Is the ceiling on a GTM agent the data, not the model?
Clay says yes. On November 3, 2026, Clay published a test it ran itself: a general-purpose agent connected to Salesforce and Gong through MCP caught a planted duplicate-record deal risk in 18 of 30 runs, and Clay's Account Agent caught it in all 30. Clay also reported 6.4 times fewer input tokens and answers 2.6 times faster. Those figures are vendor-reported. Even if they hold, the useful question is who owns the organized account context, and whether you can keep it when you change agents.
Part of Answers
Our view
Clay's argument is that as models get cheaper, more of the advantage comes from how account data is organized. That part we agree with. A capable model cannot reason about a record it never opens.
Where we differ is what to do with that observation. Clay builds both the data layer and an agent designed to read it. The risk is that the organized context, the resolved account, the filed history, the learned summary, lives inside that vendor's system. If you later want your own agent, or a second vendor's agent, to use what you already learned, you may have to start over.
Audienti is built so the learning stays with the team. Your contacts, messages, replies, the premises you tested and the results live in your CRM record and in your own agent. Human and agent work should accumulate instead of resetting each time you change tools. That context should persist across agents and be exportable, not trapped in a vendor data layer.
Nooks is making a related claim with its Revenue Agent Platform: shared context and a learning loop across calls, email and CRM, held on the vendor's platform. The buyer question is the same in both cases. What does the memory remember, and who owns it?
What Clay reported
All numbers in this section come from Clay's November 3, 2026 post, written by Linda Ye and Nisola Shobayo. WeekCRM summarized the same test on November 4, 2026 and noted that Clay published the results itself.
- Duplicate-record deal risk. Clay says it planted a deal-threatening email on a duplicate Salesforce account, with a related Gong call logged under a subsidiary name. Clay's Account Agent surfaced that threat in 30 of 30 runs. The general-purpose agent, connected to Salesforce and Gong through MCP, caught it in 18 of 30 runs. Clay also writes this as 60 percent of runs, which matches 18 of 30.
- Conflicting risk fields. Clay says a sync flipped an account's risk rating from Medium to Low while a health score looked poor and usage was down 21 percent over the past month. Clay's Account Agent caught the conflict in 16 of 30 runs. The general-purpose agent caught it in 0 of 30, because none of its queries asked for those fields.
- Champion fields cleared by a sync. Both setups caught this planted problem in 24 of 30 runs, according to Clay's appendix table.
- Tokens, time and list-price cost. Clay reports average input tokens of 250,000 for the general-purpose agent and 39,000 for its Account Agent, which is the 6.4 times gap in the headline. It reports 60 seconds versus 23 seconds per answer, which is the 2.6 times gap, and list-price cost of $0.56 versus $0.006. It also reports 15 tool calls versus 9.
How Clay says it ran the test
Clay's appendix describes five synthetic accounts sized like enterprise accounts. Each had a six-person buying group, three opportunities, 31 contacts, about 10 emails, and up to 16 Gong calls, with transcripts as long as 61,000 characters. Each account's calls sat among 416 calls from 80 other companies. Clay asked two questions per account, three times each, for 30 runs per setup: "What's my next play on this opportunity?" and "Write me a pre-call brief."
Clay says both setups had the same facts, including 11 Clay signals logged in Salesforce as Tasks. It merged the duplicate account's records ahead of time to simulate Audiences matching, so the test "measures the value of resolved context and doesn't measure matching accuracy." The two setups used different models. The token, time and cost figures leave out the cost of building and maintaining the account context. A judge model graded each answer as caught, flagged, missed, or not addressed.
Clay also compared two ways its own Account Agent reads an account, filesystem browsing versus typed tools, across 650 runs. It reports overall quality scores of 0.96 versus 0.95, with a larger gap on account history (0.92 versus 0.82). Those scores are also vendor-reported.
What those numbers do not prove
They do not prove that Clay's Account Agent will outperform a general agent on your live Salesforce and Gong data. The accounts were synthetic. The deal risks were planted. Clay designed the setup, picked the models, graded the answers and published the write-up.
They also do not prove that a vendor data layer is the only way to give an agent organized context. Clay itself says the misses came from retrieval, not from a weaker model, and that listing the right fields in a prompt is brittle. That is an argument for organized, complete account records. It is not, by itself, an argument that those records must live only in Clay.
If you are evaluating any vendor's agent demo, including ours, ask to see work on your own accounts. How do you check an AI SDR demo against what the vendor claims? covers that checklist.
Who should own the account context
Organized account memory is valuable. The buyer question is whether you can take it with you.
Ask any vendor that sells a data layer or a learning loop what you can export: the raw records, the resolved account file, and anything the agent learned. Ask whether a different agent can read that export. Ask what is deleted, and how fast, when a contract ends. The Nooks comparison lists the same questions for a revenue agent platform.
Audienti keeps that context in your CRM record and in your own agent. Every campaign starts from a testable premise, why a specific buyer should care now, and you keep what produces replies and meetings. You approve every send. Outreach goes from your own LinkedIn and inbox. Your agent's own memory still follows that agent's vendor terms if the agent is hosted, so read those terms too.
See how buyer signals work and where AI SDRs work and where they break if you are weighing a vendor-run agent against a setup you own.
Common questions
What did Clay claim about GTM agents and data?
On November 3, 2026, Clay published a post titled "The ceiling on your GTM agent is the data, not the model." Clay says it compared a general-purpose agent connected to Salesforce and Gong through MCP with Clay's Account Agent, which reads from Clay's Audiences data layer. Clay reports that its agent caught a planted duplicate-record deal risk in 30 of 30 runs, against 18 of 30 for the general agent, and that it used 6.4 times fewer input tokens and answered 2.6 times faster. Clay ran the test itself.
Are Clay's test numbers independently verified?
No. The figures come from a test Clay designed, ran, and published. WeekCRM summarized the same numbers on November 4, 2026, and noted that Clay published the results itself. Treat them as vendor-reported until someone outside Clay repeats the work.
How did Clay say it ran the test?
Clay says it built five synthetic enterprise accounts and asked two questions on each account three times, for 30 runs per setup: "What's my next play on this opportunity?" and "Write me a pre-call brief." Each account had a six-person buying group, three opportunities, 31 contacts, about 10 emails, and up to 16 Gong calls. Those calls sat among 416 calls from 80 other companies. Clay says it planted a deal-threatening email on a duplicate Salesforce account and merged that duplicate ahead of time to simulate Audiences matching. The two setups used different models. The reported token and time figures leave out the cost of building and maintaining the account context.
Who should own the account context a GTM agent uses?
The team should. Context that compounds, the contacts, messages, replies, premises and results, should live where you can keep it when you change agents, and in a form you can export. A vendor data layer can organize records well and still trap what the system learned if you cannot take that context with you.
Does this mean you should not use Clay?
No. If you already run Clay and want its Account Agent reading a resolved account file, that may be a good fit. The open question is whether the organized context stays yours when you add another agent, or leave. Ask Clay, and any other vendor, what you can export and what happens when a contract ends.
How does Audienti handle shared context?
Audienti is the harness around your own agent. Your contacts, messages, replies, the premises you tested and the results live in your CRM record and in that agent. You approve every send, and outreach goes from your own LinkedIn and inbox. Start for $5 ($20 in credits). Plans then run from $20 a month: Team $100, Division $200, Enterprise from $3,000. A forward deployed engineer is from $5,000 a month if you want an engineer to work with you.
See who is in market for your offer
Tell us what you sell and we send 10 named accounts showing buying signals this week. Free. Or start for $5 and get $20 in credits. Not useful? We refund the $5.