Head to head: Anthropic: Claude Opus 5.5 vs Meta: Muse Spark 1.3

Anthropic: Claude Opus 5.5: # QueueZero's Proxy Assistant Turns Audience Remark Into CEO Task at Launch *The AI meeting tool, priced at $30 per user per month, enters public beta next week* QueueZero on Wednesday launched Proxy, an artificial intelligence assistant that can attend meetings, produce notes and assign follow-up work to software agents. The launch demonstration also offered an unscripted look at how the system decides what counts as an instruction. During the demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the remark as an instruction and 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. ## What Proxy does Proxy goes beyond keeping a record of a meeting. In addition to producing notes, it can turn what it hears into follow-up work and assign that work to software agents. This is the part of the product most likely to save time, and the part where errors are most likely to occur. The incident on stage showed the core difficulty. Spoken conversation mixes formal requests, offhand comments and remarks that are only partly serious. The audience member's comment contained a legitimate business request alongside a pointed aside, and Proxy treated both parts as tasks to be assigned. ## Where the safeguards sit QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically, which is what happened during the demonstration. Administrators can disable automatic task creation. Under that design, a misread remark is more likely to produce an unwanted internal task than an unapproved message to a client or partner. That is a meaningful limit, but internal tasks still carry weight. Depending on how an organization uses Proxy, an automatically created assignment could land in a colleague's queue or be passed to a software agent before anyone reviews it. For companies evaluating the product, the practical question is how much automatic action they want. Teams that value speed may leave task creation on. Teams with stricter review requirements, or meetings that involve customers and outside parties, may prefer to switch it off and approve tasks manually. Of the two tasks Proxy assigned on Wednesday, the pricing sheet was the easier one to close. ## Pricing and availability Proxy costs $30 per user per month. It enters public beta next week.

Meta: Muse Spark 1.3: # QueueZero Launches Proxy Meeting Assistant QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend virtual meetings, produce notes and assign follow-up work to software agents. The announcement was made during a launch event in San Francisco, where Chief Executive Evan Wu demonstrated the system live to customers, partners and journalists. Proxy connects to major videoconferencing platforms and functions as a meeting participant. According to the company, it transcribes discussion, summarizes decisions and identifies action items. Those items can then be routed to connected software agents for execution without manual data entry. The demonstration encountered an unexpected moment when 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 Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. He then shared the pricing information with attendees and continued the presentation. The incident highlighted questions about how autonomous assistants distinguish casual remarks from directives. QueueZero said Proxy uses contextual signals to identify instructions, but informal comments made during a transcribed meeting can still trigger task creation. The company said Proxy requires confirmation before an external message is sent. Internal tasks, by contrast, can be created automatically. Administrators can disable automatic task creation and require manual review of proposed tasks, the company added. Industry analysts said the distinction is significant for enterprise adoption. Systems that can initiate internal workflows reduce administrative burden, but they also increase the risk of unwanted or incorrect actions if controls are not properly configured. QueueZero said audit logs record task origins and approvals, allowing administrators to review how each item was created. The launch comes amid growing competition among enterprise software vendors to automate routine meeting administration. Vendors have focused on transcription accuracy, summarization quality and integration with workplace systems as differentiators for corporate buyers. Proxy costs $30 per user per month. It enters public beta next week, with general availability expected after customer feedback is incorporated. The company did not disclose early customer numbers. It said additional integrations with project management and customer relationship tools will be announced during the beta period.

Anthropic: Claude Opus 5.5: **6** 1. `#Flood software moves upstream` has no space after `#`, so it does not use `# Headline`. 2. `The company is expanding its pilot.` is in the dek position but is not italicized. 3. `##Details` has no space after `##`, so it does not use `## Heading`. 4. `Read [the announcement](www.example.com/news).` uses a link that is not an absolute HTTPS URL. 5. `* The first pilot begins Monday.` starts the list item with an asterisk instead of a hyphen. 6. `<div>The second pilot begins Friday.</div>` uses raw HTML.

Meta: Muse Spark 1.3: 6

Anthropic: Claude Opus 5.5: # Cloudnote Exposed Prompt Excerpts From 1,842 Sessions on Unprotected Debugging Page *The page was publicly accessible for four hours on Monday. Cloudnote says it has found no sign that anyone other than the researcher who reported it viewed the page, but its investigation is not finished.* Cloudnote left a debugging page containing excerpts of customer prompts accessible without authentication for four hours on Monday, according to the company. Logs show the page held text from 1,842 user sessions. The company removed the page 19 minutes after a security researcher notified it of the problem. Cloudnote says it has so far found no evidence that anyone besides that researcher accessed the page. It has not yet completed its investigation. Each excerpt could include up to 200 characters from a prompt. That is a small amount of text, but it may be enough to capture names, work details or other information a user typed. Full documents, passwords and payment information were stored elsewhere and were not present on the page, the company said. The number of affected sessions is small compared with Cloudnote's 3.2 million registered users. Several key questions remain open, however. Because the logs count sessions rather than people, it is not yet clear how many individual users were affected. Nor is it known what the exposed excerpts contained, or whether the company's final review will confirm that no one else saw them.

