Five emerging AI tools from the September 2026 launch cycle, each changing a different working system: agent-native CRM, on-device models, AI observability, computer vision, and AI customer support.
Most launch coverage in 2026 keeps pointing at the same place: another chat assistant wrapped around a frontier model. The five products in this roundup were picked for the opposite reason. Each one changes a different working system — the records a sales team keeps, where model inference runs, what developers can see in production, how a camera feed becomes a decision, and whether a support ticket actually gets closed.
None of them is another chatbot or image generator. They were selected from the September 2026 launch and discovery cycle, and every claim below was checked against the vendor's official site, documentation, and pricing pages before publication. A September 10 launch page can already be out of date by September 20, so newer versions are flagged where it matters and pricing is presented as the published snapshot, not a promise.
This roundup of new AI tools skips the chat clones on purpose. The test that matters is simple: does the tool slot into a real workflow, or does it just write a clever reply? By that test, these five earn their place — for different teams, and for different reasons.
The pattern under the current launch noise is easy to miss because most new AI tools never leave the chat window. The five below do. Records update themselves, models run on a phone, dashboards trace an agent's decisions, cameras answer in plain language, and support systems resolve instead of escalate. Almost every mainstream story this month — giant personal agents, frontier model launches — points back at the same consumer chat experience. These five point somewhere else.
They share one instinct: meet AI where the work already happens. The useful test for new AI tools worth your time is whether they slot into a real workflow, not whether they can write a clever reply.
They also share a quieter theme in how they bill. None charges per chat message. The selected products meter per credit, per gigabyte, per resolution, per seat — or not at all, in Desert Ant Labs' case. That pricing shape is itself a signal: these tools are built for work that happens inside systems, where volume is predictable enough to meter.
Lightfield is the CRM entry in this group, and the easiest to misread: it is not a chatbot bolted onto a contact database. The company calls the product agent-native, and it means that as an engineering stance, not a label. Lightfield captures the stream of customer conversations — email, calendar, calls, meeting recordings — keeps them in a working context graph, and lets AI agents act on that context directly.
The difference is small to describe, large to feel. A traditional CRM is a database you feed by hand; Lightfield updates itself. Agents refresh records, draft follow-ups, prep meeting briefs, and surface deal advice cited back to the exact moment in a call. It connects Gmail, Outlook, Slack, and LinkedIn, auto-joins Zoom, Meet, and Teams meetings, and can backfill up to two years of email and calendar history — chunks of the stack that HubSpot AI and Apollo.io sell as separate products.
If the pain is records that never get updated and deals that stall for lack of follow-through, this pick is worth a look before you renew a smart-CRM contract. Most new AI tools in the CRM category layer AI onto a database; this one rebuilds the database around AI. Lightfield's own documentation describes the automation surface plainly: you do not open a node editor or write conditionals; you describe the outcome and the agent programs the trigger and steps for you. A natural-language agent builder sits alongside an API with Python, TypeScript, and Go SDKs, a CLI, and MCP access in public beta — though the deeper agent features (automations, sequences, workspace credits) live on Pro and above.
It is new, and the pricing model has two moving parts. Plans run $89 per user per month for Starter (up to three seats, billed monthly), $999 per workspace per month for Pro, and $1,999 for Growth (both billed annually), with Enterprise handling SSO and role-based access at custom pricing, plus a 14-day trial on Starter. Agentic work — workflows, automations, sequences — draws from workspace credits the vendor meters per use, so the bill tracks how much work you hand the agents, not just the seat count; Starter includes no credits.
On security claims, the company states it is SOC 2 Type II compliant and publishes GDPR guidance, and the documentation is unambiguous about training: it says models are never trained on customer data. ISO 27001 is listed as "soon" rather than delivered, so treat it as roadmap. Momentum is real but recent: a $47M Series A led by a16z (Andreessen Horowitz) announced on September 9, 2026, and more than 5,000 companies have signed up since the November 2025 launch. Full detail in our Lightfield review; wider context in AI agents for business.
