|
GTM stack intelligence, enriched.
|
|
Outreach’s Deal Agent can now write AI-recommended values straight into CRM picklist, numeric, and percentage fields, and it can do that without anyone clicking accept. Anthropic’s Microsoft 365 connector picked up tools that send email, move files, and create calendar events. Clari opened its Copilot data to outside agents over MCP, and ZoomInfo put its Context Graph on the command line so an agent can query it from a shell. Agents spent the last year reading your systems. This week the vendors let them write.
The other thread: both big labs ended the free ride on agent runs inside 48 hours of each other, and the independent metering vendor shipped daily AI credits the same week. That’s below.
|
|
|
Salesforce — Agentforce Commerce Hits GA With a Native ChatGPT Sync
Salesforce tends to get page space in every Enriched issue. It is not surprising, since it has a black-hole-like gravitational pull in the GTM stack. However, this week’s coverage is well deserved. Salesforce announced Agentforce Commerce on July 6th, its largest release in that line: Shopper, Buyer, and Merchant Agents, plus Storefront Next, all GA.
Shopper and Buyer Agents help your customers engage with your product catalog, ask questions, and place orders, while the Merchant Agent confirms inventory availability and shipping logistics in real time.
The release reaches beyond inbound customers: it includes a native ChatGPT integration that syncs a retailer’s product catalog into ChatGPT with “no extra software, no third-party tools.” A Google integration covering Search AI Mode and the Gemini app is stated for summer 2026. The shift from SEO towards GEO (Generative Engine Optimization) has been underway for quite a while now. Salesforce now has the data to back up the shift, “AI influenced 20% of global online sales, worth $262 billion,” and has built a significant product suite to help businesses capitalize on it.
The retailer stays merchant of record and orders stay on-platform, so the money path is unchanged. What changes is where discovery happens: your catalog becomes a feed into an assistant you do not own, do not instrument, and cannot A/B test.
Growth teams have a new paradigm to explore and new tools to master. Anyone who has spent a decade optimizing a storefront funnel is about to have a chunk of that funnel run inside a chat window that reports nothing back. Leveraging these agents will unlock capabilities, but customizing the user experience and garnering catalog insights will require a more nuanced stack setup.
Source: Salesforce Newsroom, July 6, 2026
|
|
|
|
AI for RevOps · Anthropic
|
Anthropic — Fable 5’s Included Access Was Extended Five Days, Then the Meter Starts
Fable 5 came back globally on July 1st, included up to 50% of weekly usage limits on paid plans through July 7th, with paid usage credits after that. On July 7th, hours before the cutoff, Anthropic extended the included allocation to July 12th at 11:59:59 PM PT on the same terms, after user backlash. At the risk of being the boy who cried wolf, Anthropic has extended Fable’s lifeline yet again, this time until July 19th.
Token cost has been the perennial hot topic in AI circles for some time now. Enterprises are leaning more and more on open-source models for cheaper token cost without much degradation in capabilities. Anthropic (and OpenAI) are keenly aware of the pricing pressure their models are facing. In that context, the repeated extension of Fable 5’s availability to existing paid plans makes sense: continue to allow innovators and early adopters to experiment on Fable 5 and embed those use cases into their workflows.
Source: Anthropic News / @claudeai
|
|
|
OpenAI — Workspace Agents Leave Free Preview, and In-ChatGPT Runs Now Draw Credits
The free preview for Workspace Agents on ChatGPT Business, Enterprise, Edu, and Teachers ended on Monday, July 6th, itself an extension of an earlier May 6th date. Agent runs invoked inside ChatGPT now draw down workspace credits. How much would this cost you? OpenAI cites a typical GPT-5.5 run at 5 to 25 credits. Runs invoked from outside ChatGPT, from a Slack channel for example, remain in free preview with no announced end date.
Cue one of Enriched’s primary pillars: usage-based pricing. Agent spend does not behave like seat spend. A run costs what it costs depending on how much context it drags in. This is where prompt engineering and experienced AI Ops teams can really get an ROI. The finance question shifts from headcount to consumption. The inside-versus-outside split is the other thing to watch: it is an arbitrage, teams will find it, and it will likely not survive contact with OpenAI’s next pricing page.
