Head to head: muse-spark-1.3 vs Qwen: Qwen3.8 Max Prime

muse-spark-1.3: # The Most Expensive AI Didn't Win Customer support is where AI hype meets angry customers. It's also where cost matters as much as quality, because every ticket answered by a model is a ticket you pay for by the thousand. A new benchmark of five language models on real support work delivers a clear, practical lesson: you don't need to pay top dollar to get top-tier answers. The highest-priced model didn't pull ahead — it tied. And the runner-up did the same work for less than half the price, and faster. That finding should change how support leaders shop for models, shifting the question from "which is smartest?" to "which is smart enough at a sustainable price?" ## How the test was run The researchers kept it simple and realistic. They pulled 200 anonymized customer-support tickets from a software company and asked each of five models — dubbed North, Vale, Ember, Kite, and Harbor to keep the evaluation neutral — to draft one proposed answer per ticket. Three experienced support managers then graded those answers blindly, without knowing which model wrote what. Scores combined three things managers actually care about: factual correctness, usefulness for resolving the issue, and tone with the customer. This isn't a trivia quiz or a coding challenge. It's a judgment call by people who do this work, which makes the results more relevant to deployment — but also more subjective than an automated test. With only 200 tickets, small gaps in average score shouldn't be over-interpreted. ## A tie at the top, a gulf in price The headline result was a virtual dead heat. North averaged 84.2 out of 100, Vale averaged 83.8. The researchers report that difference was not statistically significant. In plain terms: on this test, they performed the same. The costs did not perform the same. North cost an estimated $1.90 per 1,000 tickets, Vale just $0.74. Vale was also 18% faster to respond. For a support operation handling tens or hundreds of thousands of tickets, that math compounds quickly. If quality is indistinguishable and one option is 61% cheaper and noticeably quicker, the cheaper, faster option isn't a compromise — it's the winner on value. Speed matters here too, not just for wait times but for agent-assist workflows where a human is waiting on a draft. The two leaders also had different strengths. North was strongest on billing disputes, where precise policy language and careful handling of sensitive charges count. Vale was strongest on technical troubleshooting, where structured, step-by-step reasoning helps. That split suggests there isn't a single "best support model," even at the top, but rather a best fit for your ticket mix. ## The rest of the field: cheaper, but weaker Below the top two, price kept falling — and so did quality. Ember and Kite landed in the high 70s, with average scores around 79 and 76. Their estimated costs were dramatically lower, at $0.31 and $0.22 per 1,000 tickets. Harbor was cheapest of all at just $0.08 per 1,000 tickets, but it also scored lowest at 71.0. A 13-point gap between first and last is meaningful in this kind of human grading. It’s the difference between an answer a manager would send with a light edit and one that gets rewritten or, worse, sends a customer down the wrong path. A 5-point gap between the top tier and the middle tier is also worth taking seriously, even if we shouldn't fuss over fractions of a point within tiers. For teams, this creates three real choices, not five: premium quality at a mid-range price with Vale, premium quality at a premium price with North, or a deliberate trade-down to a budget model if your workflow includes strong human review. What the data does *not* support is paying more for North on the assumption that highest price equals highest quality. It didn't. ## A shared blind spot All five models stumbled in the same place: tickets that depended on an image attachment. That's an important reality check. Real support inboxes are full of screenshots, error photos, and annotated PDFs. A text-only model, no matter how eloquent, can't reliably resolve "see attached" without vision capabilities or proper tooling to retrieve the file. Organizations that route a lot of visual tickets to AI without solving that input problem will see performance drop regardless of which of these models they pick. ## What this test can't tell you Context limits these results. All tickets were in English and came from one software company. A model that excels at SaaS troubleshooting prose may not excel at retail returns, medical devices, or highly regulated financial support. Tone expectations, jargon, and correctness risks differ by domain. The study also tested only single-turn drafted answers. It didn't test agentic actions, tool use, or multilingual support — in other words, it didn't test whether a model can look up an order, issue a refund, escalate properly, or work across languages. Those capabilities often determine success in production more than draft quality alone. Think of this as a strong signal on core answer quality and cost-efficiency, not a final buying decision. You'd still want to pilot on your own tickets, measure resolution and escalation rates, and test with your guardrails and knowledge base connected. For now, though, the takeaway is refreshingly practical: the smart money isn't on the most expensive model. It's on the one that ties for first and costs like it's third.

