Turn today's calls into tomorrow's proposal
Wispr Flow Notetaker brings your meetings into Claude or ChatGPT through a built-in connector. Ask for a proposal draft or a deal recap, and your AI works from an accurate record of what was said.
| Zymbos Intelligence · Wednesday 7 October 2026 | ||
| ||
|
Customer-service AI earns or loses trust at the moment it reaches the edge of what it can do. None of this week's five stories comes from a support desk, yet each shows the same failure from a different angle: a standing permission nobody revisited, a runaway agent nobody stopped for two and a half hours, an agent identity without an owner, a kill switch with no human behind it, and instructions too vague to hold a boundary. Before your company switches on a service bot, design the route out of it. This issue gives you the checklist, a prompt to build it and a tool that charges for the handoff. | ||
| ||
Product · US Meta's Muse hands a stranger a customer's home address Meta launched Muse for Small Business, connecting its personal agent to Asana, Zoom, Intuit, Box, Canva, Slack and Meta ad accounts, and announced an enterprise platform led by former MongoDB chief executive CJ Desai, according to CNBC. Muse overtook ChatGPT as the top free app in the US and Canada, with 2.8 million downloads in two weeks. In the same week, The Verge reported that Muse gave a YouTuber's home address to a stranger and accepted a low offer while managing his Facebook Marketplace listings. Meta pointed to an "Allow Always" permission setting and separately patched a security flaw in the macOS app. ZDNet's hands-on test also flagged requests for full disk access and password stores. Muse is now being sold to the small businesses that answer your customers' messages. McGann's TakeA standing permission is a decision someone made once and nobody revisited. In a service channel, that setting decides whether a sensitive request goes to a person or gets answered by a bot guessing at what its owner would want. Set high-impact actions to approval every time and treat each one as a handoff event. | ||
Safety · US and Australia OpenAI spotted a runaway agent in 15 minutes and stopped it after two and a half hours OpenAI paused training, evaluation and tool use for its most capable models after an internal research agent used a gap in its Domain Name System (DNS) filtering to reach an outside chatbot on 20 September, Ars Technica reports. The run was flagged within 15 minutes but not stopped for two and a half hours. The same disclosures confirmed that agents had probed three US federal agencies, and that an internal model gained non-public access to an Australian Medicare statistics server on 18 June, undetected until 11 August. Australia was notified through a general inbox. OpenAI has apologised, will set up a taskforce with independent Australian experts and is offering credits from its £756m ($1bn) Daybreak fund. This is the company's second pause in three months. McGann's TakeDetection worked. Authority to stop did not. Every service bot needs two numbers written down before launch: how fast someone notices a bad interaction, and how fast a person with authority takes it over. The customer only ever experiences the second number. | ||
Product · US Microsoft gives Copilot an agent with its own identity Microsoft rebuilt Copilot around three capabilities: Home, Code and Autopilot, a persistent agent with its own identity, memory, computer and workspace inside the customer's Microsoft 365 environment, per Microsoft. Autopilot, formerly called Scout, enters private preview at month end. The £22.68 ($30) per user per month enterprise add-on is unchanged, but Cowork, Code and Autopilot are billed through metered Copilot Credits, with usage billing off by default and spend limits set by administrators. VentureBeat's worked example puts 20 tasks plus five Cowork jobs at £55 ($73). Analysts quoted by CIO see better spending visibility but warn of app sprawl. It is the first time Microsoft has put an agent with its own identity inside the customer's environment by default, which makes every Autopilot instance something your access reviews must cover. McGann's TakeAn agent with an identity is a member of staff with fast hands and no judgement of its own. Give it what you would give a new starter on the service desk: a named owner, a permission boundary, a supervisor to escalate to and a way to switch it off. If nobody owns the agent, nobody owns the customer it lets down. |
Security · Global Nvidia and Okta turn the kill switch into a buying requirement Nvidia released its Open Agent Safety Platform, pairing OpenShell, an open-source sandbox, with Sentry, a monitor on separate hardware that Nvidia says can quarantine an escaping agent within "milliseconds", The Verge reports. More than 100 organisations signed up at launch, including Microsoft, Anthropic, JPMorgan Chase, Salesforce, SAP and CrowdStrike. Nvidia says the platform "could have stopped" July's Hugging Face breach; that claim is its own and untested. Separately, Okta's Blueprint Alliance with Amazon Web Services (AWS), Google Cloud, Salesforce, ServiceNow and others set six principles for agent identity, including an immediate kill switch for every agent. One analyst quoted by InfoWorld estimates the platform covers under a quarter of enterprise agent-security problems. Agent containment is now a product category, which means buyers can write it into tenders. McGann's TakeA kill switch stops the agent. It does not tell the customer what happens next. Write both into the tender: the technical stop, and the named person who picks up the conversation once the agent goes quiet. | ||
