Head to head: Qwen: Qwen3.8 Max (0902) vs Anthropic: Claude Sonnet 5.5

Qwen: Qwen3.8 Max (0902): # Startup Spotlight: LedgerLark Brings AI Agents to Post-Trade Operations LedgerLark is building AI agents for a part of finance that rarely attracts attention: post-trade operations. After a trade is executed, staff often must confirm details, reconcile records, and chase down mismatches. At small investment firms, that work can consume scarce hours. LedgerLark’s software compares confirmations, custodian records and internal ledgers, then proposes corrections for human approval. It does not execute trades or move money. The founders’ paths converge around this problem. Co-founder Mina Batra spent seven years in operations at a regional brokerage, where she watched routine errors become expensive projects. One incident stuck with her: a six-person team spent three days resolving a trade mismatch caused by a copied account number. Co-founder Owen Carr built reconciliation systems at a large bank, giving him a view of how larger institutions automate similar checks but often through expensive, bespoke tools. That combination points to a clear opening: bring more systematic reconciliation discipline to smaller firms that lack large technology teams. LedgerLark’s agents read documents and records from different sources, flag discrepancies, and suggest fixes. The human-approval step is important. In financial operations, a wrong correction can create downstream reporting or client problems, so keeping people in the loop may ease adoption and satisfy compliance expectations. Early traction is modest but specific. 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 figure is a company-provided example rather than an audited benchmark, and results may differ depending on data quality, workflow complexity and how firms define an exception. The company has raised a $3.5 million seed round led by Motive Row. The funding should help LedgerLark refine its agents, add integrations and navigate the security reviews that often slow sales to financial institutions. Those reviews can be lengthy, especially when outside software touches sensitive trade and client data. Procurement cycles, controls testing and internal risk sign-offs may stretch the time between pilot and revenue. Competition is another reason for caution. Established reconciliation vendors already serve parts of this market, and some firms rely on internal bank software or spreadsheets with custom scripts. Incumbents may have deeper integrations, longer track records and existing relationships. LedgerLark’s likely advantage is focus: small investment firms may be underserved by larger vendors, and an AI-native workflow could be easier to deploy than legacy systems. Still, buyers will need evidence that the software is reliable, auditable and secure. The seed capital may also determine whether LedgerLark expands beyond exception handling into reporting, client onboarding or other operations work. The broader question is whether post-trade exception handling is a narrow tool or the start of a broader operations platform. If LedgerLark can reduce repetitive manual checks, it may free operations staff to focus on higher-value work. But the company will need to show consistent accuracy across trade types, data formats and custodians. For now, the interesting part is not the AI label; it is whether the product can make a tedious, high-stakes process faster, calmer and easier to audit.

Anthropic: Claude Sonnet 5.5: # Startup Spotlight: LedgerLark Brings AI Agents to the Back Office *The startup is betting that small investment firms need help finding trade errors more than they need another trading tool.* A trade is not finished when it executes. Afterward, the broker's confirmation, the custodian's records and the firm's own ledger all have to agree. When they don't, someone must work out why. LedgerLark builds AI agents for that work, known as post-trade operations, and it is aimed at small investment firms. ## An expensive typo Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after watching a six-person team spend three days resolving a trade mismatch. The cause was a copied account number. The error itself was trivial. The cost came from finding it. That is the problem LedgerLark is built around: in many exceptions, the fix is simple once someone locates the source. Batra's co-founder, Owen Carr, built reconciliation systems at a large bank. Together they combine day-to-day operations experience with engineering experience on the same kind of data. ## What the product does LedgerLark's agents compare confirmations, custodian records and internal ledgers, then propose corrections. A human approves each one. The product does not execute trades or move money. That limits the damage a bad suggestion can do, since a person decides what happens. It also raises a question the company will have to answer over time: how often are the proposals right, and will reviewers keep scrutinizing them once most look correct? ## Early traction Five firms are 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. LedgerLark also says one customer reduced its unresolved daily exceptions from 70 to 18. That is a meaningful change if it holds, but it is one customer's result, reported by the company. It is not yet clear over what period the drop occurred, how trade volumes changed, or whether other customers see similar results. With five customers, the sample is small. ## The competition The market is not empty. Established reconciliation vendors already sell to financial firms, and many banks run their own internal software. Both have existing integrations and track records, and buyers in this industry tend to value that. LedgerLark's opening is presumably in the gap between those options: firms too small to build their own tools, and perhaps underserved by products designed for larger institutions. That is a plausible niche, but not a protected one. Incumbents could add similar AI features, and a startup's advantage in speed or focus can narrow. ## The slow part Financial institutions may require lengthy security and procurement reviews, even for a product that cannot move money. LedgerLark's software still touches sensitive records, and a small company has to show it can protect them. These reviews can stretch sales cycles and make growth less predictable than the contract values suggest. ## What to watch Several things will show whether the idea holds up: - Whether results like the 70-to-18 improvement repeat across customers with different systems and data quality - How often human reviewers accept, edit or reject the agents' proposals - Whether early customers renew at the end of their annual contracts - How incumbents respond LedgerLark starts from a real and specific frustration, and it has chosen a deliberately constrained design. Whether that is enough to build a durable business will depend on evidence the company has only begun to collect.