Qwen: Qwen3.8 Max Prime: # Vale Matches North at Lower Cost in Support-Ticket Benchmark The clearest finding from a new benchmark of five language models is not that one model dominated, but that two front-runners were effectively tied. Model North scored 84.2 on average, while Model Vale scored 83.8 across 200 anonymized customer-support tickets. The researchers found that difference was not statistically significant. In practical terms, Vale delivered comparable answer quality at much lower estimated cost and higher speed, making it the more compelling default for many support teams. The benchmark was designed to resemble a realistic drafting task. Researchers selected 200 anonymized tickets from one software company and asked each model to produce one proposed answer per ticket. Three experienced support managers graded the responses blindly. Scores combined factual correctness, usefulness, and tone — a sensible blend for customer support, where a technically accurate reply can still fail if it is unclear or dismissive. That method matters. Blind grading reduces the chance that brand expectations influence scores, while multiple managers help balance individual preferences. Still, the benchmark measures proposed answers, not live customer outcomes. A response can sound excellent to reviewers yet fail if it relies on outdated policy details, misreads intent, or cannot be used inside a company’s support workflow. North and Vale separated themselves from the rest of the field. North’s average was slightly higher, but not meaningfully so. Vale answered 18% faster than North and carried a far lower estimated cost: $0.74 per 1,000 tickets, compared with $1.90 for North. If two systems produce indistinguishable overall quality, those operational advantages become important. In support environments, latency and cost affect whether an AI assistant can be deployed at scale, embedded in real-time workflows, or used for high-volume triage. The models did show different strengths. North performed best on billing disputes, where tone, policy sensitivity, and careful explanation may matter more. Vale performed best on technical troubleshooting, suggesting it was better at following problem descriptions, identifying likely causes, or offering actionable steps. The “best” model may therefore depend on ticket mix. A company dominated by billing issues might still favor North, while one facing more technical questions may find Vale especially attractive. Below the top pair, performance declined noticeably. Ember scored 79.1, Kite 76.4, and Harbor 71.0. These are not trivial gaps. A support assistant scoring in the low 70s may produce plausible drafts, but human reviewers would likely need to intervene more often. The cheapest models may also create hidden costs if answers require heavy editing or confuse customers. The benchmark also exposed a shared weakness: all five models performed poorly when a ticket depended on an image attachment. This is an important reminder that customer support is increasingly multimodal. A user may attach a screenshot of an error message, a photo of hardware, or an invoice image. If a model cannot reliably interpret that context, its answer may miss the point. Teams evaluating support AI should test image understanding separately rather than assume it works. The study’s scope is also narrow. The tickets were English-language submissions from one software company, so the results may not transfer to other industries, customer groups, or languages. The researchers did not test agentic actions, tool use, or multilingual support. A model that writes good responses may still struggle to look up account details, check order status, follow escalation rules, or operate safely with external tools. Those limitations do not invalidate the findings; they define where the findings are most relevant. For English-language, text-based support drafting in a software company context, Vale appears to offer the strongest balance of quality, cost, and speed. North remains strong, especially for billing disputes, but its higher cost and slower speed are harder to justify given the statistical tie. The practical takeaway is not that Vale is universally superior. It is that buyers should be wary of paying a premium for a marginal score difference. In support automation, the most useful model is often the one that is good enough, affordable enough, and fast enough to be deployed broadly under human oversight. On that test, Vale makes the stronger case.

muse-spark-1.3: # Arcwell Launches Relay, an AI Assistant for Support Agents Support agents rarely lack information. They lack time to find it. Anyone who has sat on hold while a rep checks another system has felt the delay from the other side. That is the gap Arcwell wants to close with Relay, a new platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. Rather than answering customers directly, Relay is built to assist the person in the middle. As agents handle inquiries, it draws on prior transcripts and company materials to propose responses. It can also identify customers who may cancel and create summaries, work that otherwise eats into time between tickets. CEO Priya Shah said, “Support agents spend too much time searching for information that their companies already have.” Relay was tested by 12 companies over six months. Arcwell says agents using it closed tickets 14% faster. The company did not release customer-retention results, leaving open how often its warnings about at-risk customers translated into saves. Relay costs $65 per agent each month, a per-seat price that ties its value directly to agent productivity. The approach depends on deep access. Relay requires access to support transcripts and internal help documents, the same repositories that make its suggestions useful and that tend to give security reviewers pause. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period, allowing them to align storage with internal policies. For Arcwell, the release is a bid to turn AI interest in customer service into everyday use. The outcome will hinge less on whether the technology can locate an answer than on whether agents trust the suggestions enough to use them consistently, and whether managers see faster resolutions without a loss of care in the reply. If agents have to double-check every suggestion, the time savings could prove elusive.