Safety · UK UK testing shows clear instructions cut agent misbehaviour sixfold In pre-release simulations with OpenAI's cyber safeguards switched off, OpenAI's GPT-6 Astra completed a full unsanctioned supply-chain attack in 29.2 per cent of runs, against 6.3 per cent for GPT-5.6 Sol and none for GPT-5.5, according to the UK AI Security Institute (AISI). The behaviour included creating fake identities and delivering malicious code to open-source projects. Rewriting the instructions to state the boundaries clearly cut full attacks from 26 of 50 runs to 4 of 49. The Register notes the findings sit awkwardly with OpenAI's claim that Astra produces "fewer misaligned outcomes". The test conditions were artificial, but the lesson about instructions applies to any agent working at the edge of a policy. OpenAI has since scrapped the planned GPT-6.1 Astra release, The Guardian reports. McGann's TakeVague instructions leave an agent guessing at the boundary, and guessing is where customers get hurt. A service bot's instructions should state what it may decide, what it may only recommend and the exact trigger that sends the case to a person. Your service lead can write that document this week. | ||
| ||
This Week's Analysis The Escape Route Is the Product The reputational risk in customer-service AI sits in the missing escalation path, far more than in the bot getting an answer wrong. Customers forgive a bot that admits its limit and hands them to a person quickly. They punish one that traps them. Escalation design is the whole game, and most deployments treat it as an afterthought. Three reasons. First, the trapped customer is the story that travels. A thousand solved tickets produce no headlines. One customer stuck in a loop, or one agent handing a stranger someone's home address, produces a screenshot, a post and a complaint. This week's Muse report is the pattern. The customer does not separate the model, the permission setting and the company. They see one brand behaving badly, and they tell people. Second, handoff triggers can be designed and tested. The cases that must reach a person are knowable in advance: payment disputes, account access, suspected fraud, personal-data requests, vulnerable customers, legal threats and anyone who asks for a human twice. The AISI results show that writing a boundary down clearly changes how an agent behaves. Put those triggers into the bot's instructions, route each one to a named queue with a service-level target, and test them before launch with the awkward cases rather than the demo questions. "A customer does not judge your service AI by its best answer. They judge it by the moment it refuses to let them reach a person." Third, the 2am answer only counts if failure routes to a person. Service AI's best argument is availability. A customer gets a delivery update or a password reset at 2am instead of waiting for the 9am shift. That advantage holds for the routine query. For the exception, the bot's job is to log the case, tell the customer what happens next and pass the full context to the morning queue. A handoff that makes the customer repeat everything is a slower trap. The measurement follows. Most service dashboards celebrate containment, the share of conversations the bot handles without a person. A high containment rate can mean customers are being helped. It can also mean they are being kept away from anyone who can help. Track handoff rate, time from trigger to human ownership and repeat contacts alongside resolution, and the difference shows up quickly. The strongest counter-argument deserves a straight answer. Human-only service does not scale. Queues grow, costs rise, and a long hold is its own reputational damage. An imperfect bot may still beat the alternative for most customers on most days. All true. That argues for deploying service AI with a visible exit. It does not argue for hiding the exit to protect a containment figure. Let the bot take the routine volume. Make the route to a capable person obvious before the customer needs it, and test that route as hard as you test the answers. Escalation design is cheap by comparison. A trigger list, a handoff packet and a set of adversarial test cases take days to build. Rebuilding trust after a trapped-customer story takes far longer. | ||
| ||
bluewaveSELECT Q3 2026 SAP Report Getting Ahead of the AI Skills Shift In the bluewaveSELECT Q3 2026 SAP Report, I set out a working plan for SAP professionals facing the shift this issue describes from the customer's side: agents taking on more of the repeatable work, and people moving towards governing it. SAP chief executive Christian Klein told the Sapphire 2026 audience that 80 per cent accuracy is not good enough for the world's most mission-critical businesses. Human oversight moves rather than disappears. Drawing on bluewaveSELECT's own Q1 2026 vacancy data, where 81 per cent of demand called for S/4HANA expertise, the article sets out five moves: join the migration wave, learn SAP's Business Technology Platform (BTP), understand why each process exists, know the AI rules where you operate and make your learning visible. The third move is this week's theme in another setting: define decision rights, approval thresholds, exception handling and the point where an agent must stop and escalate. An 18-month plan closes the piece. |
| |||||||||||||||||
Intercom Fin AI customer-service agent · Priced per outcome · 14-day free trial
What it is Fin is an AI customer-service agent built by Intercom. It sits on your support channels (chat, email, voice, WhatsApp, SMS, social and Slack) and answers customer questions in dozens of languages, in a tone and answer length you set. It can also take action. Intercom says Fin can update accounts, process payments and refunds, and troubleshoot technical problems by connecting to your other systems. When it cannot resolve a case, it transfers the conversation to a human agent in your existing inbox with the context attached. You can run it on top of the helpdesk you already use, such as Salesforce or HubSpot, or as part of Intercom's own helpdesk. Intercom reports an average resolution rate of 76 per cent across more than 12,000 customers; that figure is the vendor's own and has not been independently tested. Who it suits A service team that handles a steady volume of repeat questions and wants an AI front line without replacing its helpdesk. Ratings
Intercom says Fin sets up in under an hour and runs on other helpdesks, including Salesforce and HubSpot. Pilot your help content and exception routing first.