Qwen: Qwen3.8 Max (0902): C

Anthropic: Claude Sonnet 5.5: C

Qwen: Qwen3.8 Max (0902): # Raincheck Raises $8 Million to Help Cities Get Ahead of Flood Risk *The Austin software maker maps drainage networks and rainfall forecasts to guide pre-storm maintenance.* Raincheck, which develops flood-risk software for local governments, announced today that it has raised $8 million. The Austin company maps drainage systems and pairs those maps with rainfall forecasts, giving cities a way 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 now raised $10.5 million in total. > “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.” CEO Ana Flores said the company’s goal is to help cities prioritize limited resources more effectively. More detail on 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 intended as recommendations, not directives. 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.

Anthropic: Claude Sonnet 5.5: # Raincheck raises $8M to help cities decide where to send crews before floods *The Austin startup's software combines drainage maps with rainfall forecasts to help local governments prioritize storm preparation.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in a seed round, the company announced today. Raincheck maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. CEO Ana Flores said the challenge is not identifying which streets flood: > "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 has published [its methodology](https://example.com/raincheck-methodology) online. The company currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. Raincheck's forecasts are recommendations. 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.

Qwen: Qwen3.8 Max (0902): # Customer-Support Drafting: Vale Matches North While Costing Less When teams compare language models for customer support, the tempting conclusion is to crown whichever model posts the highest average score. This benchmark tells a more practical story. Across 200 anonymized support tickets, Model Vale did not differ meaningfully from Model North on overall quality, cost substantially less, and responded faster. For organizations deciding which model to deploy for draft support answers, Vale emerges as the clearer default choice, while North remains interesting for billing-heavy workflows. ## How the benchmark worked The study asked five models to produce one proposed reply for each of 200 customer-support tickets. The tickets came from a single software company and were written in English. Three experienced support managers graded the drafts without knowing which model generated them. They scored each response on factual correctness, usefulness, and tone, then combined those judgments into an average score. The researchers also estimated cost per 1,000 tickets and measured response speed. ## The clearest finding: Vale delivers similar quality for less The top-line numbers were close. Model North averaged 84.2, while Model Vale averaged 83.8. That 0.4-point gap was not statistically significant, meaning the evidence does not support treating North’s overall output as reliably better. Once cost and speed enter the picture, Vale becomes more attractive. Vale was estimated at $0.74 per 1,000 tickets, compared with $1.90 for North, and it answered 18% faster. In a production environment, those differences matter: lower cost makes broader deployment easier, and faster responses can reduce customer wait times or increase agent throughput. The benchmark also showed that the two leaders had different strengths. North performed best on billing disputes, while Vale performed best on technical troubleshooting. That suggests a nuanced deployment strategy rather than a simple winner-take-all choice. A support organization with many billing escalations might favor North for that category, while a software product with frequent technical issues might lean toward Vale. If only one model can be used, Vale’s combination of comparable overall quality, speed, and price gives it the stronger general case. The remaining models filled out a rough cost-quality curve. Ember, Kite, and Harbor were cheaper than both leaders, but their average scores were lower: Ember at 79.1, Kite at 76.4, and Harbor at 71.0. Harbor was especially inexpensive at $0.08 per 1,000 tickets, but its quality gap from the leaders was large enough that most support teams should treat it as a budget option rather than a primary answer generator. The results do not suggest that every cheap model is bad or every expensive model is essential; they do suggest that, in this test, the most compelling balance was found near the top but not at the highest price. ## Limitations that matter One limitation deserves special attention: all five models struggled when a ticket depended on an image attachment. That weakness matters because support conversations often involve screenshots, error messages, invoices, or configuration details. If a ticket cannot be understood without an image, these draft answers may be incomplete or misleading. Teams should either route image-dependent cases to human agents or ensure the workflow includes additional review before sending a generated reply. The benchmark’s scope is also narrow. It used English-language tickets from one company, so the results may not transfer to other industries, languages, or customer expectations. The models were evaluated only on proposed answers, not on whether they could take agentic actions, use tools, search knowledge bases, or update ticket systems. Those capabilities could change the value equation significantly. The grading method is useful but imperfect: three expert judges provide meaningful human judgment, yet scores still depend on the rubric and the specific tickets chosen. ## Bottom line The practical takeaway is not that Model North failed or that Model Vale is universally superior. It is that the benchmark found no meaningful quality gap between the two top performers, while Vale offered clear cost and speed advantages. For many customer-support teams, that makes Vale the safer starting point, with North as a possible specialist for billing disputes.

