The AI model is interchangeable.
Your GTM infrastructure is not.
On Monday, an agent refreshed our public jobs board with the week's new roles and opened the pull request on its own. I read the diff and merged it. That is the honest shape of AI agents in go-to-market right now: real work, shipping every week, with a human hand on the last gate.
AI agents in go-to-market read state, interpret ambiguous signals and prepare actions across research, data quality, scoring and outreach. Deterministic rules handle stable logic such as routing and execute approved changes; humans gate what reaches a buyer. The agent supplies judgment. The platform underneath it supplies everything that makes the judgment usable: the data it reads, the context it applies, the gates it cannot cross, and the record of what it did.
The platform is the part that lasts. Alma moved roughly 160 workflows onto ours in a weekend and published the writeup themselves. A commercial operator, zero engineering background, took an enrichment workflow to its tenth version within six days of first touching the platform, working through Claude Code; the full story lives on the platform page. And the jobs board above is not a demo. It ships weekly.
Bringing your own agent? Hand it this:
npx skills add getcargohq/cargo-skillsThis page also serves markdown for agents: append .md to the URL.
What do AI agents actually do in GTM today?
Six jobs run in production. Not a roadmap: this is what we run and what our customers run, with the column no vendor grid includes: where it breaks.
| The job | What the agent judges | What stays deterministic | Where it breaks | Human gate |
|---|---|---|---|---|
| Account research and qualification (TAM building, fit scoring, meeting prep) | Whether the company matches your written ideal customer definition, and what a rep must know before the call | Source fetching, scoring arithmetic, field writes | Ordering. Run news research before identity resolution and you get a beautiful brief on the wrong company | Read and recommend. Scores stay suggestions until spot-checks hold |
| Contact data work (list building, enrichment, verification) | Which of two records is the real person, and whether a title means budget ownership | Waterfall order, provider calls, verification, dedupe rules | False positives and unit economics. A title match is not a budget owner, and an ungated waterfall on a 40K-account market burns the monthly credit budget by Thursday | Write, inside caps. Spend per run is capped; merges need a confidence bar |
| Signal interpretation (sales readiness) | Whether a funding round, job change or usage spike matters, for which account, and what should happen next | Signal capture, recency weighting, decay, thresholds | Stale weighting. The signal that meant a buying window last quarter can be noise this quarter, and nobody reweights it | Recommend. The signal proposes; a workflow or a person disposes |
| CRM operations (CRM cleaning, pipeline health) | Which duplicate wins, whether a deal is genuinely stalled, who should own an account | Sync, field mapping, assignment execution | Blast radius. One bad send burns a domain; one bad owner-write quietly corrupts a quarter of reporting | Confirm-only. Identity merges and owner or stage writes wait for a click |
| Outreach drafting | The angle: which signal, which pain, which proof, for this specific person | Sequencing, send windows, suppression lists | Context-free personalization. Grounded in nothing, it reads as spam with mail-merge ambition | Send is human today. Every first touch leaves through a person |
| Building and repairing the engine itself | What the workflow should do differently, and what the run history says went wrong | Type checks, the plan diff, the tests a change must pass | Unreviewed changes. An agent that edits production without a diff is a new intern with admin rights | Pull request. The agent proposes code; a human merges it |
Deterministic steps do the moving; the agent does the judging.
The gate column is a ladder, and it is the actual control model: read, then recommend, then write, then send. An agent earns the next rung the way an employee does, with a track record you can inspect. Promotion is a decision you make against run history, not a toggle you flip on day one.
The counterintuitive part of the split: lead routing looks like judgment and should stay deterministic, because it is arithmetic over capacity and territories. CRM dedupe looks mechanical and is exactly where you want the agent, because every merge is a judgment about identity.
And none of this is sales-only. The same six jobs run customer health, renewal risk and expansion signals, the way our own customer-health loop does.
Before an agent touches production, run the agent sanity check
Four questions. If any answer is no, you do not have an agent problem. You have a platform problem.
1. Can it read real state?
Not an export from last Tuesday: the accounts, contacts, deals and events as they are right now, queryable.
No data = guessing.
2. Does it know what good means here?
Your ideal customer definition, personas and objections, written where the agent reads them, versioned so you know which definition decided.
No context = generic judgment.
3. What can it change without approval?
The answer should be a short written list. Everything not on it is a recommendation, not an action.
No guardrail = blast radius.
4. Can you replay and score its decision?
What it read, what it decided, what it cost, what happened next, inspectable after the fact.
No trace = the same mistake twice.