Desert Ant Labs sits at the deliberately small end of the field. Instead of sending every task to a frontier cloud model, this European lab ships small, specialized models that run on-device inside your product. Its September 8, 2026 launch brought 18 models — 12 stable, 6 in beta — behind one SDK for Swift, Kotlin, and JavaScript, running through Core ML on Apple devices, LiteRT on Android, and WebAssembly in the browser, with weights on Hugging Face and a CLI on GitHub.
Some jobs care more about specialization, latency, and repeated on-device runs than raw generality. Clear, at about 9MB, pulls a noisy five-minute recording to studio quality in about a second. Redact strips PII across 27 languages in a 12MB model. Gist tags content topics across 101 languages. Tongue spots 84 languages from three words in a 2MB package. Voz transcribes ten minutes of audio in about two seconds on an iPhone — currently the Apple-only model in the catalog. That is a different compute posture than paying per token for everything.
Mobile and product developers who need audio, text, or vision capabilities behind a hard privacy or latency requirement. Finished cloud alternatives — ElevenLabs, Descript, Adobe Podcast, and meeting noise cleanup via Krisp — solve similar problems the productized way. For a developer it is also the cheapest tool in this article to trial, because it is free until an app passes 100,000 active devices. Among the new AI tools here, it is the only one that gives you raw on-device inference under your own control.
The honest tradeoff is in the setup: you get the model layer, and the integration, tuning, and quality work are yours. The catalog is also young. The free tier is the headline and it is unusual — free up to 100,000 monthly active devices per SDK, no tokens, no logins, no per-run limit below the cap, with pricing above it unpublished. Privacy is the second headline: the lab states data never leaves the device, and "what's never been uploaded can never be compelled" is a real argument — but a claim to evaluate, not a guarantee. We found no independent telemetry audit. One signal worth noting since launch: the lab's own video app Detail 6 now runs entirely on these models, which is the closest thing to a public production reference. Founder Paul Veugen built Detail, Apple's 2025 iPad App of the Year, so the mobile pedigree is genuine. Our Desert Ant Labs review covers the full catalog.
Most roundups of new AI tools ignore what happens after you ship an agent. When a system makes hundreds of tool calls per request, "the model responded" stops being a useful health signal. AI Observability by OpenObserve answers the harder question: what were the agent, tools, model, latency, cost, and quality actually doing?
A naming note first. OpenObserve is a broader observability platform — logs, metrics, traces, frontend monitoring — and AI and LLM monitoring is this collection's observability pick. It launched on Product Hunt on September 10, 2026 under the name AI Observability by OpenObserve, and a day later the project shipped v1.0.0 with AI Observability v1.0, moving the feature set to general availability. It is best read as the AI layer of an existing platform, not a separate product.
It is OpenTelemetry-native, which gives it the deepest instrumentation story in this batch. Standard gen_ai spans and OpenInference conventions just route in, so instrumentation you already have — LangChain, CrewAI, LlamaIndex, OpenAI, Anthropic, LiteLLM, plus gateways and no-code builders, 80+ integrations — keeps working. The Agent Graph maps each agent to the models, tools, services, and datastores it calls, with request counts and error health on every edge, so a runaway loop or a failing tool is obvious at a glance. Session views break cost, duration, and tokens down per turn, then link to the full trace. Quality is scored in production: online evaluations grade live spans, traces, or sessions as they arrive, using your own provider as judge or a remote scorer, with built-in scorers for relevance, hallucination, toxicity, and bias. One caveat: online evaluations are an enterprise capability in OpenObserve, not part of every deployment.
It is early operational tooling — you own instrumentation and thresholds, and it targets infrastructure teams, not an office install. Pricing is per GB ingested and queried, not per span or per seat, which stays predictable when agents fan out into hundreds of spans per request. Users are unlimited, and deployment spans managed cloud, self-hosted single binary, BYOC, and BYOB, with a free self-hosted tier to 50GB. It is also the only pick here built on an open source core (AGPL-3.0) that now sits at roughly 22,000 GitHub stars, with SOC 2 Type II and ISO 27001 listed. Context in AI agent frameworks compared and AI agent security.