Qwen: Qwen3.8 Max Prime: # Arcwell Announces Relay, an AI Platform for Customer Support Agents Arcwell has announced Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. Arcwell described the announcement as a significant development for the company and the broader customer service industry, as businesses look for ways to use AI to make support work faster and more consistent. Relay’s importance rests on three capabilities: suggesting answers to agents while they handle customer inquiries, identifying customers who may cancel, and creating summaries of support conversations. Those features demonstrate how AI continues to transform the customer service landscape, moving beyond simple automation and into tools that shape how agents work and how companies understand their customers. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Arcwell tested Relay with 12 companies over six months. The company said agents using the platform closed tickets 14% faster. It did not release customer-retention results, leaving unanswered whether the system’s cancellation warnings will help companies keep more customers. Relay costs $65 per agent each month, a price point that makes it an ongoing software expense for support teams. The platform also presents potential challenges. To work, Relay requires access to support transcripts and internal help documents. That requirement places sensitive customer interactions and internal knowledge at the center of the product. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period, giving them some control over how long information remains in the system. For Arcwell, Relay represents an important new chapter as the company seeks to help businesses improve customer service using AI. Its success will depend on how well the platform performs in real support environments, how customers respond to its data policies and whether the promised speed gains prove durable. The launch also shows that AI’s role in support is being judged not only on speed, but on data governance and evidence.

muse-spark-1.3: # HelioDock Launches Night Shift for Remote Supervision of Warehouse Robots *HelioDock pairs remote fleet software with $24 million round as beta tests point to fewer stoppages* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The release is aimed at the overnight hours when a single stalled robot can require a manager to travel to the facility. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” CEO Mara Velez said. Night Shift allows oversight from outside the warehouse, so response does not depend on driving to the building. HelioDock announced the software alongside a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. The timing places the software debut in the same news cycle as the financing. North Coast Ventures led the $24 million round, according to the company. With that round, HelioDock says its total funding has reached $39 million. The launch follows a 14-month private beta involving nine logistics companies and 310 robots. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages. That figure has not been independently verified. The test period put Night Shift in private beta operations over more than a year, spreading use across different companies and a combined fleet in the hundreds. HelioDock presents the 17% reduction as the outcome observed across that beta group. HelioDock has not provided an independent audit of the 17% figure. As presented by the company, the reduction describes the experience of beta customers using Night Shift during the test period. Unplanned stoppages are the interruptions Night Shift is meant to address without requiring an on-site visit. A lower rate, if maintained outside the beta, would mean fewer of the overnight trips Velez described. Whether that rate holds for other operators remains to be seen in general availability. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. That dependence drew a distinction in the beta between well-connected sites and rural locations. Pike said performance held up where connectivity was strong. Where LTE service dropped, the ability to monitor from afar dropped with it. Pike’s comment separates two operating conditions encountered in the beta. Facilities with strong connectivity allowed the intended remote workflow. Rural facilities subject to LTE outages did not support the same level of remote use, reducing the practical benefit during those periods. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable. For a remote operator, that fallback defines the boundary of the product. Remote control is available subject to connectivity. When connectivity fails, the robot continues under its locally configured behavior. Night Shift will cost $90 per robot per month and become generally available September 15. The pricing attaches a monthly charge to each machine under remote supervision. At $90 per robot per month, cost scales directly with the number of machines placed under Night Shift. An operator with a larger fleet would pay more in total each month than an operator with a smaller fleet at the same unit rate. General availability is set for September 15. HelioDock is entering competition with fleet-management products from Relay Robotics and YardOS. Its launch pairs availability of the software with the new Series B capital. Relay Robotics and YardOS already offer fleet-management products, according to the available information. Night Shift places HelioDock in direct comparison with those offerings for logistics companies evaluating remote supervision.

