The GTM Tech Index Brief

The week in one line

HeyReach shipped AI voice notes built from cloned voices, Rox let its agents route approved prospects into sequences with nobody reviewing the list first, and Airspeed pointed its agents at decks and code. In one week three vendors moved the last human checkpoint to somewhere after the send.

This issue covers August 30 to September 5: 11 entries across nine vendors, four Verified at source and seven Partially Verified. One of them, Membrain, carries an August 29 date and landed 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 assistant got a rate limit

The first issue described a sector opening its systems of record to assistants it does not run. This week the same vendors started publishing the numbers that say how much of that traffic they expect.

Fireflies launched an MCP server piping meeting data into Claude, Devin and ChatGPT, and raised the Pro plan API limit from 50 to 500 requests a day in the same release.

The rate limit is the part to read. A tenfold increase shipped alongside the connector says the vendor expects programmatic access to become ordinary traffic rather than an edge case, and it gives a buyer asking whether conversation data can actually leave the platform at volume a published number to check rather than a promise to weigh.

Membrain shipped a built in AI Assistant and an MCP connector in one go, and the connector is the half that matters. An assistant inside the product is table stakes in this category now. Letting a customer bring their own model to the CRM record is a different posture, and it says the vendor intends to compete on the record it holds rather than on the chat window wrapped around it.

Clari added Ask Copilot, which queries up to 15 selected calls at once, and AI Themes, which analyses up to 1,000 calls into a structured report on recurring topics. Its iOS app can now record in person meetings and sync the transcript to the CRM.

Conversation intelligence has been stuck at the single call for years, and coaching a team one call at a time does not scale past a small floor. Published ceilings of 15 and 1,000 give a buyer something to test rather than a claim about cross conversation insight. In person recording closes a real gap for field teams and brings the two party consent question along with it, into rooms where nobody is looking at a notice on a screen.

Our read

Three vendors published a specific number this week: 500 requests, 15 calls, 1,000 calls. That is a healthier development than it sounds. A category that competes on adjectives cannot be compared, and a category that publishes ceilings can be. Ask for the number early in an evaluation, because the vendors that will not give you one are telling you something about where their limits actually sit.

Buyer question

For any conversation platform, ask what the API limit is on the plan you are buying, not the top plan, and what happens when you hit it. Then ask who at your company can turn on in person recording from a phone, and what that person is required to say before pressing record.

The last checkpoint moved after the send

HeyReach introduced AI voice notes: personalised audio generated from text using cloned or featured voices, with a campaign step that supports both static recordings and dynamic audio pulling in personalisation variables such as name and company.

This is synthetic voice at campaign volume on a platform where a voice note reads as a personal gesture, which is exactly why it works and exactly why it needs a disclosure answer. Nothing in the announcement addresses whether the recipient is told the voice is generated. Anyone in a two party consent state, or inside the scope of Article 50 of the EU AI Act, wants that settled before the feature is switched on.

Rox introduced auto prospecting, where agents build and evaluate prospect lists against buyer criteria, score each contact for fit, give their reasoning for the score, and can route approved prospects straight into outreach sequences with no human review first.

The published reasoning is the good half, because a fit score a rep cannot interrogate is a number they learn to ignore by the second week. The routing is the half to configure carefully. It moves the last human checkpoint from before the send to after it, and the sending domain absorbs whatever the agent got wrong in the meantime.

Airspeed widened what its agents produce from summaries to artefacts: charts, presentation decks and code, with CRM fields now displayed inside the Deal Board rather than one jump away.

Moving from summarising a conversation to producing the thing that follows it is a real widening of scope, and it is where output quality gets harder to check. A deck built from call data is easy to accept without reading, which is a different risk from a transcript nobody reads.

Our read

Three vendors removed a human step in the same week, at three different points: before the list goes out, inside the message itself, and after the call. None of the three is unreasonable on its own, and the review capacity they assume is the same person. When the list is unreviewed, the voice is generated and the deck is unread, the only remaining control is the one nobody has configured, which is what the recipient is told.