Computer vision used to be a project: datasets, labeling, model selection, training, deployment. Viso Now, the computer-vision slot in this group, makes it a conversation instead. You type what should be watched in plain language, or drop in a clip, and the platform assembles the workflow — input feed, detection logic, alerts — and runs it in front of you.
The company markets the underlying perception engine as Visual General Intelligence (VGI-1), described on the product page as one pre-trained model with general visual reasoning, while a company FAQ describes a multi-model orchestration architecture that routes to frontier models under a proprietary composite layer. Either way, the workflow promise is the same: no datasets, no labeling, no model selection — you describe the outcome, then tune rules, thresholds, and outputs before publishing to live feeds.
It sits between the cameras you already run and the tools your team uses. Sources include IP and CCTV cameras over RTSP and ONVIF, NVR systems, uploaded files, and cloud folder sync from AWS, Azure, or GCP. Outputs push to Slack, Microsoft Teams, email, or any system over the API. Templates cover loading dock exclusion zones, PPE checks, restricted-area vehicle alerts, and wait-time tracking, so operations, quality, safety, and integrator teams start from a proven base instead of a blank canvas — a working vision answer in hours, not a research project.
On accuracy we are deliberate: viso.ai publishes no public benchmark figures for its vision engine, so we will not invent any. The workflow promise — no datasets, no labeling, no model selection — is what the product actually sells, and what you buy is the platform and its policies, not a guaranteed detection rate. Pricing is credit-based and shared across unlimited users: free with 5 daily credits (up to 30 per month), Pro $25/month for 100 credits, Business $50/month adding SSO/SAML and role-based access, Enterprise quoted (details on the Viso Now pricing page). SOC 2, ISO 27001, GDPR, and AI Act compliance sit in the Suite security docs, and the company states it does not train models on customer video or image data. See also AI tools for small business 2026 and best AI tools of 2026.
Typewise Nova is the customer-support entry in this group, and the naming needs a second look. Typewise started in 2019 as a keyboard app, closed that product in 2022, joined Y Combinator, and moved to business software. Today it is solely an AI customer experience platform, and Nova is its AI operator on top.
The core idea is that support is an operational system, not a chatbot in front of an FAQ. Typewise agents resolve requests end to end — look up the order, apply the policy, close the ticket — while a supervisor routes each request to the right specialist, brings in a person when judgment is needed, and picks the ticket back up.
Nova launched publicly in September 2026 and is the fresh piece of this stack. You set it up by talking to the system: describe the desired behavior in plain language, and Nova tests and publishes the result — a billing specialist that reads Stripe, a returns-policy update, a new WhatsApp channel — in the company's words, in about 15 minutes with no developer. It also watches operations, flagging a CSAT slip or a refund spike after a product change and offering to draft a reply macro.
Resolution is the metric here, not deflection. A full resolution counts 1.0, a partial hand-off 0.5, and an unresolved ticket is free. Pricing is a base plan plus a per-resolution rate that drops with volume: Starter $99/month, Growth $599, Business $2,000/month billed annually, Enterprise custom (all on the Typewise pricing page). Resolutions run $2 pay-as-you-go down to about $1 with volume packages, and the trial includes 100 resolution credits plus $10 in AI Operator credits without a card.
Coverage is the depth play: chat, email, WhatsApp, social, voice, and in-app, in any language, over 3,500+ connectors the company describes as MCP-native — our MCP server explainer covers that term. When teams evaluate new AI tools by outcomes, per-resolution billing is the clearest signal here: the vendor only gets paid when a ticket is actually resolved. Vendor figures like a 90% autonomous resolution rate and 10M+ tickets resolved come from case studies and site statistics, so read them as marketing evidence, not independent evaluation. Security claims are substantial — ISO 27001 certified, GDPR compliant, aligned with the EU AI Act, EU data residency — and worth confirming for your needs. Related reading: AI tools for small business 2026.
