Contents
- Fast answer
- Quick facts
- What this covers
- Why it matters
- Where to go next
- Outcomes
- Step by step
- What Salesforce and HubSpot agents actually do and cost
- The data quality floor under any CRM agent
- Native CRM agents versus building your own workflow
- What changed in 2025 and 2026
- The methodology: what the score is counting
- What the native CRM agents cost, verified August 2026
- A worked example: a six-agent real estate team
- The limits, and what to use alongside it
- At a glance
- Mistakes to avoid
- FAQ
- Related guides
Fast answer
A CRM AI agent checklist audits whether your sales data, fields, permissions, activity history, and review rules are ready for agent-assisted research, updates, and follow-up. CRM agents can save sales teams time, but they can also create messy records, wrong promises, duplicate tasks, and awkward customer messages if the CRM is not ready. Start by auditing required fields, checking lead source, stage, owner, company size, ICP fit, next step, close date, and notes quality before adding automation. Define safe CRM actions, deciding whether the agent can draft notes, suggest updates, create tasks, or change stages with approval. Connect customer context from recent emails, calls, tickets, forms, and CRM notes so recommendations are grounded. Add review rules that require approval before customer communication, stage changes, or discount language, then monitor wrong updates and rep edits. Start with low-risk support work such as lead research, call summaries, and follow-up drafts, and let the agent change deal stages only after field rules and source data are reliable.
Facts last verified 2026-08
Quick facts
- Salesforce Agentforce pricing
- $2 per conversation, or Flex Credits at ~$0.10 per standard action
- Flex Credits cost
- $500 per 100,000 credits; a standard action consumes 20 (Salesforce, May 2025)
- HubSpot Breeze Customer Agent
- ~$0.50 (50 credits) per resolved conversation from April 2026
- HubSpot Breeze Prospecting Agent
- ~$1.00 (100 credits) per lead recommended for outreach
- Rep time actually selling
- About 28% of the week (Salesforce State of Sales, 7,700+ reps)
- B2B CRM data decay
- Roughly 25-34% per year (industry benchmarks incl. ZoomInfo)
On this page
What this page covers
A tool visitor should leave with a decision, not just a number: build now, prepare first, choose another workflow, or follow a course path.
- 01CRM data audit
- 02Lead routing
- 03Follow-up drafts
- 04Field updates
- 05Sales approvals
- 06FAQ
Why does this matter now?
CRM agents can save sales teams time, but they can also create messy records, wrong promises, duplicate tasks, and awkward customer messages if the CRM is not ready.
Internal path
Where to go next from this page
These links are part of the A8gent learning and conversion path. Use them to move from concept, to diagnosis, to workflow build, to course.
What you should be able to do after this
- Audit CRM fields
- Define safe actions
- Plan sales handoffs
- Create review rules
How do you do it, step by step?
1. Run the fifteen checks honestly and note the zeros
The section breakdown matters more than the percentage. A zero in guardrails is the finding, whatever the headline says.
2. Fix data hygiene before anything else
Dedupe on the key your leads actually arrive with, fill required fields on active records, and agree what each deal stage means. Every pricing model bills the same for dirty data as for clean.
3. Write the workflow paragraph
Trigger, single named owner, approval rule. If you cannot write it in one paragraph, the workflow is not defined enough to hand over.
4. Split fields into may-update and never-touch
Then turn on the audit log. Without a log, the first disagreement about what the agent did has no evidence on either side.
5. Capture the baseline before launch
Median lead response time, records touched per rep, whatever you will claim improved. A baseline reconstructed afterwards convinces nobody, including you.
What Salesforce and HubSpot agents actually do and cost
The checklist matters more now that agents are a native CRM feature rather than an add-on you build. Salesforce's Agentforce sells autonomous agents for service case resolution, sales development outreach, coaching, and CRM record updates, priced originally at 2 dollars per conversation and, since May 2025, alternatively through Flex Credits: 500 dollars buys 100,000 credits, a standard agent action consumes 20 credits, so an action costs about 10 cents. Salesforce also sells bundled editions for heavier usage, with third-party analyses putting the Agentforce 1 editions around 550 dollars per user per month including large credit allowances.
