The way to tell real AI agents for sales and marketing from staged ones is to ask for artifacts a production system generates in the course of existing: traces, override rates, rollback records, rate cards, exit exports. This handbook gives procurement 40 questions across six sections, with why each one matters, what a good answer sounds like, and what a red flag sounds like. Ask all 40 before you sign. The answers price the risk for you.

The category has earned this scrutiny. In June 2025 Gartner counted thousands of vendors claiming agentic products, estimated that only about 130 were building real agents, named the practice agent washing, and forecast that over 40 percent of agentic AI projects will be canceled by the end of 2027 on cost, unclear value, and weak risk controls. A July 2026 Forbes analysis by Robert Szczerba named the mechanism a capability-deployment verification gap: agents that perform in demos and fail on live proprietary data. The same piece cites Forrester finding 75 percent of enterprises have adopted agentic AI while few reached production. Buyers in 2026 have plenty of appetite and no instrument. This is the instrument.

Figure

The market procurement is buying into

40%+of agentic AI projects will be canceled by end of 2027Gartner, 2025

~130of thousands of vendors claiming agentic products build real agentsGartner, 2025

75%of enterprises adopted agentic AI, while few reached productionForrester via Forbes, 2026

96%of leaders say agents with full-funnel context would significantly improve executionRevSure, 2026

Sources: Gartner, June 2025 · Forbes citing Forrester, July 2026 · RevSure, The 2026 State of Agentic AI in B2B GTM (306 leaders)

The appetite is real. In RevSure's 2026 study of 306 senior marketing, sales, and RevOps leaders, 76 percent are already deploying or implementing agentic AI, and 90 percent say it is critical to their GTM goals within two years. The sponsor wants speed without regret, and the best description of that posture came from a CMO in exactly this position:

I'd much rather buy than build. I'm not a blocker here. I just want to understand what we would be buying if we did.

a CMO at a field service management software company

The same CMO also gave the reason buy beats build: "And I couldn't be able to keep up with your roadmap. So that's the logic behind it." Understanding what you would be buying is the whole job of this document. The 40 questions run in six sections: context, provenance, autonomy and control, economics, reliability, and exit.

How to run the evaluation

Send all 40 questions to every finalist in writing, then score every answer on one axis: did the vendor produce an artifact or an adjective. An artifact is a trace you can open, a rate card you can file, a rollback you can watch, an export you can inspect. An adjective is a paragraph.

Less effective: attaching the list to a security questionnaire, accepting prose responses, and scoring vendors on how thoroughly they answered.

Recommended: converting the hardest ten questions into live demo requests on the vendor's own screen ("open a trace", "show us a rollback", "run this on our data"), because artifacts do not have a marketing department.

Context: what AI agents for sales actually read

Agents act on whatever they can see, and in B2B GTM what they can see is fragmented by default: the average company runs 22 GTM tools. In the 306-leader study, 47 percent cite lead quality and data reliability as their top barriers to agentic AI, and 96 percent believe agents with full-funnel context would significantly improve execution. RevSure's founder, Deepinder Singh Dhingra, compresses the failure pattern to one line: "AI for GTM doesn't fail at the model. It fails at the context." These seven questions establish what a vendor's agents actually see.