Anthropic: Claude Sonnet 5.5: # The Most Expensive Support AI Didn't Win: What a 200-Ticket Test Shows A common assumption in AI procurement is that the priciest model is the safest choice for customer-facing work. A new benchmark of five language models on real support tickets gives that assumption little support. The top-scoring model was statistically tied with one that costs far less and answers faster. ## The central result Researchers gave five models, labeled North, Vale, Ember, Kite and Harbor, the same 200 anonymized customer-support tickets. North had the highest average score at 84.2. Vale was a hair behind at 83.8. The researchers report that this gap was **not statistically significant**, so the test cannot separate the two on quality. The cost and speed figures differ much more. North's estimated cost was $1.90 per 1,000 tickets, against $0.74 for Vale, so Vale was roughly 60% cheaper. Vale also answered 18% faster. If two models are indistinguishable on quality, a large gap in price and speed becomes the more useful basis for a decision. This does not prove the models are equal. With 200 tickets, a small real difference could go undetected. But the benchmark offers no evidence that North's extra cost bought a measurable improvement. ## How the test worked Each model wrote one proposed answer for each ticket. Three experienced support managers then graded the answers blindly, meaning they did not know which model wrote which response. Each score combined three things: factual correctness, usefulness and tone. Blind grading by experienced practitioners is a strength. The graders judged answers as people who handle these tickets would, rather than relying on an automated metric. Combining three dimensions into one number also has a cost, because a single score can hide tradeoffs. A response that is friendly but slightly wrong could score the same as one that is correct but curt. ## Where the lower-priced models fall The remaining models trail the top two, roughly in line with their prices. Ember scored 79.1 at $0.31 per 1,000 tickets. Kite scored 76.4 at $0.22. Harbor, the cheapest at $0.08, scored 71.0, about 13 points below North. The summary does not say whether the gaps among the lower tiers, or between them and the leaders, were statistically significant, so they should be read cautiously. Still, the overall pattern is plausible: spending more tends to buy better answers until the top of the range, where the returns appear to flatten. Whether a cheaper tier is acceptable depends on the use case. A team drafting replies for human agents to edit may tolerate a lower score. A team sending answers directly to customers probably has less room. ## Different strengths, with caveats The two leaders did not excel at the same things. North performed best on billing disputes, while Vale performed best on technical troubleshooting. That suggests a team's ticket mix could matter. A company whose queue is dominated by billing issues might reasonably weigh North differently from one handling mostly technical problems. The article should not overstate this. The benchmark does not report how large these category differences were, and splitting 200 tickets into categories leaves fewer examples in each. Treat them as hints worth checking on your own data, not as firm guidance. ## A weakness shared by every model The most consistent finding is a shared failure: **all five models performed poorly when a ticket depended on an image attachment.** Screenshots of error messages, billing statements and misconfigured settings are routine in support. Whatever their price or overall score, none of these models handled such tickets reliably. For anyone deploying these systems, that points to a practical safeguard. Tickets with attachments probably need routing to a human, or at least closer review, regardless of which model is chosen. ## What the benchmark does not cover The limits of the test matter as much as its results: - **One company, one language.** All tickets were English-language and came from a single software company. Results may differ for other industries, products, customer populations or writing styles. - **No multilingual testing.** The benchmark says nothing about how these models handle support in other languages. - **No agentic actions or tool use.** The models only proposed answers. They did not issue refunds, look up accounts or take other actions, which many real deployments require. - **Estimated costs.** The prices are estimates per 1,000 tickets, and real spending will depend on ticket length, usage patterns and pricing changes. ## The takeaway The benchmark's clearest message is that the best average score isn't automatically the best buy. North and Vale were statistically tied on quality, yet Vale was substantially cheaper and faster. For organizations choosing a model for English-language text responses, that makes a strong case for testing the second-ranked option before paying for the first. The conclusion is narrow, though. It applies to proposed written answers in one company's tickets, and it leaves image-dependent requests, other languages and agentic workflows untested. Teams should run a similar blind comparison on their own tickets before committing.

