The GTM Tech Index Brief

The past two weeks in one line

Two weeks ago three vendors moved the last human checkpoint to somewhere after the send. Since then it has come back, with Outreach, Pipedrive and ActiveCampaign each putting an approval step in front of the write. Clay and HubSpot handed campaign and sequence creation to code, Apollo and HubSpot taught their APIs to say no, and Pipedrive, Membrain and Close each folded a paid add on into the base plan.

This issue covers two weeks, September 6 to 19: 27 entries across 26 vendors, 21 Verified at source and six Partially Verified. One of them, LeanData, carries an August 24 date and reached the log after the last issue had gone out, so it is picked up here. Funding rounds, valuations, and awards are not logged, here or anywhere on this index.

The approval step came back

The last issue ended on three vendors removing a human step in the same week. The log since then reads like the reply.

Outreach’s September release adds approval previews before its Omni assistant changes a Prospect, Account or Opportunity record. The assistant still proposes the edit. A person now sees what will change before it does. The rollout is staged from September 10 to 24, and the entry is recorded Partially Verified.

Pipedrive launched Nova, meeting intelligence that prepares briefs from CRM data, captures the meeting, and drafts the summary and the CRM updates. Suggested changes need the user’s approval before anything is written, and an admin can switch the feature off.

Capture runs through a visible notetaker or through desktop audio, and managing participant consent is left to the user. That matters rather more on the second route than the first, since desktop audio puts no notetaker in the room for anyone to notice.

ActiveCampaign shipped two Active Intelligence updates in three days. The assistant now preserves the audiences a campaign already has and reuses existing audience IDs without overwriting their definitions. Its approval cards keep an allow or skip decision across a page reload. Recorded Partially Verified.

ActiveCampaign also documents the gap that remains, which is that a manual save made at the same moment can still be overwritten. An assistant that can rewrite who receives a campaign needs a memory of what it was told. A vendor that publishes its own remaining gap is doing the thing this index grades for.

Three smaller entries belong to the same movement. Aloware version 11.38.0 adds prompt diagnostics for its AloAi agents, where a suggested fix is accepted by hand and a test call is tied to a prompt version. Close changed Auto Update so that the values it withheld for low confidence are now shown, and a user can inspect and accept them. Planhat’s Workflow Mode creates a record for every Automation run, which makes automated customer outreach auditable one account at a time. The Planhat entry is recorded Partially Verified.

The Close change is the quiet favorite. An empty field used to mean the software failed. Now it means the software declined, and a person can overrule it.

Our read

Six vendors put a person, a record or a reason between the model and the write. None of them made the agent less capable. They made its decisions visible at the moment they can still be refused, which is the only moment that counts. Nobody was forced to. There is no regulation requiring an approval preview on a CRM field, only the memory of finding out what an assistant changed by reading the forecast.

Buyer question

For any assistant that can write to your CRM, ask to see one change end to end: what it proposed, who approved it, and what the record looked like before. Then ask whether the approval step has reached your tenant, because a control in staged rollout is a control you may not have yet.

The agent took the first meeting

Qualified made Piper Presentations generally available. Its AI SDR now selects and presents uploaded slides during a live website conversation, in English voice and video, and follows up with the PDF by email. A team controls which decks are active and can edit the talk tracks. It is limited to the Premier, Enterprise and Ultimate plans.

An agent giving the first sales presentation is a real step past chat. It also makes the content library the oversight model. Whoever keeps the deck list current is now, in effect, managing a rep who works every hour the website is up.

Consensus shipped Discovery Demo Builder AI, which assembles existing videos and tours from a prompt, using audience fit and 30 days of engagement history. Each recommendation comes with a confidence explanation and stays editable. Rox introduced Artifacts, which generate decks, documents and reports from meeting transcripts and CRM records, then restyle them to a team’s own uploaded templates.

The template upload is the practical half. A deck that lands in the house style is one a rep will send. A deck that lands in the vendor’s style gets rebuilt by hand, and the time saving leaves with it. The caution from the last issue still applies, because a document built from CRM data is very easy to send without reading.

Qwilr extended Conditional Content from whole blocks down to individual text, images, columns and contract clauses. The rules can draw on CRM, meeting or transcript inputs inside its Smart Proposal Engine, which needs the Scale plan and consumes credits. A rule that decides which contract clause a buyer sees is worth having a lawyer read once, because after that nobody will read it again.

LeanData’s 9.x release took AI Inference Nodes to general availability and added Anthropic and Microsoft Azure beside OpenAI and Google Gemini. The provider list is the detail to read twice. A buyer now chooses which model reasons over its lead data rather than inheriting one. The new assistant inside FlowBuilder is in beta and explains a routing graph without editing it, so it helps an admin understand the logic and does not yet change it.

Our read

The agent’s job description widened, from writing the message to giving the presentation, building the demo and drafting the proposal. In every case the control is a library: approved decks, existing videos, uploaded templates, clause rules. That is a sturdier design than a prompt, and it moves the governance work to whoever owns the content. Most revenue teams have not tidied the deck library since the last kickoff.