Meta: Muse Spark 1.3: # Cloudnote Left Prompt Excerpts Exposed for Four Hours *Snippets from 1,842 sessions were visible on an unauthenticated page. The company says it sees no sign of misuse so far, but its investigation is not complete.* Cloudnote left a debugging page that displayed excerpts of customer prompts accessible on the public internet without authentication for four hours on Monday, the company said after an outside researcher alerted it to the lapse. Server logs reviewed by the company show the page contained snippets from 1,842 user sessions during the exposure window, each showing up to 200 characters of a prompt entered into Cloudnote, which reports 3.2 million registered users. Cloudnote said it removed the page 19 minutes after an outside security researcher reported it and said it has so far found no evidence that anyone other than the researcher accessed it. The company cautioned that its review of access logs is still underway and it cannot yet rule out additional views. Full user documents, passwords and payment information were stored separately and did not appear on the debugging page, according to Cloudnote. But even brief prompt excerpts can contain sensitive personal or business details. The incident underscores the risk of leaving internal diagnostic tools publicly reachable, even briefly, and leaves open what was potentially visible and whether affected users will be notified as the investigation continues.

Anthropic: Claude Opus 5.5: **C**

Meta: Muse Spark 1.3: C

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup's software combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews before a storm.* 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. Basin Ventures led the round, with participation from existing investor Long Mile. The new funding brings Raincheck's total raised to $10.5 million. CEO Ana Flores said the software is meant to address a problem that comes after cities identify where flooding is likely: > "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." Details of [Raincheck's methodology](https://example.com/raincheck-methodology) are available online. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. The company's forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap and municipal engineering contractors. The company says it will use the new funding to hire eight engineers and expand into Florida.

Meta: Muse Spark 1.3: # Raincheck raises $8 million for flood-risk software for cities *The Austin company helps local governments prioritize maintenance before storms; the seed round led by Basin Ventures will fund hiring and expansion into Florida.* Austin-based Raincheck has raised $8 million to expand its flood-risk software for local governments, the company announced today. The company maps drainage systems and combines those maps with rainfall forecasts to help cities decide where to send maintenance crews before a storm. Its approach is detailed in [Raincheck's methodology overview](https://example.com/raincheck-methodology). The seed round was led by Basin Ventures, with participation from existing investor Long Mile. The new financing brings Raincheck's total raised to $10.5 million. CEO Ana Flores said in a statement: > “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, and emergency managers retain control over deployments and public warnings. 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.

Anthropic: Claude Opus 5.5: # 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.

Meta: 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.