HubSpot took the outcome route with Breeze. Effective April 14, 2026, its Customer Agent charges 50 credits, roughly 50 cents, per resolved conversation, down from a flat dollar per conversation, and its Prospecting Agent charges 100 credits, about a dollar, per lead recommended for outreach, replacing a recurring per-contact charge. Both agents are available on Professional and Enterprise tiers with a 28-day free trial. Paying per resolution or per accepted recommendation means an idle or unhelpful agent costs approximately nothing, which removes the old objection to experimenting.
The checklist's job is to tell you whether turning these on is safe rather than merely affordable. Native CRM agents inherit whatever your instance contains: your field hygiene, your stale contacts, your inconsistent stage definitions. Vendors price the agent's actions; nobody prices the cleanup of what a confidently wrong agent does with dirty data. Every item on the audit list, required fields, safe actions, context sources, review rules, exists because the native agents are only as trustworthy as the records underneath them.
The data quality floor under any CRM agent
CRM data rots at a measured pace. Industry benchmarks consistently place B2B contact data decay at roughly 25 to 34 percent per year, driven by job changes, company moves, and abandoned addresses; ZoomInfo's own hygiene guidance cites around 30 percent annually. An agent doing lead research or follow-up drafting against a database refreshed two years ago is therefore working from records where a large minority of contact facts are simply wrong, and it will incorporate them fluently, which is worse than a human who would hesitate.
The prize for fixing this is also measured. Salesforce's State of Sales research, surveying more than 7,700 sales professionals, found reps spend only about 28 percent of their week actually selling, with the rest consumed by administration, data entry, and internal process. That is the time CRM agents credibly return: research, logging, note summarization, and follow-up drafting are exactly the non-selling majority of the week. But the same research explains why the checklist front-loads field audits, because agents automating against wrong fields return that time as rework instead.
Practically, scope the cleanup to what the first agent workflow touches rather than launching a CRM-wide hygiene project. If the pilot is follow-up drafts for rep approval, the agent needs reliable contact records, recent activity history, and honest stage values for active deals only. That is usually one or two weeks of focused RevOps work, not a quarter. The checklist's field-by-field structure exists so cleanup is a scoped prerequisite, not an open-ended excuse for delay.
Native CRM agents versus building your own workflow
The build-or-buy line has moved twice since 2024, so run the math on your volume. Native agents win on integration depth and time to value: Agentforce actions live inside Salesforce permissions and audit logs, and Breeze agents read HubSpot context natively, so a checklist-passing instance can pilot in days. At roughly 10 cents per Agentforce action or 50 cents per Breeze resolution, a support desk resolving 1,000 conversations monthly pays about 500 dollars on HubSpot's model, which is trivial against the labor it replaces.
Custom workflows built in n8n, Make.com, or Zapier against the HubSpot and Salesforce APIs win on control and cross-system reach. If your workflow spans the CRM plus a ticketing system, an enrichment provider, and email, a platform workflow with an AI step often fits better than a CRM-resident agent, and you pay platform pricing, n8n's 20 euro Starter tier covers 2,500 executions, plus model usage instead of per-outcome fees. The tradeoff is that you own the logging, review queue, and maintenance the CRM vendor would otherwise provide.
The checklist result should drive the choice. Instances that pass cleanly on fields, permissions, and review rules can adopt native agents and get vendor-grade logging for free. Instances that fail on data quality but need automation now are better served starting with a custom draft-only workflow, because you can constrain it to the clean subset of records and widen scope as hygiene improves. What no result supports is granting any agent, native or custom, write access to pipeline fields before the review rules exist.
