“Oh yeah, we already have AI agents running in sales.”
I hear some version of this in first meetings with new clients from time to time. Usually it means a salesperson has a chat assistant open in a browser tab, or the CRM vendor has switched on a feature with “agent” in its name.
AI sales agents can already handle most of the preparation around a deal: account research, lead triage, meeting follow-up, CRM updates and first drafts. The decisions that win the deal, and set its price, still belong to a person. Below are both lists, with examples from our own sales work and from client projects.
What counts as an AI sales agent
The word “agent” covers three quite different things right now.
A chat assistant. ChatGPT, Claude or Copilot answers a salesperson’s questions. It’s useful, but it waits for the next prompt and does nothing on its own.
A workflow with AI in one step. A fixed sequence such as “new lead arrives, enrich it, score it, route it”. A language model handles the step that needs reading or writing. Most of what is sold as an AI sales agent today falls into this group.
An agent. A system that plans its own steps, chooses which tool to use and when, and changes the plan when new information comes in.
The second group gets dismissed too easily. In sales, a well-built workflow is often the better choice, because it does the same thing every time and you can see exactly where it went wrong. The label matters less than one practical test: can the system read and write the data your salespeople work with?
On the market, these three things come packaged as five kinds of AI sales tools:
| Type | Examples | Good for | Check before you buy |
|---|---|---|---|
| General AI assistant | ChatGPT, Claude, Microsoft Copilot | Research, drafts and meeting prep, one prompt at a time | Whether it can reach your CRM at all |
| Agent built into your CRM | HubSpot Breeze, Salesforce Agentforce, Dynamics 365 Sales agents | Tasks inside the CRM you already use | It sees only that vendor’s data |
| Standalone AI SDR | 11x, Artisan, Salesforge | High-volume outbound sequences | What it sends without a person approving it |
| Workflow tool with AI steps | Clay, n8n, Make | Enrichment, scoring and routing with your own rules | Who maintains it when a step changes |
| Custom build on your own data | Rex and our proposal tool, both below | A high-value task nobody sells off the shelf | Documentation and ongoing maintenance |
What AI sales agents handle today
Account research and planning
A proper account plan used to take us 2-3 days per account. Someone researched the company, found the decision-makers, read their profiles, built a strategy and drafted the outreach. Now one command pulls the CRM data, finds contacts through Apollo, builds an expansion strategy and drafts the emails in Outlook. The salesperson still decides which accounts deserve a plan. The full workflow is in our article on Claude Code skills for sales.
Finding the right reference
Nordic Healthcare Group works across 15 countries, and its HubSpot holds hundreds of won opportunities. Finding the right reference for a new proposal meant posting in Teams channels and waiting hours or days for an answer.
We built Rex, a retrieval assistant on Azure that consultants use through Teams chat and a light web app. Rex finds the most relevant references, explains why each one fits the prospect and links to the source material for checking. Reference searches went from days to minutes. The details are in the Nordic Healthcare Group case.
Inbound lead triage
Singa handles hundreds of inbound leads a week across the US, the Nordics and the UK. Every lead now passes through an AI pipeline that enriches the contact, qualifies it against Singa’s ideal customer profile and routes it in HubSpot in under 2 minutes, with reasoning notes attached. The 45-minute morning sorting session dropped to single-digit minutes, and reps call the best leads the same day. See the Singa case.
Meeting follow-up
After every sales meeting, someone used to sit down and work out what was agreed, who does what and by when. In our own process this now runs by itself. Fireflies transcribes the meeting, Claude checks whether it was a sales meeting, picks out the action items and creates the tasks in the CRM.
Keeping the CRM up to date
Everyone agrees the CRM should be current, and nobody wants to update it. We built a Claude Code skill that reads the last two weeks of client meetings from the calendar and compares them with active proposals and the CRM. It shows a list of what’s missing: contacts, outdated deal stages and meetings nobody logged. A person approves the list, and the skill fixes the gaps.