Anthropic: Claude Opus 5.5: # The Top Two Support Models Tied, and One Costs 61% Less When researchers put five language models through a customer-support benchmark, they did not find a single clear winner. The two strongest models, North and Vale, finished in a statistical dead heat. The real difference between them was price: Vale delivered the same measured quality at about 39% of North's estimated cost, and it answered faster. For teams choosing a model to draft support replies, that changes the question. It is less about which model is best and more about how much a fraction of a point is worth, when that fraction may not be a real difference at all. ## How the test worked The researchers collected 200 anonymized customer-support tickets and had each of the five models write one proposed answer per ticket. Three experienced support managers then graded the answers without knowing which model wrote them. Blind grading matters here: graders who know a response came from a well-known or expensive model may judge it more generously. Each answer received a combined score covering three things a support manager cares about: whether the answer was factually correct, whether it actually helped the customer, and whether the tone was appropriate. ## The headline: a tie at the top North posted the highest average score, 84.2, with Vale close behind at 83.8. The researchers found that this 0.4-point gap was not statistically significant. With 200 tickets and three graders, a difference that small could easily come from which tickets happened to be in the sample or from normal variation in grading. It should not be read as evidence that North is better. The cost gap, by contrast, is large and clear. North's estimated cost was $1.90 per 1,000 tickets, compared with $0.74 for Vale. Vale was also 18% faster, which can matter in live chat or any workflow where an agent waits for a draft. Put together, the benchmark gives no measurable reason to pay a premium for North across general support work. A note of perspective: even North's cost works out to a fraction of a cent per ticket. For a small team, the savings from switching may be trivial next to the cost of one bad answer reaching a customer. The price difference matters most at high volume, where it adds up. ## Where the models differed Averages can hide specialization, and the results hint at some. North performed best on billing disputes, while Vale performed best on technical troubleshooting. A company whose queue is dominated by one of these categories might reasonably favor the matching model. These category results should be treated as leads, not conclusions. The benchmark summary does not say how large the gaps were or whether they were statistically meaningful. A subset of 200 tickets is smaller still, which makes chance results more likely. Teams with a strong lean toward one ticket type should run their own comparison on that type. ## The budget tier The three cheaper models traded quality for price. Ember (79.1) and Kite (76.4) scored several points below the leaders at lower costs, and Harbor was the least expensive at $0.08 per 1,000 tickets but scored lowest at 71.0. Whether those trade-offs are acceptable depends on how the drafts are used. A model whose output is always reviewed by a human agent can afford to be less polished than one sending replies directly to customers. ## A shared blind spot: images One weakness cut across the field. All five models performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message. Picking a stronger model will not fix this. Teams should plan to route image-dependent tickets to human agents or handle them through a separate process. ## What the benchmark doesn't tell you The test had real limits, and they bear on how far the results travel: - **One company, one language.** All tickets were in English and came from a single software company. Results may differ for retailers, banks, or healthcare providers, and for customers writing in other languages. Multilingual support was not tested. - **Drafting only.** The models wrote answers but did not take actions. The researchers did not test tool use or agentic tasks, such as issuing refunds, looking up account records, or updating tickets. Many production support systems depend on exactly those abilities. - **Estimated costs.** Pricing figures are estimates and will vary with ticket length, vendor pricing changes, and how a system is built. ## The takeaway Within its scope, this benchmark supports one confident conclusion: Vale matched North's quality for well under half the price and with faster responses. Everything beyond that, including the category strengths and how the models would handle actions, other industries, or other languages, is a reason to test on your own tickets before committing.

Meta: Muse Spark 1.3: # When Second Place Wins on Support Customer-support leaders shopping for AI usually start with a simple question: which model gives the best answers? A new benchmark of five language models suggests that's the wrong question. The right one is: which model gives answers that are good enough, at a price and speed you can live with? In the test, each model drafted a response to the same 200 anonymized tickets from a software company. Three experienced support managers, grading blindly without knowing which model wrote what, scored them on a combination of factual correctness, usefulness, and tone. That human judgment matters, because support quality isn't just about getting facts right; it's about resolving frustration cleanly. ## A tie at the top, a rout on value The headline result was a statistical dead heat. The highest-scoring system, dubbed North, averaged 84.2, while Vale averaged 83.8. Researchers say that gap is not statistically significant, which in plain English means we can't tell them apart on quality in this test. Treat North as the “winner” and you'll miss the real story. Vale did that effectively tied job for about $0.74 per 1,000 tickets, compared with about $1.90 for North. That's less than half the estimated inference cost, and Vale was also 18% faster to respond. In a queue where seconds shape customer satisfaction and staffing costs, speed plus savings with no meaningful drop in quality is the clearest — and most actionable — finding. ## Not the same strengths Look closer and the tie breaks into trade-offs. North was strongest on billing disputes, the tickets where precision about charges, refunds, and policy language matters most and mistakes erode trust fast. Vale was strongest on technical troubleshooting, where step-by-step usefulness often decides whether a customer needs another reply. For a software company — and most buyers reading this — that split is worth more than the overall average. Choose according to your ticket mix, not the leaderboard. ## What about the cheaper models? The other three systems show why chasing bargains still hurts. Ember scored around 79 and Kite around 76, at roughly thirty and twenty-two cents per thousand tickets. The cheapest system, Harbor, managed only about 71 for eight cents. That looks tempting on a spreadsheet, but a five- to thirteen-point gap in a human-graded quality score is hard to ignore when each point represents more wrong answers, unhelpful steps, or off-tone replies that create follow-ups. ## Where all of them stumbled Just as revealing is where every model struggled: tickets that depended on an image attachment. Screenshots of error messages, annotated photos, and scanned invoices tripped them all up, a reminder that text-only testing flatters text-only systems. If your queue is full of “see attached” tickets, these scores overstate what you'll get in production. ## Read the fine print before you buy The scope is narrow by design. All 200 tickets were in English and came from a single software company, so results may not transfer to retail, healthcare, or other domains, let alone to other languages. The researchers also didn't test agentic behavior or multilingual support — no tool use, no refund issuance, no account lookups, no multi-step follow-through. That matters because modern support bots are judged less on a single polished draft and more on what they can actually do. Treat this as a test of first-draft writing ability, not autonomous resolution. For buyers, the practical takeaway is straightforward. Don't pay more than double for a statistical tie, especially when the cheaper option is faster and better at the technical cases that dominate software queues. Pilot the value leader on your own tickets, weight billing accuracy if that's your risk center, and build a separate plan for images. In support, the best model on paper rarely survives contact with cost, speed, and your real inbox.