Source: OpenAI Help Center via secondary reporting, July 6, 2026
|
|
|
ZoomInfo — Puts the Context Graph on the Command Line, Free to Read and Metered to Enrich
ZoomInfo released the GTM.AI CLI on July 9th: an MIT-licensed, open-source command-line client that lets developers and AI agents search and enrich companies and contacts, pull intent signals, and run agentic research from a shell against the GTM Context Graph.
Search and lookup are free, but enrichment and agentic research draw bulk data credits against an active subscription. That split is an important design feature: reading the graph costs nothing, and the moment you pull a record into your own systems, the meter runs.
For anyone who has built enrichment workflows, this moves the work out of a seat and into a pipeline. Enrichment can now sit in a cron job, a CI step, or an agent loop, with no human clicking through a UI and no middle layer required for the simple paths. It also removes the natural rate limit on credit burn, which was a person getting bored. An agent in a loop can spend a quarter’s credits in an afternoon, so put a hard budget guard on the credential before you hand it to anything autonomous.
Source: BusinessWire via VentureBeat, July 9, 2026
|
|
A closer look at one shift in the stack each issue.
|
|
Stack Deep Dive · Outreach
|
Outreach — The Deal Agent Now Writes Into Your CRM Fields, and One Setting Removes the Human
Outreach’s July release gives Deal Agent the ability to write AI-recommended values directly into single-select picklist, numeric, and percentage fields. Two modes: one-click acceptance, where a rep approves the value, and automatic acceptance, where the value lands in the field on its own.
There is a spectrum of human-AI interaction within workflows: insights only → human-in-the-loop → fully agentic. Until now, agentic features in this category produced a suggestion and a human did the typing. Stage, next step, forecast category, close probability: every one of those values in your CRM got there because a person put it there, however carelessly. These fall squarely into human-in-the-loop, and they are high-impact fields where you likely don’t want to grant full agentic capability on the first go. Picklists, numbers, and percentages are constrained fields with a finite set of valid answers, which is where a model can be wrong but cannot be nonsense. Free-text fields, where a model would be most useful and most dangerous, are not in scope, yet.
The operational consequence sits in your field history, and it is one that I’ve dealt with personally at Confluent. Reporting on deal hygiene, field-update timestamps, or stage progression all assume a human made the change. With automatic acceptance on, that assumption is gone, and the reports will not tell you. Whether the write lands under a service identity you can filter on is something your RevOps and Data teams should prioritize. If you cannot answer that question, you cannot answer the next one, which is whether your forecast moved because a rep learned something or because a model inferred something.
The sane rollout is a phased one to specific users or teams, comparing accuracy against manually entered data over time. Automatic acceptance is a hygiene tool if the recommendations are good and a laundering machine for bad data if they are not. Don’t roll it out to the wider GTM org until you’re confident it’s the former.
Source: Outreach Support, July 9, 2026
|
|
|
|
Clari ships an MCP Server exposing Copilot data to outside agents: Per the July release page shared with Salesloft post-merger, GA July 7th, the Clari MCP Server lets external AI agents and tools connect to Clari Copilot data through a standard MCP interface, for enterprise Clari Copilot customers, with access arranged through your account team. Ada, Clari’s chatbot, is now in-app for all users. MCP is now the socket at the revenue-data layer, not just the CRM layer.
Source: Salesloft Champions, Clari section, GA July 7, 2026
|
|
Metronome ships daily-cadence recurring credits and commits: Per its July 7th changelog, Metronome now supports recurring credits and commits on a daily cadence, so a vendor can grant a daily allocation of AI or usage credits, configured through the UI or API for duration, applicable products, and prioritization. One daily recurring credit or commit per contract by default, and it cannot be combined with seat-specific credits. If you sell AI features and have been trying to build a daily included allowance on top of monthly billing primitives, that is now a config rather than a project.
