What makes an AI sales agent different from AI features
Most sales software now has AI in it. That does not make it an agent. The test is who decides the next step. An assistant answers when you ask (summarize this call, draft this email). An automation runs the same fixed steps every time (if a form is filled, send sequence A). An agent is given a goal, reads the situation, picks the steps and takes them, then checks the result.
That is also the meaning behind "agentic AI for sales": AI that acts, not just answers. Gartner warns that the label is stretched. It describes "agent washing", the rebranding of assistants, RPA and chatbots as agents, and estimates that only about 130 of the thousands of agentic AI vendors are real (Gartner, June 2025). Ask any vendor what its agent decides on its own. If the answer is "nothing, a user triggers it", you are looking at an assistant.
| AI assistant | Workflow automation | AI sales agent | |
|---|---|---|---|
| Who starts it | A person, each time | A fixed trigger | A goal plus triggers it watches for |
| Who picks the steps | The person | Whoever built the workflow | The agent, within the rules you set |
| Handles the unexpected | Only if asked | No, it breaks or skips | Yes, it adapts or escalates to a person |
| Typical output | A summary or a draft | The same action every time | A completed task, logged with its reasoning |
An AI SDR is one kind of AI sales agent, aimed at the top of the funnel. This page covers the whole category.
The eight jobs AI sales agents do
Buying "an AI sales agent" is like hiring "a salesperson": the question is which seat. These are the eight jobs vendors package as agents today, what triggers each one and where a human should stay in the loop.
| Job | What the agent does | Typical trigger | Where a human stays | Go deeper |
|---|---|---|---|---|
| Prospecting and research | Builds account and contact lists from your ICP, enriches them, scores fit and watches for buying signals | A new target segment, a hiring or funding signal | Defining the ICP; spot-checking scores | GTM 2.0 |
| Outbound engagement | Writes and sends personalized email, LinkedIn and SMS steps, handles replies, books meetings | An account crosses a signal threshold | Approving messaging; owning replies that need judgment | AI SDR |
| Calling | Runs the dialing: lines, voicemail, phone trees and AI call screeners, then hands live people to a rep | A call list or a call step in a sequence | The live conversation | AI dialer |
| Inbound response | Contacts new form leads fast, qualifies them, books or routes to the right owner | A demo request, form fill or inbound call | Complex questions; the discovery call | Speed to lead |
| Deal management | Scores deal health, flags risks, recommends and runs the next step, drafts follow-ups | A call ends, a champion goes quiet, a close date slips | Strategy, pricing, negotiation | Next best action |
| CRM upkeep | Logs activity, updates fields from what was said, fixes duplicates and stale records | Every call, email and meeting | Field definitions; reviewing bulk changes | Agentic CRM |
| Nurture | Keeps not-ready leads and closed-lost deals warm until a signal says now, then hands back to a rep | A lead goes cold, a deal is lost, a no-show | The re-engagement conversation | Lead nurturing |
| Retention and expansion | Watches customer accounts for churn risk and growth signals, alerts owners, opens expansion opportunities | Usage drops, a renewal approaches, seats near the plan limit | The save call and the upsell conversation | Net revenue retention |
Every row has a point where a person talks to the buyer. Agents that skip it tend to fail in public, and AI voice calls carry consent rules (see the consent section of our AI SDR guide).
Autonomy levels: suggest, approve, auto
Every agent action can run at one of three levels. Suggest: the agent shows the action and its evidence, and a person decides. Approve: the agent prepares the action (the email, the field change, the call list) and a person clicks to release it. Auto: the agent acts and logs it, and a person reviews after the fact.