Anthropic: Claude Opus 5.5: # Software Update Sends $2.8 Million in Duplicate Reimbursements to Employees of 91 Companies **KiteBridge says most of the money has been recovered, but questions remain about a safeguard meant to prevent exactly this kind of error.** Ninety-one companies paid out about $2.8 million they did not owe on Tuesday after an update to KiteBridge's expense-management software issued 13,400 duplicate reimbursements to their employees. By Wednesday evening, $2.3 million had been recovered. The remaining $500,000 was still pending. KiteBridge's software handles reimbursements for corporate customers, paying employees back for work-related expenses such as travel and meals. The duplicate payments came from the employers' funds. No employees had money taken from their own accounts. ## How the error unfolded The problem began after KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. At some point, a payment processor, the outside service that moves money between bank accounts, timed out. A timeout means a request goes unanswered within the expected window, so the sending system cannot tell whether the payment went through. Payment systems usually handle this by retrying the request. To keep those retries from paying anyone twice, the system has to recognize each one as a repeat of the original. In KiteBridge's case, its fallback system, the backup process that takes over when the normal path fails, treated each retry as a new reimbursement request. Each retry produced another payment. The first customer complaint reached KiteBridge at 9:12 a.m., 27 minutes after the update went live. The company suspended automated payments at 10:03 a.m., 51 minutes after that complaint and roughly an hour and 20 minutes after the deployment. During that window, the software generated the 13,400 duplicate reimbursements. ## A safeguard that was supposed to prevent this The failure is notable because KiteBridge had told customers it was protected against this kind of error. A company changelog published in April said version 4.0 had introduced "exactly-once payment protection." In payments, "exactly once" means each transaction is carried out a single time. It should not be dropped, and it should not be repeated, even when networks fail or requests are retried. Duplicate payments triggered by retries are the problem such a feature is designed to stop. The reporting does not establish why the protection described in April did not prevent Tuesday's duplicates, or how the version 4.2 update or the fallback system bypassed it. ## A delayed disclosure KiteBridge's public account of the incident also drew scrutiny. Its public incident page first described the problem as delayed reimbursements. That description fit what customers would have seen once automated payments stopped, but it left out that thousands of extra payments had already gone out. The page first disclosed the duplicate payments six hours after the company suspended automation, around 4 p.m. Tuesday. ## Company response CEO Lena Ortiz apologized for the incident. She said KiteBridge would publish a technical report, which could explain why the exactly-once protection did not hold and how the company plans to prevent a repeat. The company did not give a date for the report. KiteBridge has not said when it expects to recover the remaining $500,000, or what happens to that money if it cannot be recovered.

Meta: Muse Spark 1.3: # KiteBridge glitch sent $2.8 million in duplicate payments to employees **91 companies affected as fallback system mistook retries for new requests; $2.3 million recovered** Thousands of workers opened their bank apps on Tuesday to find a pleasant surprise that wasn't theirs to keep: a second — and sometimes third — expense reimbursement they had never requested. The phantom deposits, totaling $2.8 million across 13,400 duplicate payments, came from KiteBridge, a maker of expense-management software used by corporate finance teams to approve and send reimbursements to employees. The company said employers, not employees, funded the erroneous payouts, and no money was withdrawn from workers' personal accounts. The error began Tuesday morning when KiteBridge deployed version 4.2 of its platform. The update went live just before business hours. Shortly afterward, its outside payment processor — the banking intermediary that actually moves money — began timing out. In plain terms, KiteBridge was asking the processor, "Did this payment go through?" and getting no answer. That silence triggered KiteBridge's fallback system, a backup designed to keep reimbursements moving when the main route fails. Instead of recognizing its repeated follow-ups as checks on the same payment, the fallback treated each retry as a brand-new reimbursement request. The result was a cascade of duplicates sent to employees who had submitted legitimate expenses. The first customer complaint arrived less than half an hour after the deployment, when a finance manager noticed double payouts. About an hour later, KiteBridge suspended all automated payments, stopping further damage. By then, 91 corporate customers had been affected. The company's early public communication understated the problem. Its incident page initially described the issue only as delayed reimbursements, a framing that persisted for hours. The page did not acknowledge that duplicate payments had been sent until about six hours after automation had been shut off, well into the afternoon. That delay matters because duplicate payouts create messy cleanups. Companies must ask banks to reverse transfers and, in some cases, ask employees who have already spent the money to return it. By Wednesday evening, KiteBridge said it had recovered $2.3 million, leaving about $500,000 still pending with banks and recipients. The incident is especially embarrassing because KiteBridge had promised this type of failure could not happen. An April changelog for version 4.0 said the company had introduced "exactly-once payment protection," an industry term for safeguards that ensure that even if a computer retries a task many times, the money moves only once. Tuesday's bug suggests that protection either did not cover the fallback path or failed under real-world timeout conditions. Chief Executive Lena Ortiz apologized for the failure and said the company would publish a full technical report explaining what went wrong and how it will prevent a recurrence. She did not give a date for the report. Automated payments were suspended while KiteBridge investigates. The episode is a reminder of how fragile automated money movement remains. A single missed signal between two computer systems turned routine expense reports into millions of dollars in unwanted deposits — and left dozens of businesses chasing their own money back.

