RC
Somewhere in your finance team's inbox is a renewal notice for an AI agent nobody has opened since August. It will auto-renew in 30 days unless someone says otherwise. Nobody will, because nobody can say what it did.
This is the quiet problem with the agent budget. The buying was easy. The keeping is where it gets hard. Every agent your GTM team added this year was justified by a pilot, a demo, or a rep who liked it. Almost none of them were justified by a second look. And the second look is the one that matters, because that is where you decide what the stack becomes.
McKinsey's 2025 State of AI found that 62% of organizations are experimenting with AI agents, but only 23% are scaling them anywhere. In any single function, scaling tops out around 10%. Most agents never cross from trial to trusted. They just keep renewing.
So here is a renewal audit. Five questions, in the order a RevOps leader should ask them. None of them require an attribution model. All of them require evidence you may not currently have.
1. Did anyone actually use it?
Start with the dumb question, because it eliminates the most seats. Not "did we roll it out," but "did reps invoke it in a normal week." Pull the usage log if the vendor gives you one. If they do not, that is your first finding.
Usage tells you something attribution never will. A rep who keeps using a tool is voting with the only currency that cannot be faked: their own time. A rep who found a workaround is telling you the sanctioned tool lost on speed or trust, and no renewal will fix that.
Be honest about the shape of the curve. Launch-week spikes do not count. The number you want is week 8 onward.
2. Did it stay inside its lane?
The second question is governance. Did the agent only do what it was allowed to do, and can you prove it?
This is less settled than most teams assume. Deloitte's 2026 State of AI found that nearly 75% of companies plan to deploy agentic AI within two years, but only 21% have mature agent governance. An agent with write access to the CRM, the inbox, and the calendar is a user with no manager. If you cannot see every action it took, the renewal decision is really a trust decision, and trust is not a control.
The audit question is specific. List the systems the agent could touch. List the actions it actually took. If the second list has anything the first did not authorize, that is not a renewal conversation. That is an incident review.
3. What happened to its output?
This is the question most teams skip, and it is the most useful one.
Every agent produces something: an email draft, a research brief, a lead score, a proposed CRM update. The output has a fate. A human approved it unchanged. A human edited it. A human overrode it. Or nobody looked.
The ratio across those four buckets is the best signal you have for whether an agent earned its seat. High approve-unchanged means the agent is trusted and useful. High edit means the agent is close but costs the rep time. High override means the agent is confidently wrong. High ignored means the agent is noise.
Most teams cannot produce this ratio, which is the real problem. LangChain's State of Agent Engineering survey found that roughly 3 in 10 agent teams aren't evaluating their agents at all, only 52% run offline evals before deployment, and just 37% evaluate live traffic. If every write goes through an approval queue, the ratio falls out for free. If the agent writes directly to the system of record, you are reconstructing it from memory.
4. Did it touch deals that moved?
Now, and only now, look at pipeline. Not "did the agent create revenue," which is a question no honest dashboard can answer. Ask the narrower one: which deals did this agent act on, and what happened to those deals afterward?
Keep it descriptive. If the agent touched 40 opportunities and 30 of them advanced a stage within the quarter, that is a pattern worth keeping. If it touched 40 and 25 went dark, that is a pattern worth cutting, even if the agent's own report says it sent 1,200 emails. Activity is what the vendor measures. Deal movement is what you are paying for.
Resist the urge to turn this into an attribution model before the renewal. Co-occurrence is enough to rank agents against each other. Causation is a longer project.
5. What does it know that nothing else does?
The last question is the one with the longest consequences. If you cancel this agent, what leaves with it?
Most agents keep their own memory. Every preference a rep taught it, every account note it absorbed, every correction it learned from, lives inside that vendor's walls. Cancel the contract and that learning is gone. Keep the contract and you are paying partly for ransom.
This is the structural reason agent sprawl is expensive. Memory is per-agent, so nothing compounds across the stack. Each renewal is evaluated alone, because each agent learned alone. The teams that get past this are the ones that moved the memory out of the agents and into a layer every agent reads from. At that point the renewal question gets easy: the agent is a replaceable skill, and the knowledge stays.
The audit is hard because the evidence is scattered
Run those five questions and you will notice the real pain. The usage log is in one vendor's admin panel. The permission list is in your identity provider. The output fates are in reps' heads. The deal movement is in the CRM. The memory is locked in the agent.
That scatter is not an accident. It is what happens when agents are added one at a time, each wired separately into the stack, each with its own memory and its own rules. The agents work. The stack does not.
We built wysdym as the operating layer for agentic GTM so that this evidence is a byproduct of running agents, not a quarterly archaeology project. Every agent on the layer acts under shared Governance, so every write is logged and queued for approval. Observe shows which agents are actually working. Cortex holds the memory, so cancelling an agent does not erase what your team taught it. The renewal audit stops being a project and becomes a view.
Whatever you run your agents on, do the audit before the notices arrive. More agents will not grow revenue. Knowing which ones earned their seat will.
If you are staring at a stack of renewals and cannot answer question three, talk to the founders. And if you want the market data behind all of this, the sources are at gtmstats.wysdym.ai.