What changed in 2025 and 2026
CRM agent pricing restructured itself around outcomes and consumption. Salesforce introduced Flex Credits in May 2025, moving from per-conversation billing toward metered actions, and HubSpot moved Breeze to pay-per-completed-task pricing in April 2026, cutting its Customer Agent price to about half its former per-conversation rate. Both shifts share a message: the vendors are confident enough in agent performance to absorb the cost of failed attempts, and buyers should model costs on successful outcomes at realistic volume rather than on seats.
Agent capabilities widened from answering to acting. Early CRM AI summarized and drafted; the 2025 and 2026 generations update records, create tasks, execute outreach sequences, and hand off between specialized agents. That widening is exactly why checklist items about permissions and review rules got more important, not less. An agent that can only draft can embarrass you; an agent that can edit pipeline fields can corrupt the revenue reporting your forecasts depend on, quietly and at scale.
The economics now punish dirty data more directly. Under outcome pricing you pay when the agent resolves or recommends, so bad data does not just produce bad outputs, it produces billable wrong outcomes: resolutions that reopen and recommended leads your reps reject. Teams that run the checklist first, fixing fields and defining review rules before enabling agents, get the new pricing models working in their favor, paying only for outcomes their own data quality made possible.
The methodology: what the score is counting
Fifteen checks in five sections of three, every check worth the same. Data hygiene asks whether required fields are filled on active records, whether contacts and companies are deduplicated, and whether deal stages mean the same thing to everyone. Access and permissions asks whether API access is enabled on your plan, whether credentials are scoped to only the objects the agent needs, and whether a sandbox or test account exists. Workflow definition asks for a clear trigger, one named owner, and an approval rule. Guardrails asks for a list of fields the agent may update, a list of fields and actions it must never touch, and an audit log. Measurement asks for baseline metrics, written success criteria, and a scheduled review.
The score is simply the number ticked divided by fifteen. The verdict bands are fixed at 85 percent and 50 percent, which in practice means thirteen of fifteen for 'Ready', eight for 'Getting there', and anything below that reads 'Not ready'. Because every check carries equal weight, you can reach 'Getting there' without a single guardrail in place, which is exactly the state most stalled CRM agent projects are in. Read the section breakdown rather than the headline number: five sections at three each is a deliberately unforgiving structure, and a zero in one section is the finding.
The next-steps panel takes the incomplete sections, sorts them by the proportion completed with the weakest first, and shows up to three canned sentences for those sections. It is a nudge rather than a plan, and it is honest about the order: dedupe and required fields before you connect anything, a scoped key and a sandbox so early mistakes stay contained, the trigger and owner and approval rule written in one paragraph, allowed versus never-touch fields plus an audit log, and baseline numbers captured before launch rather than reconstructed after it. Nothing here is scored against your CRM automatically; every answer is your own assessment, and the value of the exercise is that fifteen specific questions are harder to wave away than one general one.
What the native CRM agents cost, verified August 2026
The checklist deliberately says nothing about products, so here is the commercial context it leaves out, taken from our platform research verified on 2026-08-06 in src/content/platforms/depth-builders-4.ts. HubSpot renamed Breeze Agents to Agent Hub and prices on outcomes: $0.50 when Customer Agent resolves a conversation, $1.00 when Prospecting Agent drafts outreach for a lead, $0.10 when Data Agent answers a question about your data. Credits are a cent each, with 3,000 included on Professional, which covers roughly sixty resolutions a month. The prerequisite is the real cost: Agent Hub needs Professional or Enterprise, which starts at $90 per seat per month plus a one-time $1,500 onboarding fee.
Salesforce moved Agentforce to Flex Credits at $500 per 100,000 credits, where a standard action is 20 credits, or ten cents, and a voice action is 30. The alternative is Conversations at a flat $2 each, and the two cannot coexist in the same org. Neither is standalone: a production customer service agent sits on Service Cloud licences, and Service Cloud Enterprise is $175 per user per month billed annually. Unused Flex Credits do not roll over. For a business genuinely running on Salesforce that is defensible; for anyone else the licence is a much larger number than the agent.