Anthropic: Claude Opus 5.5: # Foundry Hall to close after losing its largest sponsor Foundry Hall, the Milwaukee nonprofit that has helped more than 200 startups get off the ground since 2018, will close Oct. 31. Its largest corporate sponsor declined to renew its support, and no replacement emerged. Blank River Bank had supplied 38% of the accelerator's annual budget. Without that money, Foundry Hall can keep its doors open through October but cannot fund another full program. Six employees will lose their jobs, and the 12 startups now working out of its space must find somewhere else to go. Executive director Camille Ross said the organization spent months trying to close the gap while hoping a new backer would step in. "We kept waiting for the replacement check that never came," she said. In a statement, a Blank River Bank spokesperson said the decision reflected a shift in the bank's community investment priorities and had nothing to do with how Foundry Hall performed. By the measures accelerators usually rely on, the organization performed well. The 212 startups it has supported now employ a combined 1,140 people. That makes it one of the more visible parts of Milwaukee's small but growing startup community, which has often struggled to keep young companies from leaving for larger markets. For many founders, though, the numbers capture only part of what the place meant. Dev Malik, who went through the program, said Foundry Hall gave him a level of credibility he could not have built alone. Before he arrived, he said, he was pitching from the back table of a coffee shop. "My first investor meeting happened in that building," he said. That kind of starting point is hard to replace, and the timing leaves little room to try. The 12 companies currently in residence have roughly weeks, not months, to relocate, and none of them yet knows where it will land. City officials have begun looking for options. Deputy Mayor Susan Cole said the city is in talks with building owners and local universities about space for the displaced companies, though she did not announce any agreements. The search for space addresses only the most immediate problem. A new desk can be found. The mentorship, the introductions to investors and the sense of being taken seriously that Foundry Hall offered are harder to move into a borrowed room. For now, Ross and her staff are focused on the final weeks: helping current companies plan their exits and winding down an organization that ran for six years largely on the strength of a single sponsor's commitment. When the doors close at the end of October, the companies Foundry Hall helped launch will carry on without it. The next founder pitching from a coffee shop table will not have the same place to go.

Meta: Muse Spark 1.3: # A Quiet Ending for Foundry Hall **Milwaukee's nonprofit accelerator will close Oct. 31 after losing its anchor sponsor** MILWAUKEE — Foundry Hall opened in 2018 with exposed brick, borrowed desks and a simple premise: Milwaukee founders should not have to leave Milwaukee to build a company that lasts. On Oct. 31, that experiment will end. The nonprofit startup accelerator will close its doors after its largest corporate sponsor, Blank River Bank, which supplied 38 percent of its annual budget, declined to renew its support. Executive director Camille Ross said staff spent the spring and summer courting foundations and regional employers to fill the gap. “We kept waiting for the replacement check that never came.” Without it, the math no longer worked. Foundry Hall has enough money to operate through October, but not enough to fund another full program without cutting into commitments to its current cohort. Rather than start something it could not finish, the board voted to wind down. In its eight years, Foundry Hall supported 212 startups, which collectively employ 1,140 people across the region — logistics firms, health software makers, food brands that began as farmers-market stalls. For a city that has long watched young talent drift toward Chicago or the coasts, those numbers became a quiet point of civic pride. The closure will be felt first in the building itself. Six employees will lose their jobs. Twelve startups now using its workspace must leave by the end of the month, clearing out conference rooms and phone booths that, for many, served as a first real office. Founder Dev Malik arrived with little more than a slide deck he used to rehearse at the back table of a coffee shop. His first investor meeting happened at Foundry Hall, and he credits the crowded hallways — the chance encounters with mentors, accountants and other sleepless founders — with giving his company credibility before it had revenue. That kind of density is hard to replace. Blank River Bank has said the move stemmed from a shift in its charitable focus, not from any dissatisfaction with the accelerator's work, a distinction that offers little comfort to the teams now searching for desks. Deputy Mayor Susan Cole said the city is in talks with landlords and nearby universities to find temporary homes for the displaced teams, an effort to keep the twelve from scattering. It may work. Alumni networks rarely disappear when a lease ends, and several founders said they plan to keep meeting, with or without the brick building. But an accelerator is more than introductions. It is rent subsidized, advice formalized, someone to answer the phone when payroll does not add up. When Foundry Hall locks its doors on a late October evening, it will leave behind more than empty desks. It will leave a question Milwaukee has asked before: who pays to give new businesses a beginning?