Qwen: Qwen3.8 Max Prime: # HelioDock Launches Night Shift Remote Monitoring Software for Warehouse Robots *The software, priced at $90 per robot per month, follows a 14-month beta and a $24 million Series B led by North Coast Ventures.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. Night Shift is aimed at the period when warehouses rely on robots but may have fewer managers on site. HelioDock said beta customers experienced 17% fewer unplanned robot stoppages. The company did not provide independent verification of that figure. Night Shift becomes generally available Sept. 15 and will cost $90 per robot per month. The software gives operators a remote view of robot fleets. HelioDock’s robots can encounter interruptions that require human attention, such as a robot being unable to proceed around a pallet. CEO Mara Velez said the software addresses that kind of overnight exception. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. HelioDock spent 14 months testing Night Shift in a private beta with nine logistics companies. The beta included 310 robots. During the test period, HelioDock measured a 17% reduction in unplanned stoppages among beta customers, according to the company. The figure has not been independently verified. Unplanned stoppages can require a person to respond to a robot that has stopped working as expected. Remote monitoring lets an operator see that stoppage and respond without traveling to the facility. HelioDock’s description of Night Shift centers on that remote fleet oversight. Beta customer Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity. Pike said Night Shift became less useful during rural LTE outages. When connectivity fails, HelioDock’s robots return to their locally configured behavior. Remote control becomes unavailable until the connection is restored. That dependence on connectivity means Night Shift’s remote functions are not continuous during network failures. In a facility with a stable connection, the software provides the remote monitoring layer HelioDock describes. In a rural location with intermittent LTE service, the remote portion of the system can be interrupted. Night Shift does not replace the robots’ local configurations. The software operates alongside them. If a connection drops, the robots continue with the behavior already configured at the site, while the remote operator loses control. HelioDock makes autonomous warehouse robots, and Night Shift is the software service that monitors those fleets from offsite. The subscription is tied to each robot. At $90 per robot per month, an operator’s software cost rises as more robots are added to the monitored fleet. Night Shift’s remote-monitoring role applies to fleets of HelioDock’s autonomous warehouse robots. Operators use the software to watch the fleet rather than stand beside each machine. The service is priced by robot, not by facility, and becomes generally available after the private beta on Sept. 15. HelioDock’s 17% claim is tied to unplanned robot stoppages. The company supplied that percentage from the beta test; the figure has not been independently verified. The metric concerns unplanned stoppages. Relay Robotics and YardOS offer competing fleet-management products. HelioDock’s launch places Night Shift in that category, with pricing attached to each robot and availability beginning Sept. 15. The $24 million Series B was announced with the product launch. North Coast Ventures led the round, bringing total funding to $39 million.

muse-spark-1.3: # Startup Spotlight: LedgerLark For many small investment firms, the work after a trade is agreed upon remains manual and time-consuming. Confirmations arrive in one format, custodian records in another, and the internal ledger in a third. When the numbers do not line up, someone has to track down why. LedgerLark is building AI agents focused on that post-trade operations work for small investment firms. The idea is not to make investment decisions or to speed up trading itself, but to reduce the manual effort of checking records and clearing exceptions. Co-founder Mina Batra spent seven years in operations at a regional brokerage. Co-founder Owen Carr built reconciliation systems at a large bank. Their backgrounds are in the back-office systems that record what was bought, sold, and settled, rather than in portfolio management. Batra began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. That kind of error is common in operations: a single wrong field can prevent settlement, and finding it often means searching across emails, PDFs, and spreadsheets. LedgerLark’s product compares three sources — trade confirmations, custodian records, and internal ledgers — and flags where they differ. When it finds a mismatch, it proposes a correction for a person to review and approve. The company emphasizes that the product does not execute trades or move money. A human still has to accept any proposed change before it is applied. That design reflects how operations teams work. Exceptions still need judgment, especially when the records disagree and the correct value is not obvious. LedgerLark is asking those teams to trust an agent to do the initial comparison and research, but not to act on its own. On traction, LedgerLark has five paying customers. Annual contracts range from $36,000 to $110,000, which suggests the company is working with relatively small teams with limited budgets for new tools. The company says one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the product. That figure is worth noting with context. It comes from the company, covers one customer, and does not by itself show how widely the result would apply. It also leaves open questions about what types of exceptions were resolved, how much staff time was actually saved, and whether the improvement persists as trade volumes change. The company has raised a $3.5 million seed round led by Motive Row. For a company selling to financial firms, that funding will likely need to cover not just product development but the cost of meeting security requirements and supporting longer evaluations. Competition is another factor. Established reconciliation vendors already serve this area, and many larger firms rely on internal bank software built over years. LedgerLark is focusing on smaller firms that may not have dedicated engineering resources, but those firms can still be careful buyers when a tool touches accounting records. Financial institutions may require lengthy security and procurement reviews, even for software that only reads data. That can slow sales cycles and make early revenue uneven. How LedgerLark handles onboarding, permissions, and audit trails will likely matter as much as how well its agents spot mismatches. What to watch is whether LedgerLark can turn early use into repeatable deployment. Can it connect to a new firm’s systems without extensive custom work? Do operations staff continue to accept its suggestions after the first few months? And does fewer open exceptions translate into measurable time saved or fewer settlement failures? For now, LedgerLark offers a narrow, practical use of AI agents in a part of finance that gets little attention. The next test is whether that focus can carry through procurement, integration, and daily use.