Set the level per action, not per tool. The same deal agent can update a next-step field on auto and still need approval to email the economic buyer. A sensible starting point:
| Job | Start at | Promote to auto when | Keep a human on |
|---|---|---|---|
| Prospecting and research | Auto | Already there; spot-check weekly | ICP changes and exclusions |
| Outbound engagement | Approve | Reps edit fewer than one in ten drafts for a month | New segments, new messaging, sensitive accounts |
| Calling | Auto for dialing logistics | N/A | Every live conversation |
| Inbound response | Auto for first contact and booking | Routing is right on a weekly audit | Enterprise and complex inquiries |
| Deal management | Suggest | Acceptance rate holds for a full cycle | Pricing, discounts, commitments |
| CRM upkeep | Approve for bulk changes, auto for logging | Reversal rate stays low | Stage and amount changes on large deals |
| Nurture | Approve | Replies stay positive and opt-outs low | Re-engaging former champions |
| Retention and expansion | Suggest | Owners act on most alerts | Save plays on top accounts |
The thresholds are judgment, not research; write your own down before the pilot. For a finer ladder (suggest, draft, act and notify, act silently) and how to measure acceptance and reversal rates, see next best action in sales.
What AI sales agents do well, and where they fail
Where agents are strong
- Watching everything at once. No rep can monitor 3,000 accounts for hiring posts, funding news and pricing-page visits. An agent can, every day.
- Speed. Contacting an inbound lead within minutes, sending a recap while the call is fresh.
- Consistency. The tenth follow-up gets the same care as the first.
- Data entry. Fields updated from what was actually said on the call, not from memory on Friday afternoon.
- Dialing logistics. Ringing several lines, navigating phone trees, answering AI call screeners, dropping voicemail. A human adds nothing to that part of the call.
Where agents fail
- Garbage in. An agent reasoning over stale close dates and blank contact roles produces confident, wrong advice. Gartner's advice to sales leaders puts data quality before scaling (Gartner, November 2025).
- Judgment and persuasion. Reading a skeptical buyer, handling an objection live, knowing when to push. Humans still win here.
- Errors at scale. One wrong assumption repeated across 2,000 emails is a brand problem.
- Alert overload. Agents that surface every signal bury reps. Gartner predicts that by 2028 AI agents will outnumber human sellers tenfold, yet fewer than 40% of sellers will report that agents improved their productivity, and warns that layering more tools onto complex workflows risks burnout (same source, a prediction rather than measured data).
- Unclear value. Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls (Gartner, June 2025). A pilot with a defined metric is the cheapest protection.
How to evaluate an AI sales agent: a checklist
Run every shortlisted agent through the same eleven questions. The answers you cannot get in writing are your risks. If you are comparing vendors, our best AI revenue agents roundup applies these criteria to named products.
- Which job does it do? Name the one job from the table above you are buying it for. "It does everything" usually means it does nothing deeply.
- What does it decide on its own? Ask for a list of the actions it can take without a person, and confirm each can be set to suggest, approve or auto.
- What does it read? CRM fields only, or calls, emails, product usage and outside signals too? The more context, the better the decisions.
- Where does it write? Native two-way write-back to your CRM objects and custom fields, or a separate database you have to reconcile?
- Can you see its reasoning? Every action should come with the evidence behind it, one click away.
- Is there an action log and an undo? Who did what, when and why, with a way to reverse it.
- How does it hand off? How fast a person takes over when a buyer is interested, confused or upset, and with what context.
- Who talks to the buyer? Writing, AI voice or a human rep. For AI voice, ask how consent is captured before each call.
- What does setup really take? Who maps your CRM, builds the rules and owns the agent after launch: you, the vendor or a partner.
- Show me unedited output. Real emails, call summaries or a week of CRM changes from a current customer, not a demo account. Ask too whether your data trains shared models.
- How is it priced, and what is the exit? Seat, credit, action or engagement, the minimum term, and what happens to unused credits.
How to pilot an AI sales agent in six weeks
- Week 0: pick one job and one metricOne job (say, nurture of closed-lost deals) and one outcome metric (qualified opportunities created). Write the success threshold and the exit clause into the order form before anything is switched on.
- Week 0: set up a holdoutHalf the deals go to the agent, half stay with the current process. Without a holdout you measure the season, not the agent.
- Weeks 1-2: run at suggest or approveEvery action passes a person. Log what they accept, edit and reject, with a reason. This is your error rate before any risk is taken.
- Weeks 3-4: promote what earned itMove the actions with high acceptance to auto. Keep the rest on approve. Review the action log twice a week for anything a person would not have done.
- Weeks 5-6: measure and decideCompare the agent group with the holdout on the outcome metric, plus cost and hours saved. Decide to expand, extend or exit, using the threshold you wrote down in week 0.