Qwen: Qwen3.8 Max (0902): # HelioDock Launches Night Shift Remote Monitoring for Warehouse Robots *The software arrives after a 14-month private beta and coincides with a $24 million Series B led by North Coast Ventures.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. Night Shift is aimed at the moment when a robot needs attention and the person responsible for it is not on site. HelioDock says the software gives operators a remote view of fleet activity, allowing them to see where robots have stopped or encountered a problem without traveling to the warehouse floor. The company says beta customers experienced 17% fewer unplanned robot stoppages. The figure comes from HelioDock and has not been independently verified. The 17% figure is a company-supplied performance claim from the beta period. HelioDock released it as part of the Night Shift announcement. The launch follows 14 months of private beta testing with nine logistics companies and 310 robots. HelioDock said the beta gave operators a way to monitor fleets remotely and produced fewer unplanned stoppages among participating customers. Chief Executive Mara Velez framed the product around after-hours response. “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. The quote identifies the problem HelioDock is targeting: a robot may stop because of a pallet, and a manager may need to travel to the site outside normal hours. Night Shift is intended to let an operator monitor that situation remotely. If the operator can view the fleet through the software, the response may not require a trip to the facility. Night Shift is priced as a recurring subscription tied to each robot. At $90 per robot per month, the cost increases as the fleet grows. The software becomes generally available Sept. 15. It enters a category where fleet-management products are already available from Relay Robotics and YardOS. Beta feedback showed that the software’s usefulness depends on network connectivity. Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. Strong connectivity allows operators to use the remote interface. LTE outages in rural facilities reduce the system’s usefulness because remote control is unavailable during those failures. HelioDock said robots return to their locally configured behavior when connectivity fails. Remote control becomes unavailable. The robots continue operating under their local configuration, but the operator loses the remote view that Night Shift provides. For warehouses with reliable connections, HelioDock is presenting Night Shift as a way to reduce the need for physical checks across a robot fleet. For sites with weaker wireless coverage, the beta feedback points to a narrower benefit. The software can help while the network is working, but its remote function stops when the connection does. The Series B was led by North Coast Ventures and brings HelioDock’s total funding to $39 million. The company announced the financing on Thursday, the same day it launched Night Shift. Relay Robotics and YardOS offer competing fleet-management products.

Anthropic: Claude Sonnet 5.5: # HelioDock launches remote monitoring software for warehouse robot fleets *Night Shift, tested for 14 months with nine logistics companies, arrives alongside a $24 million Series B.* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures. Night Shift is meant to reduce the number of times a person has to go to a warehouse because a robot has stopped working. The software will cost $90 per robot per month and become generally available September 15. It enters a market where Relay Robotics and YardOS already sell competing fleet-management products. Its usefulness, according to one beta customer, depends heavily on the quality of a facility's internet connection. ## Beta results Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages. The company supplied the figure, and it has not been independently verified. CEO Mara Velez described the problem the software is intended to address. "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. ## Where connectivity falls short Aaron Pike of Fulcrum Logistics, one of the beta customers, said Night Shift worked well in facilities with strong connectivity. It became less useful during rural LTE outages, he said. That follows from how the system is designed. When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. Operators who rely on Night Shift to step in from off-site therefore lose that ability when the connection drops, though the robots keep operating under the settings already configured on them. ## Pricing and competition At $90 per robot per month, a 100-robot fleet would cost $9,000 a month. The notes provided to this article do not say whether HelioDock offers volume discounts or whether the software works with robots from other manufacturers. Nor do they say how Night Shift's features or pricing compare with those of Relay Robotics or YardOS. ## Funding The Series B brings HelioDock's total funding to $39 million. The company did not say in the materials provided how it plans to divide the new money among product development, hiring and other areas.

Qwen: Qwen3.8 Max (0902): # 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 Sonnet 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.

