assisted complaints. It may ask the resident to resubmit, or write a short statement of the complaint for the resident to approve before the formal investigation clock starts.
Health is seeing the same thing. The Medical Protection Society surveyed almost 160 GP practice managers and reported on 12 October that 45 per cent said AI-drafted complaints were adding to workload pressure. Only 7 per cent had never received one.
The Financial Ombudsman picks up the phone
The Financial Ombudsman Service handles complaints that UK banks, insurers and lenders have not resolved. Its chief ombudsman, James Dipple-Johnstone, told the Financial Times that the worst case he had seen was a submission in which 37 of 39 references were wrong. The service is now calling more consumers to discuss their complaints, "to make sure that we get to the heart of their dispute".
In a small sample of cases from March, up to a third of responses to the ombudsman's initial assessments appeared to have been written or heavily assisted by AI. The service's own evidence to an FCA review puts the figure at around 35 per cent. Claims management companies and legal representatives are also using the tools to write submissions for consumers. In the most extreme examples these have run to more than 200 pages, often replying to a provisional decision of six to eight pages.
The service says AI also helps some consumers, including vulnerable ones, to organise a complaint and get past language barriers. It says it uses AI tools to process cases and will not use them to replace the judgement of its own staff. It received about 57,000 cases in the second quarter of its 2026-27 year, according to one report.
Where AI-written complaints break
The first cost is checking. Every rule or past decision a complaint cites has to be looked up. When most of them are wrong, the caseworker spends the time anyway. That cost falls on the company or public body receiving the complaint.
The second is that length hides the point. A 20-page complaint may contain one real grievance. The ombudsman's answer is a phone call, which costs staff time but finds the grievance faster. Croydon's answer is a short summary the resident agrees. Both put a person back between the AI text and the formal process.
The third is the clock. Complaint rules set deadlines. In financial services, the FCA's rules give a firm eight weeks to send a final response. A longer complaint does not change the deadline. A firm that wants to restate a complaint before investigating needs to check that its rules allow it, and record the customer's agreement.
The fourth is the temptation to answer AI with AI. Several suppliers sell tools that summarise long complaints. A summary is useful for triage. It can also drop the one point that matters, and the firm answers for its response whatever produced it.
What it costs and who pays
Contact centre AI is moving away from per-seat prices. Zendesk launched AI agents in May priced on resolutions, at $1.50 per automated resolution on a commitment or $2.00 pay-as-you-go, according to a comparison guide. Microsoft charges for Contact Center Intelligence in Copilot Credits by activity. RingCentral did not publish prices for this week's features.
A price per resolution needs a definition of a resolution. The buyer should get it in writing, with how the supplier counts a customer who gives up, rings back or complains later. A price per activity needs a forecast of activity. Agent-to-agent traffic and long AI-written complaints both add activity that last year's volumes do not show.
Some costs sit outside the AI line. Someone writes and maintains the quality rubric for AI agents. Someone reviews routing changes. Someone handles the calls an AI agent passes on, and the complaints an AI tool writes. In SAASiQ's view, a business case for customer service AI that counts the licence and the deflected calls, but not those roles, will understate the cost.
A framework for customer service teams
The steps below apply whichever suppliers a team uses.
Start with an inventory. List every AI agent that talks to customers, every AI tool staff use during a conversation, and every place an AI summary or score feeds a decision. For each, record the supplier, what it can change and who reviews its output.
Score AI agents on their own failures. Add lines to the quality rubric for invented policies, loops, wrong transfers and failure to tell the customer they are dealing with AI. Review a sample of AI conversations every week, as you would for a new member of staff.
Make the way out easy. Customers should be able to reach a person without fighting the AI agent. Test it yourself from a customer's phone, every quarter. Ofcom has said it expects this of telecoms providers.
Decide how you treat a customer's agent. Write down what you will accept from an AI agent acting for a customer, on the phone, in chat and through a login. Decide when you will ask the customer to confirm. Brief the contact centre and the fraud team on the same rule.
Plan for longer complaints. Agree how staff handle a complaint that runs to many pages or cites rules that do not exist. A phone call or an agreed summary works, as long as the customer's agreement is recorded and the regulatory deadline is kept.
Treat AI-made configuration as a change. Routing and flows built from a plain-language request still need an owner, a test record and a rollback route. Keep the old configuration until the new one has run through a busy day.
Get the price model in writing. Find out whether you pay per seat, per conversation, per resolution or per credit. Ask how a resolution is defined. Ask what happens to the bill if volumes double because customers' agents start calling.
What happens next
Microsoft's Contact Center Intelligence remains in public preview, with a Service Operations agent still to come. The Personal Agent Protocol is at draft 0.1, and Zendesk's Open Service Directory has no date for general availability.
SAASiQ - Intelligent Solutions for SaaS ©