The API got write access and learned to say no

Clay made its Plugin available to all customers. It bundles Clay’s CLI and API for coding agents such as Claude Code and Codex. General availability adds programmatic control of Audiences, campaign creation and analytics in Sequencer, and table queries.

That turns Clay from a tool an operator clicks through into infrastructure a team builds on. Write access from a coding agent is the part to govern. Decide which agents may create a campaign before one does.

HubSpot’s 2026.09 developer release expands the Sequences API to create, update and delete sequences, and adds an Email Template API, both authorized at the user level. New read only APIs expose forecast submissions and Notetaker recaps for Sales Hub Professional and Enterprise. Sequences were the part of Sales Hub an integration could watch but not build. Now outreach setup can be scripted.

Two vendors spent the same stretch teaching their APIs to refuse. From HubSpot’s /2026-09/ version, a request to delete a pipeline or stage that is still in use is blocked by default and returns a validation error, which is what the settings page already did. Apollo added structured error details to its REST and MCP errors, with stable codes and recovery guidance. An Organization Enrichment request with a missing or unreadable identifier now returns a 422, where it used to return a 200 and an empty organization.

For years the most dangerous response in GTM integration has been 200 OK. A workflow that treated an empty enrichment as a quiet success will now see a failure, which is the honest answer and also a breaking one. Apollo removes the legacy error fields on February 16, 2027, so integration owners have five months.

The connector work continued underneath. Clari’s September release adds Forecast MCP, seven permission scoped tools that let Claude and ChatGPT analyze forecast data under the user’s existing access. Membrain added individual OAuth authentication for Claude through MCP. Airspeed rebuilt its MCP server to run from the browser rather than a desktop install, which removes the step that stalls most rollouts, namely getting an installer past IT. Airspeed is recorded Partially Verified.

Attio’s App SDK version 1.1.0 lets developers build full app pages and custom record tabs inside the CRM, recorded Partially Verified. Lusha API release 2.14.0 lets a team define a table’s columns from code and enrich every row asynchronously, and the calls currently consume no credits. The word doing the work in that sentence is currently.

Our read

Permission scoped, user level, individual OAuth. The vocabulary of these connector entries is the vocabulary of identity, and that is new. The August entries were about whether an assistant could reach the system of record. These are about whose access it borrows when it does. An API that says no belongs to the same story, because an agent cannot be supervised through an interface that reports success either way.

Buyer question

Ask each integration vendor in your stack two things. Has it moved to the HubSpot /2026-09/ version, and does it handle the new 400 if it deletes pipeline objects? And does anything you run treat an empty Apollo enrichment as a success? September is the month those two questions became worth an afternoon.

The add on became the base plan

Three vendors folded something they used to charge for into the subscription. Pipedrive’s Nova is included on every plan at no extra cost. Membrain now includes Insight Engine Basic for every customer, with 500 credits and 100 MCP calls a month, while advanced enrichment and automation stay in paid packages. Close removed the split between personal and group phone numbers and gave every number the former Premium Number features at no extra charge, including shared ringing, round robin and lead based routing.

Nobody has ever needed persuading to read a release note that says the thing now costs nothing. The part to read is the allowance. Membrain’s monthly credits, and the line between included and paid automation, are what decide the real cost. Close’s change arrives with a separate caller ID permission for voice agents, which decides which AI may call out under which identity.

Two smaller entries point the same way. Gainsight’s Skilljar made multiple login methods a self service setting, where they used to be a Professional Services add on. Clay opened its Plugin and other newer products to customers on legacy Starter, Explorer and Pro plans until December 31, 2026, which is a free look with a hard end.

The counterweight is that the newest agent work went the other way. Piper Presentations is limited to three of Qualified’s plans. Qwilr’s clause level rules need the Scale plan and consume credits. Several of HubSpot’s new APIs depend on the Sales Hub tier.

Our read

Meeting capture, basic AI assistance and telephony routing all crossed from add on to table stakes at the same time. Last year’s AI feature becomes this year’s base plan, and the price moves to the allowance, the credit and the tier above. For a small team paying separately for a notetaker, the math just changed. For everyone else, the unit to compare is no longer the seat.

The unglamorous half

HeyReach released workspace blacklists that hold up to 1,000 people and 1,000 companies, with CSV import and additions through the API and MCP. New campaigns turn on both exclusion checks by default, and an excluded lead never enters the sequence.

The first issue of this Brief complained that compliance controls arrive as toggles with a default somebody else will set. This one arrived with the default pointing the safe way. A user can still switch it off for a campaign, and a list stops at the edge of its workspace.

QuickProsp shipped four releases between September 8 and 19. They add a separate tracking domain for each sending domain, warmup limits per mailbox with an API endpoint to set them, and wider automatic suppression. Version 1.26.0 adds Microsoft permissions that let warmup mark messages as read and move them out of Junk. Existing connections pick those up when reconnected, so an admin should know what is being granted before clicking through.