Question	Why it matters	A good answer sounds like	A red flag sounds like
1. What exactly does this agent read before it acts, and from which systems?	Agents act on the context they can see; missing systems mean missing judgment.	"Here is the connector list and the exact objects and fields each agent reads."	"It integrates with everything." No list follows.
2. How do you resolve identity when the same person appears differently in the CRM and the marketing automation platform?	Unresolved identities double-count people and split one buyer into several.	"A named resolution layer with match logic you can inspect record by record."	"The model figures that out."
3. What happens when two of our systems disagree about the same account, stage, or owner?	Enterprise systems disagree constantly; the agent needs a ruling, and you need to know it.	"Precedence rules you configure, plus a log of every conflict we resolved."	"That rarely comes up in practice."
4. How much of our history does the agent use, and how does its context stay current?	Stale context produces confident actions against last quarter's reality.	"Continuous sync with stated latency per source, and stated history depth per object."	"We take a snapshot during onboarding."
5. Who maintains the funnel definitions the agent reasons over, and where do they live?	If stage definitions drift across teams, every downstream action inherits the drift.	"One canonical funnel definition, versioned, with a named owner on your side."	"Each team can keep its own definitions."
6. Can we see the exact context an agent had at the moment of a specific decision?	Debugging an agent without its context view is guesswork.	"Open any run and inspect the inputs the agent saw at that moment."	"We can pull logs if you file a ticket."
7. Will you demo on our live data before we sign?	The verification gap lives between curated demo data and your production reality.	"Connect read-only this week and we will run the demo on your records."	"Our sample dataset represents typical customers."

Provenance: a higher bar than attribution software ever met

Attribution software spent a decade producing numbers that could not survive cross-examination, and buyers remember. Agentic systems raise the stakes, because now the number acts: it sends the email, moves the budget, reprioritizes the queue. A marketing operations leader at a software testing company set the field's real standard for any reported figure: "I would rather have a worse model that I can actually explain to people if they ask." Provenance is that standard made mechanical.

Question	Why it matters	A good answer sounds like	A red flag sounds like
8. Can every number you report be decomposed to the underlying records on demand?	A figure that cannot be decomposed cannot survive a finance review.	"Click any figure and see the deals, contacts, and events behind it."	"The methodology is proprietary."
9. Does every agent action carry a trace of the evidence that triggered it?	Untraced actions make audit impossible and blame political.	"Every action writes a trace: inputs, evidence, output, timestamp."	"You can trust the model's judgment."
10. Could our finance team audit an agent's output without your engineers on the call?	Evidence that requires the vendor to interpret is testimony rather than evidence.	"Finance gets read access to traces and can export them on their own."	"We can prepare a report for audits."
11. When your number disagrees with our CRM, how do we find out which is right?	Discrepancies are certain; unexplained ones destroy trust in both systems.	"A reconciliation view shows the delta and the specific records causing it."	"Our number is the accurate one."
12. How long are decision records retained, and in what format can we export them?	Audits arrive years later; retention decides whether you can answer them.	"Retention terms are in the contract, exportable in open formats."	"We keep what we need for the model."
13. What third-party standard covers your AI management system, and what does it attest?	Certification proves process discipline; overclaiming it signals the opposite.	"ISO/IEC 42001 covers our AI management system, and here is what it does and does not prove."	"We are fully certified and compliant." No standard named.

Autonomy and control: who approves what

The reference pattern to test every vendor against is a four-step loop: the agent proposes, a human approves, the system commits, and anything committed can roll back. RevSure ships this loop as the GTM Harness; whatever your vendor calls theirs, the loop has to exist as software, with permissions, logs, and reversals you can watch. One published field story shows why the approve step earns its keep: a CMO was shown a 4x ROI reallocation of $2M away from LinkedIn and declined it because "LinkedIn is strategic." The human knew something the data did not, and the system recorded the override. A vendor that cannot show you an agent it killed is showing you a roadmap.