Buyer question

Before enabling any of this, write down where the human review now happens and who does it. If the honest answer is that review happens when a prospect complains, that is not a workflow, it is a deliverability strategy, and your domain is the thing being tested.

The unglamorous half, which is where the money is

Chili Piper shipped a Tie Break Order setting that picks a team member at random when reps are tied on credits for a lead, and a retry matching option that lets Marketo and Salesforce nodes sit on the same routing path by absorbing sync delays.

The tie break sounds trivial and is not, because a tie resolved by list order quietly hands volume to whoever sits first, every time, for as long as nobody checks. The retry change addresses the more expensive failure: a lead falling out of a routing path entirely because two systems disagreed for a few seconds about whether a record existed.

Qwilr connected its Salesforce integration directly to its Smart Proposal Engine, so proposals generate from CRM data instead of being assembled from it by hand. Proposal tools live or die on whether the data arrives without someone retyping it, and the question a demo will not answer is which fields the engine actually reads.

The same vendor also released an account wide data export covering account data, proposal performance and platform usage. This index grades that as portability rather than reporting, and the distinction is the point.

A proposal library and its performance history are among the least reproducible assets a revenue team accumulates. A vendor that ships a self serve export is answering the exit question before a buyer has to raise it in a renewal, which is a stronger signal than anything on the pricing page.

Autobound added a podcastAppearance signal subtype, separating the case where an executive appears as a guest from the existing podcast signals. Signal vendors mostly compete on breadth, so a new subtype at that granularity is a claim about precision instead, and precision is the difference between an opening line that is personalised and one that only looks it.

Membrain also shipped Product Revenue Graphs, aggregating product rows from Sales Projects, Quotes and Orders into one view. Useful mainly to teams whose quotes originate in an ERP and sync in, which is where the product level view usually fragments across two systems.

Our read

Every entry in this section is a fix for something a customer had already worked around, and worked around silently. Nobody files a ticket about a tie break, they just notice one rep is having a good quarter. Nobody asks for an export until the day they are leaving. These are the changes that never appear in a competitive comparison and decide how the platform feels in year two.

What the week says about the category

Two weeks ago the theme was access: who can reach the system of record and with whose permissions. This week it is autonomy, and the two are the same argument arriving from opposite ends. An assistant with write access needs a permission model. An agent that sends without review needs a disclosure posture. Both are controls the vendor ships and the buyer configures, and both default to open.

The pattern worth carrying forward is that this sector keeps shipping the capability and the control in the same release, which is more than most sectors this index covers manage. What it does not ship is the default. Rox publishes its reasoning and lets you skip review. HeyReach clones a voice and says nothing about telling the recipient. In both cases the setting that matters is left for somebody downstream to choose, and that person is usually measured on volume.

Index Answer

Which GTM vendors tell the recipient they are dealing with software?

Four. Of the 314 vendors the GTM Tech Index has graded on Recipient Disclosure and Authenticity, 4 document that automated interaction discloses itself by default, with real or clearly labeled sender identities, and address the Article 50 obligation by name where it applies to the product.

Another 35 publish a disclosure posture that is available but not the default, or cover the ground with one real gap. The other 275, 208 at C and 67 at D, publish nothing a buyer can read on whether the person on the other end is told.

That is not a gap in the marketing. It is the answer. For a product whose AI writes to prospects, or now speaks to them, silence on disclosure is a regulatory posture rather than a style choice, and it is the posture the vendor has chosen.

The adjacent axis reads the same way. On AI Disclosure and Model Transparency, 10 of 314 name their models, address AI generated content disclosure and explain their scoring or routing logic well enough for a buyer to interrogate it.

The reason this matters more this month than last is that the output stopped being text. A generated email that reads slightly off is a quality problem. A cloned voice in a note that the recipient believes a person recorded is a different category of thing, and the obligation to say so does not attach to the vendor. It attaches to the sender, which is the buyer.

The full grading, the axis definitions and where each vendor sits are on the outreach compliance page.