Source: Metronome changelog, July 7, 2026
|
|
Alta raises $25M to sell an orchestration layer above your GTM stack: Alta, founded in 2023 by ex-monday.com operator Stav Levi-Neumark, announced a $25M Series A on July 8th led by IN Venture, with Mindset, Skywell, LeumiTech77, Entrée, Target Global, and Verissimo. It pitches an “AI System of Actions” that centralizes context across 50+ data sources and 60+ GTM tools, Salesforce, HubSpot, Attio, and Clay among them, through an ontology layer, then runs agents for prospecting, research, outbound, inbound qualification, and AI calling. Company-stated and unaudited: $1M revenue within months, 800% projected 2026 growth, with Snowflake, Atlassian, Deel, Riverside, and Sabio named as customers.
Source: PR Newswire via SiliconANGLE, July 8, 2026
|
|
Salesloft/Clari — July release tiers call-recording access and auto-rotates idle Salesforce tokens: Salesloft’s July release went GA on July 7th. One of the changes that might catch your team off guard: Salesforce refresh tokens that sit idle now auto-rotate, to meet Salesforce’s connected-app security requirements. Tokens that sit idle for 30 days will be rotated automatically. Scripts, syncs, or low-traffic integrations leaning on a long-lived token will break quietly (and maybe over a weekend).
Source: Salesloft Champions, GA July 7, 2026
|
|
|
👁️ Noticed
Anthropic says more than 90% of Cowork usage is not software development, and the leading categories are business operations and content creation: reconciling spend, building renewals trackers from contracts, turning call transcripts and pipeline data into decks. That is a description of a RevOps job. The AI lab built a coding tool and the customers turned it into an ops hire.
|
|
|
|
|
When OpenAI released ChatGPT, “agentic” was not part of the lexicon in the way that it is today. Over time, as capabilities grew and structured methods for interacting and building with LLMs evolved, agents became the pathway for them to move beyond a chatbot and into a coworker. This new coworker can operate along a spectrum of autonomy: insights only → human-in-the-loop → fully agentic.
In the Stack Deep Dive, we covered Outreach’s Deal Agent gaining write access to CRM fields. This is a “fully agentic” workflow, even if it is for very select fields with constrained values. But it opens the question that I, and lots of operators, are asking: when to use fully agentic versus human-in-the-loop?
When RevOps teams build processes and workflows, we often center the use cases on providing value to someone else, and most often it is sellers. For example, RevOps wants MEDDPICC fields updated to help predict pipeline progression, so we highlight that MEDDPICC completeness leads to sellers closing more deals. It’s true, but also self-serving.
Designing lightweight, frictionless workflows to meet sellers where they work certainly helps. Providing value to the seller so they want to engage with the workflow is ideal. But if we’re honest, there are workflows that will not directly help an AE make their next sale, and RevOps still needs them to run the business. You can build a human-in-the-loop workflow to get 90% of the way there, but if the user won’t engage in the last 10%, you’re nowhere. If you’ve exhausted other options, it’s worth exploring whether a fully agentic workflow can unlock RevOps and give more time back to the seller.
So the next question is: how do we do it safely? Full disclosure, I don’t have a magic answer here. Or an answer at all. But a few things come to mind.
The control point here is not the CRM; it’s in the agent architecture.
|
1.
|
Auditability is plumbing, and nobody has ever been promoted for plumbing, but plenty of people have been fired for a flood. Make sure each agent run has its findings logged, time-stamped, and stored in a SOX- and MNPI-compliant data warehouse.
|
|
2.
|
Your CRM’s field history has always been a record of what people did. Now it’s a record of what people and models did. Make sure to create distinct user IDs for agents so you know which agent updated which field or took which action.
|
|
3.
|
Production workflows are a layered cake. Don’t have an agent write directly into a production table or system. Create staging environments or external tables and grant the agent write access there. This acts as a control point and quality check before nonsense data gets written into production systems.
|
|
4.
|
Every agent needs a rollback plan. Field history, versioned artifacts, partitioned tables, and anything else that stores the status quo is worth its weight in gold in a fully agentic workflow. Have a plan that can revert the agent’s output and restore the correct data if needed. That is another reason to be cautious with fully agentic workflows that reach customers. You can’t roll back what a customer has already seen.
|
There are some workflows that should always have a human in the loop. But if your RevOps team is exploring fully agentic workflows, I’d love to hear the safety considerations you’re making along the way.
See you next week. — Andrew
|
|
|
|
|