These five new AI tools span different industries and different lanes — the point of grouping them is the direction they confirm, not the competition between them. The comparison below is factual, not a ranking.
| Tool | Core role | What it changes | Who owns the work |
|---|---|---|---|
| Lightfield | Sales engagement | Record-keeping and follow-up | Reps and sales ops |
| Desert Ant Labs | On-device inference | Where compute runs | Mobile and product developers |
| AI Observability by OpenObserve | Agent production visibility | How teams debug AI systems | SRE and platform teams |
| Viso Now | Agentic computer vision | How vision workflows get built | Operations and safety teams |
| Typewise Nova | Customer operations | How tickets get resolved | Support and CX teams |
None of the 5 new AI tools competes with another for the same buyer, which is why the reader match in the next section matters more than "which is best." All five also share the pricing theme that separates this crop from last year's wrappers: none charges per chat message — they meter per credit, per GB, per resolution, per seat, or free per device.
Here is who each of these new AI tools is actually for.
Two cautions apply to all five new AI tools here. They are young: four are launch-week or launch-month products from early September 2026, so APIs and pricing may move fast. And the strongest claims — accuracy, resolution rates, certifications — come from vendors; we verified what we could against documentation and flagged what we could not.
Before picking any of them, check the parts that actually matter:
These products serve different jobs, and some are early enough that they may change quickly. Verify pricing and access close to the decision, keep security and permissions in mind when a tool touches production data, and pick for the workflow — not for the novelty.
Short answers to the questions people ask most often about these new AI tools, with pricing current as of the September 2026 snapshot.
What are the 5 new AI tools featured in this article? The five products featured are Lightfield, Desert Ant Labs, AI Observability by OpenObserve, Viso Now, and Typewise Nova. They were selected from the September 2026 launch and discovery cycle because each solves a different operational problem — CRM records, on-device models, agent monitoring, computer vision, and customer support — instead of adding another chatbot or image generator to the pile.
Is Lightfield an AI CRM? Yes. Lightfield is an agent-native CRM that captures emails, calls, and meetings, keeps a customer context graph, and uses AI agents to update records, draft follow-ups, and prepare for meetings. Pricing starts at $89 per user per month on Starter, with agent work metered through workspace credits on plans from Pro upward.
What is different about Desert Ant Labs? Desert Ant Labs builds small, specialized models that run on-device through one SDK for Swift, Kotlin, and JavaScript. The lab launched 18 models on September 8, 2026, and offers them free up to 100,000 monthly active devices per SDK, with pricing above that cap unpublished.
Why does AI observability matter for AI agents? Once agents start making tool calls and generating operational traffic, teams need to see what the agent, its tools, models, latency, cost, and quality were actually doing. AI Observability by OpenObserve provides this with OpenTelemetry-native tracing and online evaluations on live traffic.
What is Viso Now used for? Viso Now is an agentic computer vision platform. You describe what should be watched in plain language and the platform assembles the vision workflow, connects cameras, and pushes alerts to Slack, Teams, email, or any system over its API. It has a free plan and paid plans from $25 per month.
What is Typewise Nova? Typewise Nova is an AI operator inside Typewise's customer experience platform. You describe how your support should work in plain language and Nova builds, tests, and publishes the specialists and workflows, with plans from $99 per month plus per-resolution pricing.
Get our best AI tool reviews and guides delivered to your inbox each week. No spam, unsubscribe anytime.
Written by
Editorial Team
The AI Tools Vault editorial team researches, tests, and reviews the best AI tools across every category.
More reading on ai news
AI NewsAnthropic's September 16, 2026 update folds Claude Cowork and chat into one Claude and adds Claude Docs and Claude Slides in beta. What each tool does, who has them, exports, sharing, and the limits still in place.
AI Tools Vault Team
AI NewsGPT-6 Astra changes what it means to delegate computer work, not just query a model. Here is what it actually does, what it costs per task, the security record nobody is quoting fully, and where supervision still belongs.
AI Tools Vault Team
AI NewsThe AI agent news that matters this month: Cursor's OpenAI cut-off, the UK ICO's evidence call and Australia's automated-decision rules all land within seven weeks.
Jawad Fiaz