Qwen: Qwen3.8 Max (0902): 6

Anthropic: Claude Sonnet 5.5: Going through the draft line by line: 1. `#Flood software moves upstream` violates the headline rule, because `# Headline` needs a space after the `#`. (1) 2. `The company is expanding its pilot.` is meant to be the dek, but it is not italicized. (2) 3. `##Details` violates the section heading rule, since it has no space after the `##`. (3) 4. `Read [the announcement](www.example.com/news).` has a link without an absolute HTTPS URL. (4) 5. `* The first pilot begins Monday.` uses an asterisk where a hyphen is required. (5) 6. `<div>The second pilot begins Friday.</div>` is raw HTML. (6) Every violating line counts once, and the blank lines have no violations. **6**

Qwen: Qwen3.8 Max (0902): # Cloudnote Debugging Page Exposed Customer Prompt Excerpts for Four Hours *Cloudnote says a temporary debugging page leaked short excerpts from 1,842 user sessions, though it has not yet determined whether anyone other than a security researcher viewed it.* Cloudnote said Monday that it discovered a debugging page containing excerpts from customer prompts was accessible without authentication for four hours. The company said it removed the page 19 minutes after a researcher notified it. Logs show the page included excerpts from 1,842 user sessions. Each excerpt could contain as many as 200 characters from a prompt, meaning the exposed material may reveal fragments of user queries but not full files. Cloudnote has 3.2 million registered users. The company said passwords, payment information and full documents were stored elsewhere and were not present on the page. It also said it has found no evidence that anyone besides the researcher accessed the page, but it has not completed its investigation. The incident highlights the privacy stakes for prompt-handling services, where even truncated prompts can include workplace plans, legal questions, medical details or other sensitive material. Cloudnote has not said whether the page was indexed, cached or viewed by another party. It also has not said whether affected users will receive notice, what additional steps may follow, or when the investigation will ultimately conclude.

Anthropic: Claude Sonnet 5.5: # Cloudnote Debugging Page Exposed Snippets of User Prompts for Four Hours *The company says only the researcher who reported it is known to have viewed the page, but its investigation is not finished.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, meaning no login was required to view it. Logs show the page displayed excerpts from 1,842 user sessions, each containing up to 200 characters of a prompt. Cloudnote has 3.2 million registered users, but it is not yet clear how many individual people those sessions represent or what the excerpts contained. The company removed the page 19 minutes after a security researcher notified it. Cloudnote says it has found no evidence that anyone other than the researcher accessed the page. It has not completed its investigation, so that finding could change. Cloudnote says full documents, passwords and payment information were stored elsewhere and were not on the page. That limits the exposure, but even short prompt excerpts can include personal or professional details, depending on what users typed. Cloudnote has not yet explained how the page became reachable without a login, or how long it was exposed before the researcher's report. Whether its review of access records holds up as the investigation continues will determine how far the incident reaches.

Qwen: Qwen3.8 Max (0902): # KiteBridge Update Meant to Streamline Expenses Instead Created $2.8 Million in Duplicate Payments When KiteBridge released a routine update to its expense-management platform on Tuesday morning, the intended result was smoother reimbursements for corporate clients. Instead, the failure turned ordinary employee expense claims into a cascade of duplicate payments, leaving finance teams to untangle 13,400 mistaken reimbursements worth $2.8 million. KiteBridge, which builds software companies use to manage employee expenses and reimbursements, deployed version 4.2 at 8:45 a.m. Tuesday. At 9:12 a.m., the company received its first customer complaint. By midmorning, support teams realized that the problem was not merely a delay. A payment processor had timed out, meaning the outside service responsible for moving money did not respond quickly enough. KiteBridge’s fallback system then tried to recover by retrying the affected payments. But because of a flaw in the update, those retries were not recognized as second attempts at the same transaction. The software treated them as new reimbursement requests and sent money again. The damage spread quickly across customers that relied on KiteBridge’s automated payment feature. Ninety-one corporate customers were affected, and the duplicate payments were drawn from employer-funded accounts rather than employees’ personal bank accounts. According to the company, no employees had money withdrawn from their accounts; the employers bore the cost of the extra transfers. KiteBridge suspended automated payments at 10:03 a.m., cutting off the system before more duplicate transactions could be issued. Yet the company’s public explanation lagged behind the technical severity of the incident. Its public incident page initially described the problem as delayed reimbursements. Only about six hours after automation had been suspended did the page disclose that duplicate payments had occurred, giving customers a clearer picture of the cleanup ahead. The episode is especially embarrassing because KiteBridge had previously advertised safeguards designed to prevent this exact failure. In an April changelog, the company said version 4.0 had introduced “exactly-once payment protection,” a technical control meant to ensure that a reimbursement is paid one time even if a payment system retries a request after a timeout. Version 4.2 was supposed to build on that foundation, but the safeguard failed under real-world conditions when the payment processor stopped responding as expected. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the duplicate payments. The remaining funds were still pending, and affected companies faced reconciliation work across payroll, accounting and expense systems. For many customers, the immediate task was identifying which employees or departments had received extra funds and arranging returns or adjustments without disrupting regular business operations. Chief Executive Lena Ortiz apologized for the incident and said the company would publish a technical report explaining what went wrong. For customers, that report will need to answer two basic questions: how a feature intended to prevent duplicate payments failed, and why the public description of the problem did not match what was happening inside the system. Until those answers are provided, KiteBridge’s recovery effort is about more than retrieving money. It is also about rebuilding trust with the finance teams that depend on its software to handle sensitive payments accurately.