The reason this belongs on a readiness page rather than a pricing page is that every one of those models bills for activity on data you have not cleaned yet. An agent that answers from duplicate records bills the same as one that answers from good ones. Prospecting Agent charges when it drafts outreach, not when the outreach works, which is a weak outcome definition on a dirty database. The order that saves money is dedupe, then define, then guardrail, then connect. Doing it the other way around is how a pilot produces a bill and a mess at the same time.
A worked example: a six-agent real estate team
Take a residential team running Follow Up Boss with lead sources feeding in from portals and Facebook Lead Ads, which is the shape our real estate page describes. First pass through the checklist typically scores four or five of fifteen. Data hygiene fails on deduplication, because the same buyer has registered on three portals with two email addresses and a misspelled name. Access passes on API but fails on sandbox, because there is not one. Workflow definition fails on ownership, because 'the team' owns follow-up. Guardrails scores zero, which is normal and is the finding. Measurement fails because nobody wrote down the current speed to lead.
The fix order follows the weakest section, not the easiest. Dedupe on normalised phone rather than email, because portal leads frequently arrive with masked or missing email addresses and the phone number is the reliable key. Then write the paragraph: the trigger is a new lead from any source, the owner is the agent on duty, the approval rule is that the first outbound text sends automatically and anything after the second reply waits for a human. Then the guardrails: the agent may write to lead source, last contacted and next step, and may never change the owner, the price expectation or the stage. Then capture the baseline, which here is median lead response time over the last thirty days.
That sequence takes about a week of part-time work and moves the score to eleven or twelve, which is 'Getting there' rather than 'Ready'. The two checks usually still open are the audit log and the scheduled review, and both are worth closing before launch rather than after, because without them the first argument about what the agent did has no evidence on either side. Speed to lead is the metric that pays in this business, and it is only a result if you wrote down what it was before you started.
The limits, and what to use alongside it
This is a self-assessment. Nothing connects to your CRM, nothing is measured, and ticking a box is a claim rather than a proof. The checklist has no notion of your industry, your compliance obligations, your record volume or your team's size, and it does not distinguish between a two-person operation where the owner is obvious and a twenty-person sales floor where ownership is the entire problem. It also treats all fifteen checks as equal, which is defensible as a design choice but is not how risk works: shipping without an audit log is worse than shipping without a written success criterion.
Pair it with tools that ask different questions. The risk scorer covers what happens when the agent is wrong, which this checklist only gestures at through its guardrails section. The readiness quiz covers whether the business as a whole is ready, which is a broader question than whether one CRM is. If the CRM itself is about to move, do the migration first and the agent second, because an agent built against a system you are leaving is work you will throw away. And for the first fourteen days after launch, the onboarding planner covers the sequence this checklist stops at.
One routing note we will make plainly: if your CRM holds only a third of the context that matters, because bookings live in a scheduling system and invoices live in accounting software, a native CRM agent will give confidently partial answers however clean the CRM is. That is not a data hygiene problem and no amount of ticking fixes it. In that case a workflow built across systems on a general automation platform is the better shape, and our done-for-you automation services exist for exactly that case.
At a glance
Native CRM AI agent pricing (verified July 2026)
| Offering | Pricing model | Cost | Availability |
|---|---|---|---|
| Salesforce Agentforce (conversations) | Per conversation | $2 per conversation | Salesforce orgs with Agentforce |
| Salesforce Agentforce (Flex Credits) | Per metered action | About $0.10 per standard action; $500 per 100,000 credits | Launched May 2025 |
| HubSpot Breeze Customer Agent | Per resolved conversation | 50 credits, about $0.50 | Pro and Enterprise, 28-day trial |
| HubSpot Breeze Prospecting Agent | Per recommended lead | 100 credits, about $1.00 | Pro and Enterprise, 28-day trial |
| Custom workflow (n8n plus CRM API) | Platform plus model usage | From EUR 20/mo for 2,500 executions plus model costs | Any CRM with an API |
What mistakes should you avoid?