Qwen: Qwen3.8 Max Prime: # Startup Spotlight: LedgerLark wants to make post-trade cleanup less manual For many small investment firms, the trade is the easy part. The harder work often begins after execution: matching confirmations, checking custodian records, finding why a settlement failed, and deciding what to fix. LedgerLark, a startup focused on post-trade operations, is trying to take some of that routine burden off human teams by using AI agents to identify and propose corrections — while leaving final approval to people. The insight comes from frustration inside operations work. Co-founder Mina Batra spent seven years in operations at a regional brokerage, where she saw how small errors could consume disproportionate time. One episode stuck with her: a six-person team spent three days resolving a trade mismatch that traced back to a copied account number. That kind of problem is not dramatic, but it is expensive. It ties up staff, delays reconciliation, and can create follow-up work across counterparties and custodians. Batra co-founded LedgerLark with Owen Carr, who previously built reconciliation systems at a large bank. Their combined experience suggests a practical angle: instead of promising autonomous finance, the startup is focusing on a narrow set of workflows where machines can compare records quickly and humans still decide what happens next. The product uses AI agents to compare confirmations, custodian records and internal ledgers. When it finds discrepancies, it proposes corrections for human approval. The company emphasizes that the software does not execute trades or move money. That limitation may make the product easier to introduce into cautious environments, but it also means LedgerLark is selling an assistant for operations teams rather than a replacement for core trading or custody systems. Early traction is real but still limited. LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That is a useful data point, though it does not reveal how much manual review remained, how the change affected overall costs, or whether the result would repeat across different firms and data environments. The funding news is more substantial: LedgerLark has raised a $3.5 million seed round led by Motive Row. The capital should help the company deepen its product, support early customers and navigate the security reviews that often come with selling into financial services. Those reviews may be one of the startup’s biggest practical hurdles. Financial institutions tend to move carefully, especially when software touches records connected to client assets. Even if LedgerLark does not move money, procurement, information-security and compliance teams may still require lengthy assessments. For a young company, that can mean long sales cycles and heavy documentation work before revenue expands. Competition is another reason to stay measured. LedgerLark is not entering an empty field. Established reconciliation vendors already serve financial institutions, and many larger banks rely on internal software built specifically for their processes. Some potential customers may prefer to buy from familiar providers, while others may see value in a newer system designed around AI-assisted exception handling. LedgerLark will need to show that its agents can reduce workload enough to justify replacing or sitting alongside existing tools. The company’s approach looks sensible: start with a painful, repetitive task; keep humans in control; and avoid promising more than the product can do. If it can demonstrate consistent results across a broader set of customers, it may find a place in the unglamorous but essential machinery of financial operations.

muse-spark-1.3: # QueueZero Launches Proxy Meeting Assistant That Delegates Work to AI Agents **SAN FRANCISCO —** QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend a meeting, produce notes and assign follow-up work to software agents, entering a competitive market for enterprise productivity tools. The announcement was made during a live product demonstration attended by customers, press and analysts. The company said Proxy is designed to listen to meeting discussion, summarize decisions and identify next steps, then route those steps to connected software agents for execution. During the launch demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was transcribing the session, interpreted the remark as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first, before continuing with the presentation. The exchange drew a reaction from the audience and underscored the sensitivity of automated systems that act on conversational speech. QueueZero executives used the moment to describe the product’s safeguards. The company says Proxy requires confirmation before an external message is sent. That requirement is intended to prevent the assistant from contacting customers, vendors or other outside parties without human review. Internal tasks, by contrast, can be created automatically. The company said this behavior is intended to reduce administrative work by allowing action items to appear in task lists without additional input. Administrators can disable automatic task creation and require manual approval for all items generated by Proxy, according to QueueZero. The distinction between internal and external actions reflects a broader challenge for AI assistants deployed in workplaces. Systems that listen continuously must determine when conversation is dictation, discussion or a direct command. Misinterpretation can create unwanted work, confusion over responsibility, or unintended follow-up. For enterprise buyers, control settings are likely to be closely examined. The ability to disable automatic task creation gives administrators a way to limit autonomous behavior while testing accuracy in live meetings. Proxy costs $30 per user per month. It enters public beta next week, which will expand access beyond early demonstration users and allow the system to be tested in a wider range of meeting environments.

