An AI sales agent is software that takes a sales task, such as replying to an inbound lead, briefing a call, drafting the recap or keeping CRM fields current, and carries it out within rules you set. Today agents are dependable at research, drafts and record updates a person can check. They fail at commercial judgment and at anything built on bad CRM data.
Most of what ranks for this term is a vendor list, and most of those lists stop at prospecting. Prospecting has its own guides here: our AI BDR guide covers outbound, and our AI SDR entry covers qualifying inbound leads. This page covers the rest of the sales cycle, where agents now do the most hours of work: answering demand fast, preparing and following up on calls, and keeping the record clean. We run these processes for B2B companies with agents doing the repeat work and a person approving anything that reaches a buyer or changes a deal.
Which sales jobs can AI agents run today?
Six jobs, each now shipped as an agent by at least one major sales platform. The vendor column lists what each vendor says its agent does on its own product page. We have not tested these claims. Vendor pages checked on 2 November 2026.
| Job | What the agent does | Vendor example | What a person keeps |
|---|---|---|---|
| Inbound response | Researches a new lead, routes it, replies or drafts a reply, books the meeting | Salesforce Agentforce routes inbound leads, answers questions from company content and books meetings | Any reply that mentions price, terms or roadmap |
| Account and call research | Builds a brief on the account and the people on the call | Salesloft lists Account Research and Person Research agents | Checking the facts the seller will repeat out loud |
| Call recap and follow-up | Drafts the recap email and the notes after each call | Dust reports its customer PayFit cut follow-up email time from 7 minutes to 2 | Sending the recap, and the next step it promises |
| CRM updates | Fills stage, next steps and notes from calls and emails | Salesforce updates opportunity fields such as Stage and Next Steps, in a suggestive or an autonomous mode | Approving changes to stage, amount and close date |
| Stalled-deal alerts | Flags deals that went quiet and suggests a reason to re-engage | Salesloft’s Stalled Deals Insight surfaces stalled deals for managers | Deciding whether the deal is worth another try |
| CRM questions | Answers “what happened on this account?” from records, calls and documents | HubSpot’s Data agent pulls insights from the CRM, calls and documents | Reading the answer against the record before acting |
The pattern is the same one we see in outbound and in AI marketing agents: agents are strong where the output is a draft or a flag that someone checks before it matters, and risky where the output reaches a buyer or rewrites a deal record directly.
How much sales work can agents take on?
A lot, because most of a seller’s week is not selling. Salesforce’s 2026 State of Sales report, a survey of 4,050 sales professionals in 22 countries run in August and September 2025, found:
- The average seller spends 40% of their time selling.
- 54% of sellers say they have used agents, and nearly 9 in 10 plan to by 2027.
- Sellers expect agents to cut prospect research time by 34% and email drafting by 36%.
- 51% of sales leaders using AI say disconnected systems are slowing their AI projects.
Figures re-checked on 2 November 2026. Salesforce sells sales agents, so read the adoption numbers as a vendor’s survey. The 40% figure is the useful one: it puts a size on the research, notes and record-keeping that agents now draft.
Where do AI sales agents fail?
Five failures show up in practice. A better model fixes none of them. Each is fixed by data, permissions or a review point.
They act on whatever the CRM says. An agent that briefs a call or routes a lead reads your records first. If the account owner is wrong or the deal stage is six weeks stale, the agent repeats the mistake at speed. In Validity’s State of CRM Data Management in 2026, a survey of 500 marketing professionals, only 21% said their CRM data was very well prepared to support AI, and 62% reported losing revenue directly to poor CRM data quality. Clean the fields an agent reads before you switch it on. (Figures re-checked on 2 November 2026.)
They write next steps the buyer never agreed to. A recap agent summarizes a call from the transcript. If the buyer said “send me something and I’ll look,” the draft can turn that into “agreed to a technical review next Tuesday.” A buyer who receives a recap with a commitment they did not make stops trusting the thread. The seller sends the recap, not the agent. Our sales call follow-up email guide lists what a recap needs.
They overwrite deal records quietly. An agent that moves a deal to “Negotiation” because pricing came up once changes your forecast without anyone noticing. Salesforce’s own page offers a suggestive mode in which reps review, edit and approve opportunity updates instead of letting the agent write them. Use it for stage, amount and close date.
They answer commercial questions they should hand off. Discounts, contract terms, security reviews and roadmap dates commit the company. An inbound agent that answers “yes, we can do annual billing with a 20% discount” has made an offer. Give the agent a list of topics it must route to a person.
Many “agents” are rebranded chatbots. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls. It also estimated that only about 130 of the thousands of vendors selling agentic AI offer real agentic capabilities. Ask to see the agent take a step it was not scripted to take.
Which actions should an AI sales agent take without approval?