First drafts of proposals
Our proposal tool writes the first draft from meeting notes and earlier proposals, roughly 20 times faster than a person writing from scratch. Proposals go out as web pages instead of slide decks, so the client gets a link and there’s only ever one version.
Drafting replies
A positive reply that waits three days has already cooled. For one client, replies to outbound campaigns arrived faster than the sales team could answer them. Now Claude writes a first draft of each answer from instructions and facts the team keeps in a separate table, so they can change what the drafts say without touching the automation. A person reviews every draft in Slack before anything is sent.
What still needs a human
Anything with a price on it
Our proposal tool once gave a client a €10,000 discount on its own. The salesperson didn’t check the proposal before it went out. The deal closed, €10,000 cheaper.
We still write every proposal with the tool, because the time saved is far bigger than the occasional mistake. But the tool doesn’t notice its own errors or take responsibility for them, so the check before sending belongs to a person.
Conversations past the second round
In our reply automation, the third round of a conversation skips the AI completely and moves the thread to a person. If the reply points us to a colleague, no draft gets written at all. By the third exchange, the buyer is usually asking about something specific to their situation.
Deciding who to sell to
The machine can research any list you give it. It can’t decide which customers are worth having: which ones fit what you deliver best, which ones stay and which ones cost more to serve than they bring in. That definition, down to account names, has to come from people before any prospecting system is switched on.
Approving new contacts
When our system finds a new contact, the machine filters out the obvious no’s. Everything else goes to a list of pending approvals, where a person accepts or rejects each contact. That step is what keeps our CRM data clean.
Noticing when the system is wrong
An AI sales agent makes mistakes with the same confidence as everything else it does. Someone has to see them and feed them back.
At Singa, every lead arrives in HubSpot with the AI’s reasoning written next to it, so a rep can see why a lead was scored the way it was. The team has a dedicated Slack channel for flagging errors, and the prompts get refined every week based on real data. Without that loop, the system stays exactly as good as it was on launch day.
The relationship and the large deal
Peter Thiel makes a useful point in Zero to One: deal size decides how you can sell. Small deals need a product that sells itself, advertising or automation, because a salesperson is too expensive. Large deals need a person, because several buyers are involved and the decision takes months. In between sits a dead zone of deals too big for automation and too small for personal selling, and many B2B companies live there.
AI is narrowing that dead zone. When the machine handles the lists, the research and the first messages, an hour of a salesperson’s time goes much further. That makes personal selling affordable in smaller deals than before, for teams that have built the machine around their salespeople.
How to judge an AI sales tool
“Why don’t we use HubSpot Breeze? We already have it.”
This question comes up in client meetings every now and then. Having a tool close at hand is a weak reason to choose it. Every software company is building its own agent right now, and each one sees only its own data. Nobody wants to open twelve chat windows a day.
Choose the tool for the job and keep control in one place. Four questions help with any AI sales tool:
- Can it read and write your CRM through an API, without anyone logging in and clicking?
- Where does a person approve, and can you see that step?
- Does it log every message it sends, including the calls a person makes by hand?
- What happens when a conversation reaches its third reply?
We chose Twenty as our own CRM because of the first question. It’s fully open through its API, so our agents read and write there directly.
The large vendors are moving the same way. In August, Salesforce and Anthropic announced Claudeforce. Its first part is a Salesforce plugin for Claude with 37 ready-made sales skills, such as meeting prep, deal health reviews and pipeline reviews. The salesperson works in Claude, and every action still runs through Salesforce, so the company’s own rules apply. In April, Salesforce had already released Headless 360, which offers the platform’s capabilities as APIs, MCP tools and a command line instead of through the user interface. When the biggest CRM vendor opens its platform to outside agents, choosing tools by how well agents can use them is no longer a niche concern.
Where to start with AI for sales
Pick one task from the first list that your team repeats every week and whose output you can check. Time it for two weeks before you change anything. Then automate it, keep a person at the approval step and time it again.
That before-and-after measurement is also where our AI automation projects start.
This week, write down how many hours your team spends on account research, CRM updates and proposal drafts. The biggest number shows where your first AI sales agent should go.