Anthropic: Claude Opus 5.5: # HelioDock Launches Remote Fleet Software for Warehouse Robots, Raises $24 Million *The Night Shift platform lets operators oversee autonomous robots from afar, though one early customer says its value drops when the network does.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor their robot fleets remotely. The company also announced a $24 million Series B funding round led by North Coast Ventures. The launch gives HelioDock a recurring, per-robot software product to sell alongside its machines. It enters a fleet-management market where Relay Robotics and YardOS already offer competing tools. HelioDock's pitch is that remote oversight can cut costly disruptions during off-hours. That claim rests on the company's own beta figures, which have not been independently verified, and on network connections that at least one early customer says cannot always be counted on. Chief Executive Mara Velez framed the product around a familiar pain point for warehouse operators. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," she said. ## Fourteen months of testing Night Shift spent 14 months in a private beta with nine logistics companies operating a combined 310 robots. According to HelioDock, those customers saw 17% fewer unplanned robot stoppages during the test. The figure has not been independently verified, and HelioDock is the only source for it. Even so, it is the central evidence the company is offering that remote monitoring changes outcomes on the warehouse floor, rather than simply moving the monitoring to a different screen. ## Connectivity is the constraint The software's usefulness depends on a reliable connection between the robots and the people watching them. 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. When connectivity fails, HelioDock's robots return to their locally configured behavior, and remote control becomes unavailable. A network dropout therefore does not leave the machines without instructions. It does, however, remove the remote intervention that Night Shift is built to provide. That trade-off matters most in the scenario Velez described. A robot stuck on an unfamiliar pallet in the middle of the night is exactly the situation remote control is meant to resolve. At a site with an unreliable rural connection, an operator may be unable to reach the robot during an outage, and the drive across town could still be necessary. For operators in well-connected facilities, Pike's experience suggests the system performed as intended. For those running warehouses in areas with spotty cellular coverage, network reliability is likely to be as important a consideration as the software itself. ## Pricing and availability Night Shift will cost $90 per robot per month and will become generally available on September 15. That rate works out to $1,080 per robot per year. A fleet the size of the entire beta program, 310 robots, would cost about $27,900 a month at that price. Because the fee scales with every robot an operator adds, the software's cost will grow alongside fleet size. Customers weighing it against products from Relay Robotics and YardOS will be comparing not only features but also how those costs add up across a full deployment. ## New funding The Series B brings HelioDock's total funding to $39 million, meaning the company had raised $15 million before this round. North Coast Ventures led the new investment. The raise arrives as HelioDock moves from selling robots toward a model that also includes ongoing software revenue. Night Shift is the first product in that shift, and its general release in September will give the company its first test outside the nine logistics firms that helped shape it.