Question	Why it matters	A good answer sounds like	A red flag sounds like
14. Which actions can each agent take without human approval, and who sets those tiers?	Autonomy is a permission decision you own, never a vendor default.	"Approval tiers per agent and per action type, configured by your admins."	"The agent is fully autonomous out of the box."
15. Can we change approval thresholds ourselves, mid-quarter, without a support ticket?	Trust changes with evidence; the control surface must move at your speed.	"Self-serve controls, with a change log of who altered what and when."	"Threshold changes go through our success team."
16. Show us a rollback: one real action, reversed end to end.	Reversibility is the difference between an experiment and an incident.	"Here is a committed action, its trace, and the rollback that undid it."	"Rollback has never been necessary."
17. What do approval and override rates look like across your customers in the first 90 days?	Real fleets get overridden while trust builds; the pattern proves live usage.	"Override rates start high and fall as calibration improves; here is the shape."	"Our customers rarely override anything."
18. Tell us about an agent you or a customer turned off, and why.	Every production fleet has retired an agent; a spotless record describes a slideshow.	"One agent was retired over weak downstream conversion; here is the postmortem."	"All of our agents perform well."
19. Exactly what can the agent write back to our systems, and what is it forbidden to touch?	Write access is where agent mistakes become CRM damage.	"A field-level write manifest per agent, enforced by permissions."	"It writes wherever it helps."
20. Who on our side may change what an agent is allowed to do, and is that logged?	An ambiguous permission graph turns agent management into a political fight.	"Role-based permissions with an audit log of every change."	"Anyone with a license can adjust settings."

Economics: what an action costs

This is the section where most vendors go quiet, and the silence is information. Real unit numbers exist in this category. A single personalized email consumes about 31,000 LLM tokens. Published per-million-token rates run from $2.50 for a light model to $70 for a heavyweight one, with the customer choosing the model per agent. An agent run costs $0.25. A credit is $0.02. RevSure publishes these because metered systems can; the questions below test whether a vendor's economics are metered or improvised.

Figure

The prices a real vendor can put in writing

Platform, annual, by unique leads and contacts

Under 50K contacts$20,000

Under 200K contacts$32,000

Under 600K contacts$56,000

Under 1M contacts$76,000

Per action, from published credit rates

Personalized email (20 credits)$0.40

Account research brief (25 credits)$0.50

Event follow-up (40 credits)$0.80

Agent run$0.25

Source: RevSure published pricing model, 2026. Credit rate $0.02; per-action costs derived from published credit rates.

Question	Why it matters	A good answer sounds like	A red flag sounds like
21. What does one unit of work cost: one personalized email, one researched account, one run?	Budgets built without unit costs cannot be governed after signature.	"A published rate card: a personalized email about $0.40, a research brief about $0.50."	"Pricing depends on your use case."
22. How many tokens does a typical action consume, and at which model price per million?	Token appetite decides whether usage pricing stays predictable at volume.	"A personalized email runs about 31,000 tokens; you pick the model per agent from a posted price list."	"You will never need to think about tokens."
23. What is a credit worth in dollars, and what does each agent action cost in credits?	Credits without a posted exchange rate are a price increase waiting to happen.	"One credit is $0.02, and every agent's credit consumption is listed."	"Credits are flexible."
24. How does the platform fee scale with our database size, and where are the tier breaks?	Contact growth should carry a known price before renewal negotiations start.	"Published tiers by unique leads and contacts, from under 50K to over 1M."	"We will price your growth at renewal."
25. What happens when we exceed our usage: alert, throttle, hard stop, or invoice?	Overage behavior at quarter-end is a production incident and a finance incident at once.	"You choose the behavior, and overages bill at posted rates."	"Overages true up at the end of the term."
26. Do unused credits roll over, and do purchased credits expire?	Expiry terms quietly change the effective unit price you modeled.	"Rollover and expiry terms are written into the order form in plain language."	"Talk to your account team at renewal."
27. Will you put an all-in first-year number for our volumes in writing before the pilot?	A worked example forces every hidden cost into daylight.	"Here is a worked example: 200,000 contacts, 10 managed agents, $74,880 all-in."	"Total cost depends on adoption."
28. Which levers must move for your ROI claim to come true, and against what baseline?	An ROI multiple without levers is marketing; with levers it is a testable plan.	"The claim decomposes into named lifts you can check against your own history."	"The ROI is proven across our customer base."