Not one of the four comes from the agent.
Every one is the platform underneath it.
What are you actually buying when you buy AI for GTM?
Four buying models get sold under one label, and most stacks run more than one. The test that separates them: when you switch, what remains portable?
AI SDR vendors
You rent the coworker. The judgment, the learning and the history accumulate in their system. When you leave, what remains portable is the contact list, at best.
Workflow tools
They run decisions people already made in a builder. Fast, dependable, and often the right home for stable logic. What remains portable: the diagrams, if you rebuild them elsewhere.
Research tables
Agents reasoning over rows you rent. Strong for prospecting research, but the history lives in the table, and table-level monitoring is not a replayable decision trace. What remains portable: exports.
GTM infrastructure
Agents and humans operating one system you own: the data, the context, the gates and the run history in one place, defined as code. Portable by construction, because it is yours. The model this page is about.
Why does the platform matter more than the agent?
Here is the part of the agent conversation vendors skip. The model inside your agent will be replaced within quarters, but the swap still has to pass real evaluations. And do not confuse it with swapping the agent: a production agent is tools, state, retries, evaluations, approval logic and failure history, and none of that moves in an afternoon. The model is the replaceable part. The operating system around it is not.
What compounds is everything the agent stands on. The account graph you resolved. The context you wrote down. The gates you tuned. The run history that tells you what actually worked. Change agents and that layer stays. Lose that layer and no agent, however good its model, gets you back to even.
Agents come and go. The revenue engine must be owned, versioned, and continuously repairable.
This is also where I part ways with the loudest claim in the category. Apollo announced "the Industry's First Fully Agentic End-to-End GTM Platform" in October 2025, then "the first fully agentic GTM operating system" in March 2026. I will take the other side of that trade: in our production work, agents earn autonomy one rung at a time, from read to recommend to write to send. We have not seen enough evidence to remove gates where a mistake can burn a domain or corrupt CRM ownership.
How we hold ourselves to it: Cargo is GTM infrastructure built for exactly this split, and we run our own revenue engine on it. Our market is continuously rescored as buying signals land; coverage scales with credits, not headcount. Claude, ChatGPT and Cursor connect to the platform over OAuth. The honest gap: our fully agent-run engagement lane is still in build, gated exactly the way this page prescribes. The platform case, receipt by receipt, is one page over: why AI agents operate well on Cargo.
How do you put an AI agent into GTM production?
Five rules, in order. Start with hygiene and identity, never outbound.
1. Write the ICP and the do-not-contact list where the agent can read them
Judgment starts from a definition. Put your ideal customer profile and your suppression rules in a place the agent reads at run time, versioned, so you always know which definition made a call.
2. Cap spend per run
Credits, sends, API calls: hard caps enforced by the platform, not left to the agent's judgment. A cap turns a bad run into a small bill instead of a blown monthly budget.
3. Make identity merges and CRM owner or stage writes confirm-only
These are the writes with the widest blast radius and the slowest recovery. The agent prepares the change and states its reasoning; a person clicks.
4. Route every outbound message through a person
Sending is the last rung of the ladder, not the first. Drafts are agent work; the send belongs to a human until the review record says otherwise.
5. No trace, no production
If you cannot replay what the agent read, decided and changed, it does not run unattended. This rule is what makes every other rule enforceable.
The order is the point. The first two rules make judgment possible, the middle two make mistakes cheap, and the last one makes the whole thing improvable. Teams that start at outbound run the protocol backwards and pay for it in burned domains.
Keep reading
The reading path on agents in go-to-market, in reading order.
GTM Software Was Built for Humans. Agents Are the New Users.
The category thesis: agents need infrastructure underneath them, not a prettier interface on top.
Read it →AI Loop Engineering for GTM
Coding agents inherited a free verifier: tests pass or fail. Revenue has no test suite, and building the missing verifier is most of the work.
Read it →Revenue Orchestration: The Decision Layer of GTM Infrastructure
Where commercial decision logic lives once both people and agents act on it: explicit, versioned, and readable by both.
Read it →Cargo with Claude Code
Install the CLI and skills with one prompt in Claude Code, then launch a first workflow in minutes.
Read it →Frequently asked questions
The questions operators ask about agents in go-to-market, answered to stand alone.
Sources: Apollo ApolloNEXT announcement, October 2025 · Apollo AI Assistant announcement, March 2026 · Alma migration writeup, Medium
Last verified: August 26, 2026.
Give your agents a runtime
Bring the agents you have.Start free, deploy in one command.