Meta: Muse Spark 1.3: # HelioDock Launches Night Shift to Monitor Warehouse Robots Remotely *Software emerges from 14-month beta with nine operators as company adds $24 million in new funding* HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of its autonomous warehouse robots remotely. The introduction connects an operational problem in overnight logistics with a business expansion. HelioDock, which makes autonomous warehouse robots, also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. Night Shift will cost $90 per robot per month and become generally available September 15. Competing fleet-management products are available from Relay Robotics and YardOS. CEO Mara Velez described the issue as travel for tasks that could be handled off site. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet.” That scenario — a single stopped machine holding up work during overnight hours — is the focus of Night Shift. Rather than keeping managers on site or on call for travel, the software provides remote supervision of the fleet. Remote control is part of that supervision, according to the company’s description of what is lost when connections drop. HelioDock builds the robots the software supervises. Night Shift spent 14 months in private beta before Thursday’s launch. Nine logistics companies participated, using the software with a combined 310 robots. Averaged across participants, that beta pool represents about 34 robots per company. The test covered multiple logistics operators over 14 months and provides the basis the company cites for its performance claim. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages while using Night Shift. The figure has not been independently verified. The claim remains attributed to HelioDock alone. Unplanned stoppages include interruptions of the kind Velez cited, in which a robot encounters a pallet it cannot resolve on its own. Reducing those interruptions, even when the robot remains in the building and the operator does not, is the central assertion behind the product. Beta customer Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. Pike’s account points to the role of network access in remote supervision. Where connectivity was strong, monitoring from afar functioned as intended. Where LTE service dropped in rural areas, the value of the remote layer declined. His comments came as a beta customer. When connectivity fails, HelioDock says robots return to their locally configured behavior. Remote control becomes unavailable at that point. That fallback means the machines continue to operate under rules set on site, without input from a distant operator. For managers considering overnight coverage, the distinction matters: remote monitoring is available when the link holds, and local behavior governs when it does not. The pricing structures the product as an ongoing fleet expense rather than a one-time addition. At $90 per robot per month, charges scale directly with fleet size. A fleet of the same size as the beta pool, 310 robots, would represent $27,900 per month at list price. The rate applies per robot each month. General availability is set for September 15. Until then, the software remains outside broad release following the private test period. The financing adds $24 million and brings total funding to $39 million. North Coast Ventures led the Series B round, which was announced alongside the product launch. Relay Robotics and YardOS also offer fleet-management products, giving warehouse operators alternative suppliers for supervision software. No feature or price comparison was provided for those offerings.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark and the Quiet Work After the Trade Most conversations about AI in finance focus on the moment a trade happens: the signal, the model, the execution. LedgerLark is interested in what comes afterward, when confirmations, custodian records and internal ledgers have to agree with one another, and often don't. ## Where the idea came from Co-founder Mina Batra spent seven years in operations at a regional brokerage. One episode in particular stayed with her: a six-person team spent three days resolving a single trade mismatch. The cause was a copied account number. The error itself was small. What struck Batra was the cost of finding and fixing it. Post-trade operations at smaller firms often depend on people cross-checking records from several sources, and a minor data-entry mistake can absorb days of skilled labor. Her co-founder, Owen Carr, approached the same problem from the other side. He built reconciliation systems at a large bank, where tooling and budgets for this work are far more substantial. The pairing suggests a specific thesis: small investment firms face many of the same reconciliation problems as large institutions but lack the internal engineering resources to build their own solutions. ## What the product does LedgerLark builds AI agents that compare trade confirmations, custodian records and internal ledgers. When the agents find a discrepancy, they propose a correction. A human reviews and approves it. The boundaries matter here. The product does not execute trades or move money. That design choice narrows what LedgerLark can do, but it may also make the product easier to adopt. Firms that are wary of handing operational authority to software can treat the agents as an assistant that surfaces and drafts fixes rather than one that acts on its own. Whether the human-approval step remains a meaningful check or becomes a rubber stamp over time is a fair question, and one that will likely depend on how each customer configures its workflows. ## Early traction LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row. The most concrete performance figure comes from the company itself: LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That is a notable change if it holds, but it describes a single customer, and the figure has not been independently verified. It also leaves open questions about the time period, the types of exceptions involved and whether other changes at the firm contributed. Results across five customers, measured over longer periods, would offer a clearer picture. ## Competition and friction LedgerLark is entering a field that already has incumbents. Established reconciliation vendors sell into this market, and large banks often rely on internal software. The startup's bet is that AI agents can handle messier, less structured mismatches than rules-based tools, and that small firms are underserved by existing options. Both claims will be tested as competitors add their own AI features. Sales cycles are another challenge. Financial institutions frequently require lengthy security and procurement reviews before granting a vendor access to sensitive records. For a seed-stage company, long reviews can slow revenue growth and stretch runway, even when prospective customers are interested. ## What to watch LedgerLark's founders are working on a problem they know firsthand, and the product's limited scope is a deliberate, understandable choice. The next phase will show whether that insight translates into durable demand. Key signals include whether the company can show consistent results across more customers, how quickly it moves through procurement at new firms, and whether its agents can stay meaningfully ahead of incumbents adding similar capabilities. For now, LedgerLark is a small company making a specific wager: that much of the cost in post-trade operations hides in ordinary errors, and that catching them sooner is worth paying for.