Reliability: revenue intelligence under quarter-end load

LangChain's June 2026 survey of 1,340 agent builders found quality is the number one blocker to shipping agents, 89 percent of teams run observability, and only 52.4 percent run offline evals. The industry watches its agents more than it tests them. GTM adds a scheduling problem: load spikes exactly when mistakes cost the most. And volume metrics can hide a quality collapse completely:

you might get 100 meetings but none of those meetings are converting... you reduce the cost but that agent is still generating the same quality meetings, the conversion of the meetings is not going to improve just because you deployed an agent.

from a working session with a customer team

Question	Why it matters	A good answer sounds like	A red flag sounds like
29. What evaluations run before an agent version ships, and can we see the suite?	Only 52.4 percent of teams run offline evals; demand to be in that half.	"Versioned eval suites per agent, with results you can review."	"We test everything thoroughly." No artifact.
30. When you swap or upgrade the underlying model, what regression testing protects our workflows?	Model swaps silently change behavior in ways a demo never reveals.	"Every model change reruns the eval suite, and regressions block release."	"Newer models are better, so we upgrade automatically."
31. What observability do we get into live runs, and where does it live?	You cannot manage a fleet you cannot watch.	"Run-level dashboards: volumes, outcomes, overrides, cost per run."	"We monitor everything on our side."
32. What happens under quarter-end load, when everything spikes at once?	Reliability at average load is the wrong question for a seasonal business.	"Stated capacity behavior and priority rules under peak load."	"We have never had an outage."
33. How many production runs does your platform execute, and over what period?	Production volume separates a shipped product from a staged one.	"Run counts over time from live deployments, shared under NDA."	"Adoption is growing fast." No number.
34. How do you catch quality drift your volume metrics miss, like meetings that book but never convert?	Activity metrics hide quality collapse; downstream conversion is the test.	"Guard metrics track downstream conversion per agent, with alerts."	"Meeting volume is up across our customers."

Exit: what leaves with you

The strongest argument for a context platform is also the buyer's largest exit risk: the harmonized graph of your accounts, contacts, funnel history, and resolved identities becomes more valuable every month it runs. That asset was built from your data with your budget. Establish who keeps what before signature, while the vendor still wants the logo. Buyers negotiate pilot opt-out clauses routinely; far fewer specify what a wind-down does to agents mid-flight in their systems.

Question	Why it matters	A good answer sounds like	A red flag sounds like
35. If we leave, what do we take with us, and in what format?	Your harmonized data graph is an asset you paid to build.	"Full export of your data, resolved identities included, in open formats."	"You keep whatever is in your source systems."
36. What happens to the agents' accumulated context and configurations at termination?	Agent tuning represents months of institutional learning.	"Configurations and learned settings export as documentation."	"Configurations are part of our platform."
37. What is deleted at exit, on what schedule, and can we get certification of deletion?	Unreturned, undeleted data is your risk long after the relationship ends.	"A deletion timeline in the contract, with written certification."	"Deletion follows our standard policy." Nothing in writing.
38. What are the mid-term exit terms if the deployment misses our pilot criteria?	An exit clause tied to agreed metrics keeps both sides honest.	"Opt-out language tied to the pilot metrics we set together."	"Our standard term is annual, paid up front, no exceptions."
39. At termination, what happens to in-flight actions, scheduled runs, and pending approvals?	A sloppy shutdown can email your entire database on the way out.	"A wind-down runbook: pending actions halt, approvals cancel, owners get notified."	"We turn off your access."
40. Can we export decision traces and audit records after we leave?	Audits of the agent era will arrive after the vendor relationship ends.	"Trace archives export before shutdown and stay readable offline."	"Traces live in the platform."

RevSure's answers: the short version

A handbook like this invites one fair objection: the author sells one of the platforms it interrogates. So here are RevSure's answers to five of the hardest questions, including the uncomfortable one.

What do your agents read (question 1)? Everything the Full Funnel Data Graph harmonizes from native integrations, Salesforce and HubSpot and Marketo through the ad platforms, with identities resolved and funnel definitions made canonical before any agent runs. One published deployment connected 7 data sources in under 30 minutes on day one.