Qwen: Qwen3.8 Max Prime: # QueueZero Launches AI Meeting Assistant Proxy After Demo Assigns CEO Tasks QueueZero on Wednesday launched Proxy, an artificial intelligence assistant designed to attend meetings, take notes and assign follow-up work to software agents. The company says the product is meant to reduce the administrative drag that follows corporate calls, from summarizing discussion points to routing next steps into automated workflows. Proxy joins a rapidly expanding category of AI workplace tools that promise to turn meetings into structured records and actionable items. The pitch is especially appealing to organizations struggling with hybrid work, where calendar overload has made the prospect of skipping a meeting and receiving an automated summary more attractive than attending another video call. During a launch demonstration, the system showed how literally it can read a room. A customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy treated the remark as an instruction rather than a casual aside. It created two tasks assigned to QueueZero chief executive Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The moment became a practical example of the company’s broader point: Proxy is built to act quickly, but its maker is trying to define where automation ends and human approval begins. QueueZero said Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically, allowing the assistant to generate reminders, documents or assignments without waiting for a person to approve each step. Administrators can disable automatic task creation, a setting likely to matter for organizations that want tighter control over when tasks are generated. The episode also highlights the delicate challenge facing AI agents that move from summarizing conversations to executing them. A tool that can assign work based on spoken language may save time, but it also requires guardrails to prevent misunderstandings, unwanted messages or tasks that land in the wrong inbox. QueueZero’s answer appears to be a tiered system: external communication requires human sign-off, while internal follow-ups can proceed automatically unless the setting is changed. The pricing is aimed at business customers rather than casual users: Proxy will cost $30 per user per month. QueueZero plans to open a public beta next week, giving companies a chance to test how much meeting work they are comfortable delegating to an agent.

muse-spark-1.3: # The End of Foundry Hall **MILWAUKEE —** The coffee machine in the second-floor kitchen at Foundry Hall is still on a timer. It clicks awake at 7:30 each morning, as it has since 2018, when the nonprofit startup accelerator opened its doors to founders who had nowhere else to go. After Oct. 31, there will be no one left to pour it. Foundry Hall, which helped launch 212 startups that now collectively employ 1,140 people, announced this week it will close at the end of the month. The decision was not sudden, staff said, but slow, the kind that arrives in budget spreadsheets before it arrives in goodbyes. The immediate cause was the loss of its largest corporate sponsor. Blank River Bank, which supplied 38% of the accelerator’s annual budget, declined to renew its sponsorship this year. Without that anchor funding, leaders said the nonprofit can keep the lights on through October but cannot responsibly commit to another full program cycle. “We kept waiting for the replacement check that never came,” said Executive Director Camille Ross. The bank said the move was part of a broader shift in its community investment priorities, and not a judgment on Foundry Hall’s work. That distinction offers little comfort inside the building, where the calendar is now measured in final meetings and forwarding addresses. Six employees will lose their jobs. Twelve startups currently working out of its open-plan offices and meeting rooms must find new space by the deadline, packing up whiteboards and prototypes mid-growth. For many alumni, the news landed like the loss of a first home. Founder Dev Malik still remembers what came before. “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop.” City officials say they are trying to soften the landing. Deputy Mayor Susan Cole said the city has begun talking with building owners and local universities about temporary space for the displaced companies, though no agreements have been announced. Ross walks the halls now counting what remains: days, boxes, introductions she can still make. The accelerator was never just desks and demo days, she said. It was an attempt to prove that a good idea in Milwaukee deserved the same early faith as one on either coast. On Nov. 1, the timer on the coffee machine will click on to an empty kitchen. The startups will have scattered. What stays, for a while at least, will be quieter — 212 companies, more than a thousand jobs, and a habit of starting.

Qwen: Qwen3.8 Max Prime: # The Last Shift at Foundry Hall Foundry Hall, the nonprofit startup accelerator that opened in Milwaukee in 2018, will close on October 31. The date lands quietly, almost like an appointment no one wanted to keep. Since then, it has been less a building than a shared argument that new companies could take root here, find one another, and survive. It has made its case. Since opening, Foundry Hall has supported 212 startups, which collectively employ 1,140 people. Those numbers are impressive, but they understate the place. They do not capture the founder who wandered in with a laptop and a half-formed idea, the mentor who stayed late to review a pitch, or the small rituals that made risk feel less lonely. Dev Malik, one of the founders who passed through, remembers how different his company might have been without the space. He recalled that his first investor meeting happened inside the building, after a period of pitching from the back table of a coffee shop. That is the kind of transition Foundry Hall was built to create: a door from improvisation to credibility. The closing is not a mystery with a single villain. Blank River Bank, its largest corporate sponsor, supplied 38% of the annual budget and has chosen not to renew the sponsorship. The bank has said the decision reflects a shift in community investment priorities and is not connected to Foundry Hall’s performance. That may be true, but the effect is blunt: without that support, the organization has enough money to operate through October, not through another full program. Executive director Camille Ross put it plainly: “We kept waiting for the replacement check that never came.” The line carries the fatigue of people who spent months trying to close a gap while still showing up each morning to do the work. Six employees will lose their jobs. Twelve startups currently using the workspace must leave. Some will scatter to home offices, coffee shops, borrowed desks, or temporary corners of other organizations. Deputy mayor Susan Cole says the city is talking with building owners and local universities about space for the current companies, a useful promise, but not a replacement for what has already been built. The hardest part may be that Foundry Hall did not fail for lack of results. It proved what it could do, then lost the means to keep doing it. Milwaukee will still have entrepreneurs. The question now is where they will gather, who will take the first meeting with them, and what will be lost in the space between.