Anthropic: Claude Sonnet 5.5: # Expense-Software Glitch Sends $2.8 Million in Duplicate Reimbursements to 91 Companies A software failure at KiteBridge, which makes expense-management software, generated 13,400 duplicate reimbursements worth $2.8 million on Tuesday. The payments were funded by the 91 corporate customers affected, not by their employees. By Wednesday evening, $2.3 million had been recovered, and the remaining $500,000 was still pending. ## How it unfolded KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m. According to the reporting, the problem began with a timeout at a payment processor, the outside service that moves money between accounts. A timeout means the processor did not answer in time, leaving KiteBridge's software unsure whether a payment had gone through. In that situation, KiteBridge's fallback system, the backup logic meant to keep payments moving when something goes wrong, retried the payments. But it treated each retry as a new reimbursement request rather than a repeat of one already submitted. Each retry therefore produced another payment, and the software generated thousands of duplicates. KiteBridge suspended automated payments at 10:03 a.m., less than an hour and a half after the deployment. By then, 13,400 duplicate reimbursements had been issued. ## Who paid No employees had money withdrawn from their accounts. The employers funded the duplicate payments, so the companies, not individual workers, bore the immediate financial exposure. By Wednesday evening, $2.3 million of the $2.8 million, about 82 percent, had been recovered. ## A feature promised, a problem described differently The failure drew attention in part because of what KiteBridge had said about its own product. An April changelog, the company's public list of software updates, said version 4.0 had introduced "exactly-once payment protection." That phrase generally refers to safeguards intended to ensure a payment is made one time, even if a request is repeated. The reporting does not say why those protections did not prevent Tuesday's duplicates. KiteBridge's public incident page also drew scrutiny. It initially described the problem as delayed reimbursements, which did not match what was happening: duplicate money was going out. The page first disclosed the duplicate payments six hours after automation was suspended. ## The company's response KiteBridge CEO Lena Ortiz apologized and said the company would publish a technical report on the incident. Several questions remain unanswered in the available reporting. They include how the retry behavior got past the protections described in the April changelog, whether the Tuesday deployment played a role, and when the remaining $500,000 will be recovered. KiteBridge's promised report is likely to be the first full account of what went wrong. For the 91 affected companies, the immediate task is reconciling their accounts and getting back the money still outstanding.