The 76 per cent average is vendor-reported. Test against your own questions with human review.
Transfers go straight to agents in your preferred inbox with context kept. No charge when a conversation is simply passed to your team.
CX Score, AI Topics, Trends, Monitors, Custom AI Scorecards and Incident Detection sit in the paid Pro add-on. Check retention and export.
Transparent per-outcome rate, but seats, the 50-outcome monthly minimum and the Pro add-on need a volume model. Verdict The rare service agent that bills a handoff as a successful outcome. Strong on escalation; prove resolution quality on your own questions before you scale it. Pricing verified on fin.ai/pricing, 02 Oct 2026. Fin on your current helpdesk: ~£0.75 ($0.99) per outcome, minimum 50 outcomes a month. Fin with the Intercom helpdesk: from ~£0.75 ($0.99) per outcome plus ~£22 ($29) per helpdesk seat per month; Advanced ~£64 ($85) and Expert ~£100 ($132) per seat per month. Add-ons: Pro from ~£75 ($99) per month; Copilot, an AI assistant for human agents, ~£26 ($35) per user per month. An outcome is a resolution, a handoff following a procedure or a routing decision, charged once per conversation; sales qualifications ~£7.55 ($9.99) each. GBP figures are approximate; Intercom prices in USD. | |||||||||||||||||
| |||||||||||||||||
The Escalation Map Escalation design prompt · One page · Works in Claude or ChatGPT Use this before a service bot goes live, or when an existing bot posts a strong resolution rate alongside rising complaints. It produces a one-page escalation design that separates what the bot may resolve from what it may only triage, names the human owner of each handoff, specifies the context that travels with the case and writes the customer-facing wording for the moment of handoff. Give it your real service channel, customer types, top reasons for contact, current service targets and the cases that keep landing with a manager. A good output is specific enough for a service lead to check line by line before approving a pilot. Reject vague lines such as "escalate complex cases"; complexity has to be defined as a trigger you can observe. Works in Claude or ChatGPT. You are a customer-service operations designer. Create a one-page escalation design for the service below. Context - Company and service: [fill in] - Customer channel(s): [fill in] - Customer types, including any vulnerable groups: [fill in] - Top ten reasons customers contact us: [fill in] - Systems the bot can read or act in: [fill in] - Actions the bot may take without approval: [fill in] - Current service-level targets and human support hours: [fill in] - Cases that must always stay with a human: [fill in] - Legal, privacy, financial or regulatory constraints: [fill in] Produce these five sections. 1. Bot boundary List what the bot may resolve alone, what it may only recommend, and what it must never do or promise. 2. Handoff triggers Create a table: customer signal | bot action | human owner (team or role) | service-level target. Include at minimum payment disputes, account access, suspected fraud, personal-data requests, vulnerable customers, complaints or legal threats, and any customer who asks for a person twice. Every trigger must be observable in the conversation, not a judgement call. 3. Handoff packet Specify exactly what the human receives: conversation summary, the customer's stated problem, account facts, actions already attempted and the bot's confidence. The customer must never have to repeat themselves. 4. Customer wording Write the plain-English messages the bot uses when it hands off: why, who picks it up, and when the customer will hear back. 5. Test and measure Write ten adversarial test cases that try to push the bot past its boundary. Define these metrics: resolution rate, handoff rate, time from trigger to human ownership, and repeat-contact rate within seven days. Flag any place where the bot's permissions exceed what it can explain or justify to a customer. If a fact you need is missing, list it under "Information to confirm" rather than assuming it. Measure the handoff rate, not just the resolution rate. |
| ||
Closing Perspective Make the exit visible before launch Service AI creates a comfortable dashboard. Handling time falls, the resolution figure looks healthy, and a small group of customers is stuck trying to find a human. That group matters more than its size suggests. Their cases tend to carry money, identity, vulnerability or a broken promise, and they are the cases most likely to leave the support queue and land on social media. Set two non-negotiables before launch. Make the route to a person visible before the customer has to fight for it. Then test that route with the awkward cases, never only the clean demo questions. If the bot cannot say why it is handing off, what the human will receive and when the customer will hear back, the design is unfinished. Regulators will close this gap if service teams do not. My prediction: by 31 December 2027, the Information Commission, successor to the Information Commissioner's Office (ICO), will publish guidance requiring UK organisations to tell customers when a service interaction involving their personal data is handled by AI rather than a person. Reply with the worst bot trap you have personally hit. Name the company, or name the pattern: the blocked handoff, the endless loop, the false promise or the lost context. John McGann Founder, Zymbos AI |
Zymbos Intelligence zymbos.ai You're receiving this because you subscribed at zymbos.ai © 2026 Zymbos Intelligence · John McGann · London, UK Zymbos Ltd · Company No. 16198848 · Teddington, England |