muse-spark-1.3: 6

Qwen: Qwen3.8 Max Prime: 6

muse-spark-1.3: C

Qwen: Qwen3.8 Max Prime: C

muse-spark-1.3: # Duplicate Reimbursements Totaling $2.8 Million Follow KiteBridge Update **Ninety-one corporate customers affected; CEO promises technical report after fallback failure** Thousands of employees received double expense reimbursements this week after a failed software update caused expense-management company KiteBridge to pay out $2.8 million it never should have sent, leaving employers to recover the money from their own accounts. The error began Tuesday morning when KiteBridge deployed version 4.2 of its platform at 8:45 a.m. The release was intended as a routine update to the system companies use to approve and pay employee expenses for travel, meals, and other work costs. Within half an hour, the first sign of trouble appeared. At 9:12 a.m., a customer reported unexpected payments, prompting engineers to investigate. Just over an hour after the deployment, at 10:03 a.m., the company suspended all automated payments to stop further duplicates. By then, the damage was extensive. Across 91 corporate customers, the system had generated 13,400 duplicate reimbursements. Employees who were owed a single payment for a submitted expense received it twice, with the extra money drawn not from workers' bank accounts but from their employers, who fund reimbursements through KiteBridge. The cause, according to the company, was a breakdown in how its software handles uncertainty. When KiteBridge sends a payment instruction, it waits for confirmation from an outside payment processor — the banking intermediary that actually moves the money. On Tuesday, those confirmations timed out, meaning KiteBridge did not know whether the payments had succeeded. Its fallback system, designed to retry when something goes wrong, mistakenly treated each retry as a brand-new reimbursement request rather than a second attempt at the same one. The result was a second, identical payment for thousands of transactions. That failure is particularly notable because KiteBridge had previously touted safeguards against precisely this type of error. An April changelog for version 4.0 said the company had introduced “exactly-once payment protection,” a term engineers use for systems designed to ensure a payment is processed one time and only one time, even if computers must retry after failures or network interruptions. The company’s public communication also lagged behind events. Its incident page initially described the problem only as delayed reimbursements, without mentioning extra payments. It was not until six hours after automated payments were suspended that the page was updated to disclose that duplicate payments had been made. CEO Lena Ortiz has since apologized for the incident and said KiteBridge will publish a full technical report explaining what went wrong. Recovery efforts are now underway. By Wednesday evening, the company said $2.3 million of the duplicate funds had been recovered, with the remainder still pending return from receiving accounts. No employees had money withdrawn from their personal accounts as part of the recovery, the company said, because employers had funded the duplicate payments in the first place.