Meta: Muse Spark 1.3: # Startup Spotlight: LedgerLark Wants to Clean Up Post-Trade Messes Post-trade operations rarely make headlines, but for small investment firms, they can consume entire weeks. LedgerLark, a young startup, is building AI agents aimed at that back-office work — the matching, checking, and correcting that happens after a trade is agreed. The premise is straightforward: smaller firms face many of the same settlement and record-keeping requirements as larger institutions, but without large operations teams to handle exceptions. When records disagree, staff must track down the source across multiple systems before anything can settle cleanly. ## Why founders focused here Co-founder Mina Batra spent seven years in operations at a regional brokerage before starting the company. Co-founder Owen Carr previously built reconciliation systems at a large bank. Their backgrounds cover both sides of the problem — living with manual exception queues, and building the internal tools meant to reduce them. The initial idea, Batra said, took shape after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. It was not a complex failure of strategy or markets, but a small data entry error that rippled across confirmations and ledgers. That kind of brittle, time-consuming work became the focus for what would become LedgerLark. ## What the product does The product compares three sources — trade confirmations, custodian records, and a firm’s internal ledgers — to flag inconsistencies. When it finds a likely mismatch, it proposes a correction for a human to review and approve. The company emphasizes that the system does not execute trades or move money, a boundary that matters in a regulated environment where automation raises questions about control and accountability. That human-in-the-loop design reflects how the work actually happens. Exceptions often require judgment about which record to trust and whether a client needs to be contacted. LedgerLark is positioning its agents as a way to assemble the relevant context faster, rather than as an autonomous decision-maker. ## Early traction and open questions On traction, LedgerLark says it has five paying customers, all small investment firms. Annual contracts range from $36,000 to $110,000, suggesting deal sizes that align with limited operations budgets but also leave questions about how quickly revenue can scale. As one early data point, the company says one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the product. That figure has not been independently verified, and it is unclear how representative it is across different firm types or trading volumes. The company has raised a $3.5 million seed round led by Motive Row, which it plans to use for product development and early go-to-market work. Competition is a real factor. Established reconciliation vendors already serve this space, and many larger firms rely on internal bank software built over years to fit specific workflows. For LedgerLark, the question is whether smaller firms — often underserved by heavyweight systems — will prefer a lighter agent-based tool, and whether that tool can integrate reliably with the custodians and accounting platforms those firms already use. Another open question is sales cycles. Financial institutions may require lengthy security and procurement reviews, even for software that only reads data and suggests fixes. For a seed-stage company with a small team, extended diligence processes could slow deployment and make support demands heavier than contract sizes might suggest. Still, the problem LedgerLark has chosen is concrete and easy to observe: too many hours spent chasing mismatched numbers. If its agents can reliably reduce that queue without introducing new risks, small operations teams might notice the difference in their daily workload. Whether that translates into a durable business will depend on integration breadth, accuracy under messy real-world data, and the ability to earn trust in a cautious industry.

Anthropic: Claude Opus 5.5: # Arcwell's Relay Puts an AI Assistant Beside Customer Support Agents Arcwell is betting that much of the work of answering customer questions is really the work of finding information a company already has. Its new platform, Relay, uses artificial intelligence to examine customer support conversations and suggest answers to the agents handling them. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Suggested answers are only part of the package. Relay can also flag customers who may be preparing to cancel and generate summaries. Together, those features cover tasks that often pull agents away from the conversation in front of them. The early evidence comes from a six-month test with 12 companies. Arcwell says agents using Relay closed tickets 14% faster. That gain is the only performance figure the company has made public. Arcwell did not release customer-retention results, so there is no data yet on whether the cancellation-warning feature helped those businesses hold on to customers. For many buyers, that could be the most valuable claim of all. Relay costs $65 per agent each month. For prospective customers, the math will come down to whether faster ticket handling, and any eventual retention benefit, justifies a fee that grows with every seat on the support team. There is also the question of what companies must hand over. Relay requires access to support transcripts and internal help documents, records that can include customer details and information businesses would rather keep private. Arcwell says customer data is encrypted and is not used to train shared models, an assurance aimed at companies reluctant to let their records improve a system other clients also use. Companies can choose a 30-day or 90-day retention period. Relay is Arcwell's attempt to help companies improve customer service with artificial intelligence, and the company is starting with a claim that is easy to measure: speed. Proving that the platform can also keep customers from leaving will require the retention data Arcwell has not yet shared.

Meta: Muse Spark 1.3: # Arcwell Launches Relay, an AI Assistant for Support Agents Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. Instead of responding to customers directly, Relay listens in the background, scanning the unfolding exchange and pulling from past transcripts and internal help documents to recommend what the agent should say next. The platform can also identify customers who may cancel and create summaries of those conversations. The design reflects a familiar compromise in customer service software. Rather than promising fully automated support, vendors are increasingly selling AI as a real-time copilot — one that keeps a human in the loop while shaving seconds off every interaction. 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 linked to its ability to flag at-risk customers. Relay costs $65 per agent each month. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. That problem is easy to recognize on a busy support floor, where agents often toggle between ticket queues, knowledge bases and chat windows while customers wait. Relay is meant to collapse that search into a single suggestion, reducing both handling time and the strain of finding the right language under pressure. The promise is less about replacing judgment than removing drudgery. Such convenience depends on deep access. Relay requires access to support transcripts and internal help documents to function. Arcwell says customer data is encrypted and is not used to train shared models, and that companies can choose a 30-day or 90-day retention period. How enterprises weigh that exchange — faster resolutions against entrusting a vendor with detailed records of customer complaints, cancellations and confusion — will shape Relay’s reception in the months ahead as Arcwell looks to establish itself in AI-assisted customer service.