Planhat added permitted sending hours and days for email steps. It is the small control that keeps a customer success automation from emailing a client at three in the morning, which is technically proactive.

On the trust side, Cognism announced that its ISO 27001 and ISO 27701 certifications are now issued by British Assessment Bureau, a certification body accredited by UKAS, and published the certificate number and dates. Who issued a certificate matters as much as the certificate. Gainsight extended KMS encryption and masking of saved secrets across six connectors, so a saved credential no longer displays when an admin edits a connection. That is a classic audit finding closed with no migration work.

Three entries change something an admin will otherwise discover later. Clari caps recording storage at three years from September 15, while transcripts, AI insights and metadata stay available beyond that. Coaching libraries and retention policies both tend to assume the audio is still there. Clari’s notes now publish under the Salesloft name, following the two companies’ move to one brand on September 2.

Aloware’s new Inbox Metrics report counts unstaffed inboxes, which lifts reported totals by roughly a tenth, and its Random dial mode retires after September 30. Qualified introduced a guided migration that moves affected mailboxes from SendGrid to Amazon SES, which needs a DNS administrator and about a week and a half.

Somewhere a dashboard is about to show a ten percent lift, and somebody is about to take credit for it. Note the date before comparing months.

Our read

Every entry in this section is a change somebody will otherwise meet in a month, at a bad moment. A retention cap, a retired dial mode, a report that counts differently and a permission prompt on reconnect are not features. They are dated facts about a platform, and the sector deserves credit for writing them down before they bite.

Market notes

Klenty published an analysis of 22,195 retry attempts across 112,225 US calls. Retries made the next day were answered 14.67 percent of the time, against 12.27 percent for retries within 15 minutes. Meeting rates were 9 percent against 2 percent, and negative reactions 22 percent against 35 percent.

It turns out the prospect who did not pick up ten minutes ago has not, in the intervening ten minutes, changed their mind. The data is observational, from calls placed through Klenty, with no sampling period and no controlled comparison. Treat it as a reason to run your own test rather than a promised result.

Bigin by Zoho CRM released updates for Apple’s newest operating systems, with Siri queries, Apple Intelligence summaries of records and deals, and an AI task list on Apple Watch, for anyone who felt their wrist was not yet applying enough pipeline pressure. The gains depend on the device fleet being current.

Reevo updated its macOS app with local meeting transcription in more than 90 languages, recorded Partially Verified. Its documentation says supported languages can vary by workspace, so confirm the ones you need.

Three smaller items round out the log. Qwilr added Zoho CRM automations that update deal fields and stages when a proposal is accepted, declined or expires. Lusha release 2.15.0 added Instagram, YouTube and TikTok URLs to company enrichment on all plans. Consensus now uses Salesforce record IDs to attribute demo engagement, which is the plain change that makes the numbers trustworthy inside CRM reports.

What the past two weeks say about the category

The last issue described a sector that ships the capability and the control in the same release and leaves the default open. This time the defaults started to close. Outreach previews the write before it happens. HeyReach starts its suppression check switched on. HubSpot refuses a delete unless the caller insists. Apollo returns an error where there used to be a polite silence.

None of that slowed the agents down. In the same two weeks an agent gave a sales presentation, a coding assistant got the keys to campaign creation, and the sequence became something a script can build. The category is widening what the software may do and narrowing what it may do unobserved, at the same time, and the two movements need each other.

The commercial signal points the same way. When meeting capture and basic AI assistance fall into the base plan, the feature has stopped being the product. What is left to sell is the allowance, the tier and the control surface. What is left for a buyer to compare is how much of that control surface the vendor has written down.

Index Answer

Which GTM vendors publish what their AI does without a human?

About 4 in every hundred. Of the 314 vendors the GTM Tech Index has graded on Autonomy and Oversight Model, 13 reach the top grade. That grade means the autonomy boundary is on the public record. A buyer can read what acts without review, what waits for approval, what a bad run can and cannot touch, and which control stops a batch mid flight.

Another 107 describe the human in the loop posture in substance, usually draft versus auto send or an approval flow, and leave the failure story incomplete. The remaining 194 assert oversight without documenting it, or publish nothing, and a buyer cannot tell from public sources what runs unsupervised.

Set that beside the same 314 vendors on Ecosystem and Integration Depth. There, 196 document their integrations at a checkable standard, against 120 on autonomy. The field writes down how software gets into the system of record far more readily than it writes down what that software may do once inside.

The past couple of weeks made that gap matter more, because the thing getting in is increasingly an agent somebody else built. It also moved a handful of vendors the right way, each by one documented control. None of those controls is a published boundary. Each is a piece of one, and a piece a buyer can watch working in a trial rather than take on trust.

The practical move is to ask for the boundary in writing before the pilot. Ask which actions run without review and which wait for a person. Ask what a bad run can touch. Ask where the stop button lives. A vendor graded A can answer from its documentation. For most of the rest the honest answer is a setting, and the setting is yours.

What each grade means on every axis, and how often anyone earns it, is on the what to look for page.