Can every number and action be audited (questions 8 and 9)? Yes. Decision Attribution is computed on the graph, and every agent decision records a Decision Trace a finance team can open without us on the call. The AI management system behind it is certified to ISO/IEC 42001:2023, which attests to management process discipline and does not by itself prove outcomes.

Show us an agent you killed, and your overrides (questions 17 and 18)? The live fleet is public: the Campaign Reallocation agent, the Deal Risk agent, and the Deanonymization agent, with over 1.12M agent runs logged at a single AI software customer. Overrides are visible per deployment in the GTM Harness, and they happen: the declined $2M LinkedIn reallocation above is ours, recorded. The record of that no is worth more than the recommendation was.

The all-in number in writing (question 27)? Published: 200,000 contacts and 10 managed agents runs $74,880 a year, from platform ($32,000), managed deployment ($20,000), and net credits ($22,880), with 80,000 testing credits included.

Time to value, honestly (question 7 and the pilot clause)? Our stated onboarding is live in under 4 weeks. Our field record also holds the other kind of story, and you should hear it from us rather than from a reference call:

we got too focused on configuring the system correctly, but we are not really able to use the system in our day-to-day right now.

an operations leader at a contract lifecycle management company

A marketing leader at a healthcare benefits company put the same friction in contract terms: "we are almost more than halfway through our contract for the year and we're still in setup mode." Setup friction is real in this category, ours included, which is exactly why questions 7 and 38 belong in your process. Hold us to them.

What to do next quarter

Put the 40 questions into your RFP template this week, unchanged, and require written answers from every finalist before demos are scheduled.
Convert questions 6, 8, 16, and 27 into live demo requirements: open a trace, decompose a number, run a rollback, and show the worked cost example on screen.
Refuse any pilot that does not run on your live data. Sample datasets are how the verification gap gets past procurement.
Build a per-action cost sheet for every finalist (email, research brief, enrichment, run) and hand it to finance before the shortlist.
Write pilot exit criteria and wind-down terms into the order form. Questions 38 and 39 cost nothing before signature and everything after.
Score all 40 answers as artifact or adjective, and let the ratio pick your shortlist.

Where this comes from

Built from our work inside enterprise GTM teams: indexed, verbatim working sessions with marketing, RevOps, and revenue leaders, quoted here at descriptor level with permission discipline. Pricing figures are RevSure's published 2026 pricing model. One limit to weigh: the question set was written by a vendor that has to answer it, and it skews toward the failures we have watched up close. Use that. Every question works as hard against us as it does against anyone else on your shortlist.

Frequently asked questions

What questions should you ask before buying AI agents for sales?

Ask 40 questions across six areas: what context the agent reads, whether every number and action traces to evidence, who approves and rolls back actions, what each action costs in tokens and credits, how the vendor tests reliability, and what you take with you at exit. Score vendors on artifacts produced, never on prose.

What is agent washing?

Agent washing is Gartner's term for vendors rebranding conventional automation as agentic AI. In June 2025 Gartner estimated that of thousands of vendors claiming agentic products, only about 130 were building real agents, and forecast that over 40 percent of agentic AI projects would be canceled by the end of 2027. Procurement counters it by demanding production artifacts.

How much do AI agents for sales and marketing cost?

Published reference prices exist. In RevSure's pricing model, a personalized email costs about $0.40, a deep account research brief about $0.50, and an agent run $0.25, with platform fees tiered by database size from $20,000 a year under 50,000 contacts. A worked example prices 200,000 contacts and 10 managed agents at $74,880 a year all-in.

What should an AI agent contract cover at exit?

Four things: export of your harmonized data and resolved identities in open formats, documentation of agent configurations and learned settings, certified deletion on a stated schedule, and continued access to decision traces and audit records after termination. Exit terms are cheapest to negotiate before signature, when the vendor still wants the logo.

The agentic GTM procurement handbook: 40 questions before you sign