- Connecting an agent to a CRM with unresolved duplicates, then paying per outcome for answers drawn from the wrong record.
- Reaching 'Getting there' with zero guardrails ticked and treating that as permission to launch.
- Letting the agent write to fields that feed forecasting or commission before a month of supervised running.
- Buying a CRM subscription tier to unlock an agent, when the licence costs more than the agent will ever save.
- Skipping the sandbox because the CRM plan makes one awkward to create.
- Launching without a baseline, then arguing about whether anything improved.
- Expecting a CRM-native agent to answer questions whose data lives in your scheduling or accounting system.
FAQ
- What score means we are ready?
- Thirteen of fifteen, which is the 85 percent band the tool calls 'Ready'. Eight of fifteen reaches 'Getting there' and anything below is 'Not ready'. Because every check is worth the same, you can hit the middle band with no guardrails at all, which is why the section breakdown matters more than the headline. Our own bar before connecting an agent to live records is simpler: all three guardrail checks and all three data hygiene checks, whatever the total says.
- Why is data hygiene first?
- Because every commercial model in this category bills for activity regardless of whether the underlying record was right. HubSpot's Prospecting Agent charges $1.00 when it drafts outreach for a lead, not when the outreach works, so a duplicated contact list generates duplicate drafts and a duplicate bill. An agent inherits whatever mess lives in the CRM and then acts on it faster than a person would. Deduplicating and filling required fields is unglamorous and it is the highest-return hour in the whole project.
- What do CRM AI agents actually cost in 2026?
- Verified in August 2026: HubSpot's Agent Hub, formerly Breeze Agents, bills $0.50 per resolved conversation, $1.00 per lead drafted by Prospecting Agent and $0.10 per Data Agent answer, but requires Professional or Enterprise from $90 per seat per month plus a one-time $1,500 onboarding fee. Salesforce Agentforce bills either $2.00 per conversation or about $0.10 per action through Flex Credits at $500 per 100,000, on top of Service Cloud licences at $175 per user per month billed annually. The prerequisite licence is usually the larger number.
- Do we really need a sandbox?
- You need somewhere a mistake is cheap. On Salesforce that is a sandbox; on HubSpot it is often a test portal or a small set of clearly marked test records; on a lighter CRM it may be a duplicate pipeline. The requirement is not the feature name, it is that the first two weeks of agent behaviour can be watched without a real customer receiving a wrong message. If none of those is available, restrict the agent to read-only and drafting until the correction rate is near zero.
- Which fields should an agent never touch?
- Anything that feeds forecasting, commission or a contractual commitment. In practice that means deal stage, amount, close date, owner, and any custom field a report depends on. Safe starting fields are activity logs, last contacted, next step, research notes and enrichment data, because a wrong value there is visible and cheap to correct. Widen the list only after a month of supervised running with a low correction rate, and keep the audit log on permanently rather than for the pilot only.
- Can we skip the baseline if we are confident it will help?
- You can, and then you will spend the following quarter unable to prove it did. Baselines take an afternoon: median lead response time, records touched per rep per week, follow-ups sent, deals with a next step set. The reason this is a check rather than advice is that reconstructing a baseline after launch is not possible in most CRMs without export archaeology, and a claimed improvement with no before number is the fastest way to lose budget for the second workflow.
- What if our CRM is not where most of our data lives?
- Then a CRM-native agent is the wrong shape and no score on this checklist will fix that. An agent grounded in a system holding a third of the relevant context gives confidently partial answers, and you will have paid a licence for the privilege. If bookings sit in a scheduling system and invoices in accounting software, build the workflow across systems on an automation platform instead, or have it built for you. The checklist still helps: the hygiene, guardrail and measurement questions apply to any system of record.
Sources & further reading
Was this page helpful?