Qwen: Qwen3.8 Max (0902): # The Last Program at Foundry Hall By the end of October, the front doors of Foundry Hall will lock for the last time. The Milwaukee nonprofit startup accelerator, which opened in 2018 as a home for early-stage founders, will close Oct. 31 after losing the sponsorship that held together much of its budget. The closure does not arrive with a dramatic crash, but with a calendar. Executive Director Camille Ross said the organization has enough money to operate through October, but not to fund another full program. Six employees will lose their jobs. Twelve startups that have been using the building’s workspace will have to find somewhere else to work. Foundry Hall’s numbers tell a larger story than its empty rooms. Since opening, it has supported 212 startups, which together employ 1,140 people. Those companies became part of a local network that gave new businesses a place to meet investors, borrow advice, and learn from one another. For many founders, the Hall was less an office than a first address. Dev Malik remembered the building as the place where his company became real. He said his first investor meeting took place there, after years of pitching from the back table of a coffee shop. That sense of shelter was fragile. Blank River Bank, Foundry Hall’s largest corporate sponsor, supplied 38% of the annual budget. When the bank decided not to renew its sponsorship, the organization began looking for a replacement. Ross put it plainly: “We kept waiting for the replacement check that never came.” The bank, for its part, said the decision was tied to a shift in community investment priorities and was unrelated to Foundry Hall’s performance. That distinction may matter to the institution’s public record, but it offers little comfort to the founders and staff who built their routines around the Hall’s conference rooms, event nights, and shared desks. City officials are now trying to soften the landing. Deputy Mayor Susan Cole said Milwaukee is talking with the building’s owners and local universities about possible space for the companies currently inside. Whether those conversations will produce leases, lab space, or temporary corners is not yet clear. For the startups still packing laptops and prototypes into bags each evening, the closing is practical before it is symbolic. They need Wi‑Fi, mail, and rooms where they can meet customers. They also lose a place where being early did not mean being alone. Foundry Hall will survive in the payroll of the companies it helped shape and in the founders who learned to pitch there. But on Halloween, the lights will go out, and Milwaukee will have one fewer room where someone’s first idea could become a business.

Anthropic: Claude Sonnet 5.5: # Foundry Hall, Milwaukee Startup Accelerator, to Close Oct. 31 Foundry Hall, the Milwaukee nonprofit accelerator that has supported 212 startups since opening in 2018, will close October 31 after its largest corporate sponsor declined to renew its funding. Blank River Bank had supplied 38% of the organization's annual budget. Without that money, Foundry Hall has enough to operate through October but not to run another full program. Six employees will lose their jobs, and the twelve startups now working from its space will have to leave. A Blank River spokesperson said the decision reflects a change in the bank's community investment priorities and is unrelated to Foundry Hall's performance. Executive director Camille Ross said the organization had hoped to replace the money. "We kept waiting for the replacement check that never came." The companies that passed through Foundry Hall now employ 1,140 people between them. For some founders, the building mattered long before those jobs existed. Dev Malik remembers it as the place where his company first looked like a company. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Malik said he has been thinking less about the program than about the founders who would have come next, the ones who will start out at the coffee shop table without a place like Foundry Hall to move into. The most immediate problem is the twelve companies in the building. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about space for them. She did not say whether a location had been identified, and the October 31 deadline leaves little time to find one. For the six staff members, the timeline is just as short. Ross did not describe their next steps, and the organization has not announced any plans to continue its work in another form. What Foundry Hall leaves behind is harder to count than the money it lost. There are 212 startups, some thriving and some not, and 1,140 people who go to work at companies that, in many cases, got their start there. On November 1, the building will no longer be a place where a founder can walk in with an unfinished idea and a borrowed slide deck. The companies will remain, scattered across the city, in whatever space they can find.

Qwen: Qwen3.8 Max (0902): # Arcwell Launches Relay, an AI Platform for Support Agents Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. The product places Arcwell in the broader customer service software market, where AI tools are increasingly being aimed at everyday support operations and the workflows around them. Relay is built around three functions. It suggests answers to agents, identifies customers who may cancel and creates summaries of support conversations. Those capabilities demonstrate how artificial intelligence continues to transform customer service, moving from high-level promises into the practical work of answering questions and tracking account risk. The company said Relay was tested by 12 companies over six months. Arcwell said agents using the platform closed tickets 14% faster. It did not release customer-retention results, leaving the effect of its cancellation alerts undisclosed. For buyers, that may raise a straightforward question: whether the faster ticket times and churn signals translate into measurable business gains. Relay costs $65 per agent each month. Priya Shah, Arcwell’s CEO, said the product addresses a familiar inefficiency inside support teams. “Support agents spend too much time searching for information that their companies already have,” Shah said. The platform also carries operational requirements that customers will have to weigh before adopting a tool that reads customer conversations. Relay needs access to support transcripts and internal help documents to produce its suggestions and summaries. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose either a 30-day or 90-day retention period, a control that may matter to organizations reviewing how long sensitive support records remain available. For Arcwell, Relay is an important new chapter as the company seeks to help companies improve customer service using artificial intelligence. Its pitch combines productivity gains with data controls, but the absence of retention results will likely shape how closely the product is scrutinized by potential customers.