Set the permission by what a mistake costs and who sees it first.
| Action | Let it act alone | Draft for approval | Keep with a person |
|---|---|---|---|
| Log calls, emails and meeting notes | Yes | ||
| Research brief before a call | Yes, the seller reads it before the call | ||
| Confirmation and calendar link to an inbound lead | Yes | ||
| Personal reply to an inbound lead | Yes | ||
| Call recap email to the buyer | Yes | ||
| Change deal stage, amount or close date | Yes, suggestive mode | ||
| Re-engagement email on a stalled deal | Yes | ||
| Pricing, discounts, terms, roadmap | Yes | ||
| Closing a deal as lost | Yes |
Widen a permission only after the drafts for that action have needed few edits for several weeks.
What does an AI follow-up agent do on a real deal?
Illustrative, not a benchmark. The company is fictional.
Brightwell Freight sells dispatch software to trucking companies with 50 to 300 drivers. A seller has a demo on Tuesday with an operations director at a 120-truck carrier.
| When | What the agent does | What the seller does |
|---|---|---|
| Tuesday, 7:30 am | Sends a one-page brief: the carrier’s fleet size, two open dispatcher roles, the tools on its careers page, and that a colleague downloaded a pricing guide last month | Reads it, notices the download, plans to ask about it |
| Tuesday, 11:00 am | Demo | Runs the call |
| Tuesday, 11:40 am | Drafts the recap from the transcript and proposes moving the deal to “Evaluation” | Deletes a line saying the buyer “agreed to a trial” (they said they would think about it), approves the stage change, sends the recap |
| Following Friday | Logs that the buyer opened the recap twice and forwarded it to a second address | Nothing to do |
| Day 9, no reply | Flags the deal as stalled and drafts a note tied to a new job post for a fleet manager | Edits the opening line and sends it |
The agent did about an hour of work across the week. The seller made three decisions, and one of them caught a commitment the buyer never made.
How do you choose an AI sales agent?
Ask every vendor these questions and ask to see the answer in the product.
- Can every action that reaches a buyer or changes a deal be held for approval? If approval resets or costs extra, treat it as missing.
- Does each brief and recap show where its facts came from? A seller cannot check a claim with no source.
- Which CRM fields does it read and which can it write? Get the list, and turn off write access to fields that drive the forecast until you trust it.
- Does it log what it did? Every email, update and flag should be visible on the record, with the agent named as the author.
- What does it route to a person? It should hand off pricing, legal and security questions by rule.
- What is the pricing unit? Seats, conversations, credits or a platform bundle. Price it at month-three volume.
Should you build agents, buy a platform, or use a service?
Build when you have someone who can maintain the prompts, the CRM permissions and the logs. Buy when your CRM vendor’s agents cover the job and your data is clean enough for them. Use a service when you want the process running before you hire for it.
Growigami runs these sales processes as services, each proven by hand first and then run by AI every week. With inbound lead follow-up, each inbound lead is researched, scored and sent to the right seller with a reply drafted, at any hour. With deal follow-up, we brief each call and draft the recap, and stalled deals get a new reason to talk. Automations connect the tools around your CRM, and we fix any break before your team feels it. In every case the seller checks what goes to the buyer.
Frequently asked
What is an AI sales agent?
Software that takes a sales task and carries it out within rules you set: replying to an inbound lead, researching an account before a call, drafting the recap after it, updating CRM fields or flagging a deal that went quiet. Unlike a fixed workflow, it decides its own steps toward the task, which is why its output needs a review point.
Is an AI sales agent the same as an AI SDR?
An AI SDR is one kind of AI sales agent: the one that follows up on inbound leads and books meetings. An AI BDR is the outbound kind. AI sales agents also cover the work after the first meeting: call briefs, recaps, CRM updates, stalled-deal alerts and quotes.
Can an AI sales agent update the CRM on its own?
It can, and some vendors offer both modes. Salesforce describes a suggestive mode, where reps review, edit and approve opportunity updates, and an autonomous mode, where updates happen automatically. Start in the suggestive mode for stage, amount and close date, and let the agent write activity logs and call notes on its own.
Will AI sales agents replace sales reps?
They are taking over the typing: research, first drafts, notes and record updates. Salesforce's 2026 State of Sales survey found the average seller spends 40% of their time selling, so there is a lot of non-selling work to take. Discovery, negotiation and the judgment about which deals deserve time stay with people.
How much do AI sales agents cost?
Pricing varies by unit: per seat, per conversation, per credit or bundled into a CRM plan. HubSpot, as of November 2026, includes its agent builder in its Starter, Professional and Enterprise editions and charges custom agents in HubSpot Credits each time they run. Model the cost at the volume you expect in month three, and add the hours of the person who reviews the agent's output.
Which AI sales agent should a team deploy first?
The one whose mistakes are cheapest to catch. Call briefs and recap drafts are good first jobs: a seller reads each one before it matters, so an error costs a minute. Agents that change deal records or send to buyers without review should come later, once the drafts need few edits.