A worked example with hypothetical numbers: 400 closed-lost deals, 200 to the agent and 200 to the holdout. If the agent group produces 14 qualified opportunities and the holdout produces 6, the agent added 8. At a hypothetical $3,000 for the pilot, that is $375 per added opportunity, which you can compare with your current cost per opportunity.
How AI sales agents are priced: four models
Agent pricing is still settling, and many vendors mix models. These are the four you will see, with an example of each from a vendor's own page (checked 2026-11-18). Prices change often, so treat them as examples of the model, not quotes.
| Model | How it works | Example (vendor's own page) | Watch out for |
|---|---|---|---|
| Per seat | A monthly fee per user, often with agent usage included or capped | Salesforce lists an Agentforce add-on at $125 per user per month for Sales, Service and Field Service (Salesforce) | Paying for seats when the agent, not the user, does the work |
| Credits | You prepay a pool of credits; each action or task burns a set number | Salesforce Flex Credits at $500 per 100,000 credits, "pay per action"; unused credits do not roll over (Salesforce) | Unpredictable burn from always-on monitoring; expiring credits |
| Per action or outcome | You pay when a defined task or result is completed | HubSpot's Prospecting Agent charges $1 per lead recommended for outreach, paid in HubSpot credits (HubSpot, April 2026); Rox meters work in "Agent Actions" (Rox) | How the outcome is defined; a "recommended lead" is not a meeting |
| Per engagement | A scoped project price, usually after a proof of value, often with an engineer who sets up the agents | Not usually published; common for agents that need deep CRM mapping | Scope creep; make the deliverables and exit terms explicit |
Whatever the model, convert it to cost per qualified opportunity using your pilot numbers. A cheap per-action price on a low-quality action is expensive. Count the hidden costs too: enrichment data, approval review time and CRM admin.
Frequently asked questions
What is an AI sales agent?
Software that takes a sales goal, decides the steps and carries them out across your tools: research, outreach, dialing, follow-up and CRM updates. Unlike an AI assistant, it does not wait to be prompted for each step.
What is the difference between an AI sales agent and an AI SDR?
An AI SDR is one kind of AI sales agent, focused on prospecting and booking first meetings. AI sales agents also cover inbound response, deal management, CRM upkeep, nurture, retention and expansion. See our AI SDR guide for the SDR-specific detail.
Can AI sales agents make phone calls?
Some vendors offer AI voice agents, often aimed at inbound and opted-in conversations, because the FCC treats AI-generated voices as artificial and those calls generally need prior consent. Many teams use agents for the dialing logistics instead and keep a human rep on every live conversation.
How much autonomy should an AI sales agent have?
Set it per action. Research and activity logging can usually run on auto from day one. Buyer-facing emails should start on approve, and pricing or strategy should stay on suggest. Promote an action only when its logs show it is right often enough.
How much do AI sales agents cost?
It depends on the model: per seat, prepaid credits, per action or outcome, or a scoped engagement. Published examples range from $1 per recommended lead to $125 per user per month, but compare options on cost per qualified opportunity, not list price.
- 1.Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (prediction; agent washing estimate), checked 2026-11-18 (2025)
- 2.Gartner, Gartner Predicts By 2028 AI Agents Will Outnumber Sellers by 10X, Yet Fewer Than 40% of Sellers Will Report AI Agents Improved Productivity (prediction), checked 2026-11-18 (2025)
- 3.Salesforce, Agentforce pricing (per-user add-on, Flex Credits), checked 2026-11-18 (2026)
- 4.HubSpot, Customer Agent and Prospecting Agent: now you pay when the task is complete, checked 2026-11-18 (2026)
- 5.Rox, Pricing (Agent Actions), checked 2026-11-18 (2026)
- 6.FCC, FCC Makes AI-Generated Voices in Robocalls Illegal (2024)
Spend your reps' day on conversations, not ringing phones.
TwinsAI parallel-dials up to 10 lines at once, validates every number with the carrier before it rings, and bridges a live prospect to your rep in under 400ms. It works inside HubSpot, Salesforce, Close, Attio, Outreach, Zoho, HighLevel, Pipedrive, Apollo and SmartMoving, with no seat minimum.