Anthropic: Claude Sonnet 5.5: # Arcwell Launches Relay, an AI Tool That Feeds Answers to Support Agents Arcwell has launched Relay, a platform that uses artificial intelligence to analyze customer support conversations and suggest answers to agents as they work. The product is the company's bid to help businesses improve customer service with AI. Relay does more than propose replies. It can also identify customers who may cancel and create summaries of support interactions. The pitch starts with a familiar frustration. "Support agents spend too much time searching for information that their companies already have," said Arcwell CEO Priya Shah. Arcwell has early numbers to back the idea, though they are limited. Twelve companies tested Relay over six months, and the company says agents using it closed tickets 14% faster. That figure comes from Arcwell, not an independent evaluation. The company did not release customer-retention results. That leaves a gap, since spotting customers who may cancel is one of Relay's selling points, and Arcwell has not said whether the feature helped companies keep them. Relay costs $65 per agent each month. The tool also requires access to sensitive material. To work, Relay needs support transcripts and internal help documents, which together hold both customer conversations and a company's own guidance for handling them. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a retention period of either 30 or 90 days. For now, the case for Relay rests on speed. Whether it also helps companies hold on to customers is a question Arcwell has yet to answer publicly.

Qwen: Qwen3.8 Max (0902): # QueueZero’s AI Meeting Assistant Turns Remark Into Tasks **SAN FRANCISCO** — QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant that can attend meetings on a user’s behalf, produce notes and assign follow-up work to software agents. The company says the product is designed to reduce the administrative drag of meetings and keep projects moving after the call ends. But the launch itself offered an early lesson in how literal such systems can be. During a 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 rather than a casual aside. It created two tasks and assigned them to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The moment briefly turned the launch into a live example of the problem Proxy is meant to solve — and the complications that can arise when software treats spoken language as executable work. QueueZero said Proxy requires confirmation before any external message is sent. Internal tasks, however, can be created automatically, allowing the assistant to generate to-do items, route follow-ups and trigger agent-based workflows without waiting for a human to approve each step. Administrators can disable automatic task creation, a setting that is likely to appeal to organizations wary of unexpected assignments appearing in employees’ queues. The incident underscores a broader challenge for AI agents: deciding when a comment is a request, when it is a joke and when it is simply part of the background noise of a meeting. Vendors in the rapidly growing assistant market are betting that workers want software capable of turning talk into action, but the technology also raises questions about oversight, accountability and how much authority users are willing to hand over to automated systems. Proxy is priced at $30 per user per month, placing it in the midrange of productivity tools aimed at teams that spend much of their day in calls and status updates. The product enters public beta next week, giving QueueZero a chance to test how the assistant behaves in ordinary workplace settings rather than a controlled stage presentation. For QueueZero, the episode may become a useful proof point: Proxy understood the room, even if it needed a chief executive to resolve one of the tasks by hand.

Anthropic: Claude Sonnet 5.5: # QueueZero Launches Proxy, an AI Meeting Assistant That Assigns Work to Software Agents QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. The product enters public beta next week and costs $30 per user per month. ## A live demonstration, and an unplanned test During the launch 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. 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. It was, in the narrowest sense, a faithful reading of the sentence. ## Where the guardrails are The episode points to a central design question for assistants that act on what they hear: how to separate a passing remark from a directive. QueueZero says Proxy is built with limits on what it can do without human approval. According to the company, Proxy requires confirmation before any external message is sent. That means a task like the pricing request would not reach a customer until a person approved it. Internal tasks, such as those assigned to a colleague or an executive, can be created automatically. Administrators can disable automatic task creation, which would require a person to review each item before it enters a team's workflow. QueueZero did not say in the reporting available for this story how Proxy distinguishes requests from casual commentary, or how often it misreads them. Those details will matter as customers begin using the product in real meetings, where offhand remarks, jokes and disagreements are common. ## Why the distinction matters Meeting tools that only transcribe or summarize carry limited risk, since a person reads the output before acting on it. Proxy goes a step further by turning conversation into assigned work and passing some of that work to software agents. A misread sentence can then become a task in someone's queue, or potentially an action taken on their behalf. The confirmation requirement for external messages and the administrator controls are meant to contain that risk. Whether those defaults are enough will likely depend on how organizations configure them. A company that leaves automatic task creation on is accepting more cleanup in exchange for less manual entry. One that turns it off gets a more cautious tool that still produces notes and proposed tasks. ## Pricing and availability Proxy costs $30 per user per month. The public beta begins next week. In the demonstration, the system's mistake was harmless and was corrected within seconds by the person it affected. The beta will show whether that holds when the meeting is less friendly and the task is not a joke.