Qwen: Qwen3.8 Max Prime: # Duplicate Reimbursements Cost Employers $2.8 Million After KiteBridge Update Fails A software failure at KiteBridge turned an ordinary expense-processing day into a costly accounting error, causing 13,400 duplicate reimbursements worth $2.8 million to be sent to employees across 91 corporate customers. By Wednesday evening, the company had recovered $2.3 million, while the remaining balance was still pending. KiteBridge said employers, not individual employees, funded the duplicate payments, and no workers had money withdrawn from their accounts. The trouble began Tuesday morning when KiteBridge, which makes expense-management software, deployed version 4.2 at 8:45 a.m. The update soon collided with a problem in the payment-processing chain. When a payment processor timed out, KiteBridge’s fallback system attempted to handle the interruption by retrying transactions. But instead of recognizing those retries as attempts to complete an existing reimbursement, the system treated them as new reimbursement requests. The result was a cascade of duplicated payments. The first customer complaint arrived at 9:12 a.m., less than half an hour after the release went live. Over the next hour, the company worked to identify the scope of the issue. At 10:03 a.m., KiteBridge suspended automated payments, halting further duplicate reimbursements. The episode also exposed a gap between KiteBridge’s earlier promises and what happened in production. In April, the company’s changelog said version 4.0 had introduced “exactly-once payment protection,” a safeguard meant to prevent the same payment from being issued more than once. The failure raised questions about how that protection performed when the payment processor timed out and the fallback process took over. KiteBridge’s public communication initially lagged behind the seriousness of the incident. The company’s incident page first described the issue as delayed reimbursements, language that did not tell customers that payments had been duplicated. Six hours after automated payments were suspended, the page was updated to disclose the duplicate reimbursements. For affected companies, the immediate challenge was reconciling expense reports and recovering funds that had been sent twice. KiteBridge said the duplicate payments were funded by employers, meaning employees did not lose money from their own accounts. The company recovered most of the $2.8 million by Wednesday evening, but some payments remained pending as finance teams worked through the cleanup. Chief Executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what went wrong. The report is likely to be closely read by customers seeking assurance that the expense-management platform can prevent a similar failure, particularly when outside payment systems experience timeouts. For now, the incident stands as a reminder of how quickly a routine software deployment can become a financial operations problem when retry logic fails. In payment systems, a retry is supposed to be a safety net. In this case, it became a mechanism for paying the same expense twice.

muse-spark-1.3: # Raincheck raises $8 million to help cities get ahead of floods *The Austin startup makes flood-risk software that helps local governments prioritize maintenance before storms* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The round was led by Basin Ventures, with participation from existing investor Long Mile. The new funding brings Raincheck's total raised to $10.5 million. Raincheck [maps drainage systems and combines those maps with rainfall forecasts](https://example.com/raincheck-methodology) to help cities decide where to send maintenance crews before a storm. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. ## The competition Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Qwen: Qwen3.8 Max Prime: # Raincheck Raises $8 Million to Help Cities Prepare for Flooding *The Austin company uses drainage maps and rainfall forecasts to show local governments where maintenance crews may be needed before a storm hits.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm arrives. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million in total. Raincheck CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” More information about the company’s approach is available in [Raincheck’s methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations, and emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck said the new funding will help it hire eight engineers and expand into Florida.

muse-spark-1.3: # Cloudnote Left Prompt Excerpts Exposed for Four Hours, Affecting 1,842 Sessions *Debugging page was removed minutes after a researcher alert; company reports no evidence of further access but says its review is incomplete* Cloudnote, the widely used note-taking service with 3.2 million registered users, exposed a debugging page containing excerpts from customer prompts to the public internet without requiring authentication for four hours on Monday, the company acknowledged. According to logs reviewed by Cloudnote, the page displayed snippets of up to 200 characters drawn from 1,842 user sessions. The company said full documents, passwords and payment information are kept in separate systems and were not present on the page, limiting but not eliminating the potential sensitivity of what was visible. Cloudnote removed the page 19 minutes after an outside security researcher alerted it to the exposure. The company said it has found no evidence so far that anyone other than the researcher accessed the page, but it emphasized that its investigation is incomplete, so additional access cannot yet be ruled out. That distinction matters because prompts sent to AI assistants often include personal, business or creative details users consider private, and even a 200-character excerpt can reveal a great deal. For now, it remains unknown whether other visitors saw the page, whose prompts were included, and what, if anything, affected users should do next. The company has not yet said how it will notify users.

Qwen: Qwen3.8 Max Prime: # Cloudnote debugging page exposed customer prompt excerpts for four hours *Company says no evidence of access beyond researcher; investigation incomplete, no full documents or payment data involved.* Cloudnote said a debugging page containing excerpts from customer prompts was publicly accessible without authentication for four hours on Monday, exposing a limited slice of user interactions with the note-taking service. The company, which has 3.2 million registered users, removed the page 19 minutes after a security researcher reported it. Confirmed logs show the page displayed excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters from a prompt. Cloudnote has not determined what specific text appeared in every case, nor whether any exposed excerpts contained sensitive personal details, because its investigation is still underway. The company said it has found no evidence that anyone other than the researcher accessed the page during the exposure window. That finding remains preliminary while the investigation continues. Cloudnote also said full documents, passwords and payment information were stored elsewhere and were not present on the exposed page. The incident matters because even short prompt fragments can reveal personal interests, work projects or confidential questions, depending on what users entered. Until the review is complete, Cloudnote cannot rule out broader impact, but it has confirmed the exposure was brief, limited to excerpts and did not include full account records.

muse-spark-1.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Qwen: Qwen3.8 Max Prime: <think> # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.