AI for Sales: Agents, Add-ons, Prompts

In sales, use AI to prepare, not to send: research an account, brief the rep before a call, draft a first email from facts the rep supplies and write up the meeting. A rep reads and sends every message. On Claude's published prices a draft costs well under a cent, and the real costs are your CRM, your email tool and rep time. Checked 31 October 2026.

The job in one table

What the job is

Sales teams spend much of their day on work that is not selling: reading about a company before a call, writing the same kind of email forty times, typing notes into a CRM. AI is good at that reading, drafting and summarising. It is not good at knowing your buyer, judging whether a deal is real or keeping a promise.

The sensible split is that the agent prepares and the rep decides. A rep who walks into a call with a one-page brief of recent news, the buyer's role and three likely questions sells better than one who read ten tabs. A rep who sends fifty unread AI emails sells worse and may break the law.

Which agent and tools fit

Like support, sales work is usually a small pipeline, not a coding task. Pick by team shape.

The build guide uses Claude Sonnet for drafting and mentions Haiku as the cheaper test. See Claude Sonnet and Claude Haiku.

Add-ons that help

Research servers have their own price lists. Exa lists a free tier, then Hobby at $19, Standard at $99, Growth at $399 and Scale at $749 a month billed monthly. Firecrawl lists a free tier of 1,000 credits a month and Hobby at $19 ($16 billed annually) with 5,000 credits. These are the figures on our add-on pages, checked 31 October 2026. For the full set of servers, browse MCP servers, and see what MCP servers are for the idea.

Prompts and kit items

The prompt library holds sales prompts filterable by task. Two kit items help with building the agent itself: the CLAUDE.md starter template, where you write your sales rules, and the plan-before-coding prompt, for scoping the script before an agent writes it. For repeatable procedures like "prepare a call brief," use a skill.

A safe workflow, step by step

The build guide has the code and the checks. The shape:

1. Decide what the agent may and may not do. May: research, brief, draft. May not: send, change CRM stages, quote prices.
2. Fix the input. The rep supplies the facts: who, what they did, why now. Garbage in, garbage out.
3. Draft from the facts only. Tell the model to mark any claim it cannot support.
4. Add a compliance gate. A check for subject line honesty, a postal address, an opt-out line and suppression-list matches.
5. Hand off to a rep. The draft lands in the rep's drafts folder.
6. Test with real accounts and real reps. Twenty accounts, scored blind.
7. Roll out slowly. One team, then more.
8. Reuse the pattern for call prep and meeting notes.

What it costs

Model cost is small. From Anthropic's pricing page, checked 31 October 2026, Claude Sonnet 5.5 is $2 per million input tokens and $10 per million output tokens, and Haiku 5.5 is $0.10 and $0.50.

Worked example (token counts are our assumptions). A draft with 1,000 input tokens and 600 output tokens costs 1,000 x $2 / 1,000,000 plus 600 x $10 / 1,000,000, which is $0.002 plus $0.006, about $0.008 on Sonnet. On Haiku it is $0.0001 plus $0.0003, about $0.0004.

A research step that reads long web pages changes the picture, because input tokens then dominate. Rep review time is the real cost: at five minutes per draft, 5,000 drafts is more than 400 hours of reading. If the draft is not better than a rep's own, you are paying for nothing. See Claude API pricing and Claude Code pricing if you use a terminal agent to build.

Risks and how to review them

The FTC's CAN-SPAM compliance guide lists what a US marketing email needs: accurate header information, subject lines that do not mislead, a disclosure that the message is an ad, a valid physical postal address, a clear way to opt out and honouring opt-outs within 10 business days. Other countries have their own rules, and some are stricter about cold email. Check each regulator's guidance for the countries you sell into. An AI draft does not change who is legally responsible: your company is.

Measure what the agent changes

Compare to a baseline from the previous month and keep a control group of reps working the old way for a short period.

What a good call brief contains

The brief is the highest-value output, because it is read and used. Ask for a fixed layout so reps learn where to look.

Insist that every claim in "Company" carries a link. A claim with no link is a guess, and reps should skip it. This one rule removes most of the invented-fact risk, because the rep can open the link in seconds.

Rules for the first-touch draft

Write these into the prompt or the agent's instructions file, so every draft follows them.

• Use only facts from the rep's input and the linked sources.
• One reason for writing, one ask, under 120 words.
• No flattery you cannot support, no fake familiarity, no invented mutual connections.
• No pricing, discounts or delivery promises.
• End with a plain opt-out line and the company's postal address, as the FTC guide expects for marketing email in the US.
• If a fact is missing, write a question for the rep in brackets instead of making one up.

Reps will edit the output. Count the edits: the places they always change are the places to improve the prompt.

A month-one plan

Do not automate sending in month one. Do not scale to a second team until the first team has used the briefs for a full cycle.

Common mistakes

• Using AI for volume instead of quality. More bad emails is not more pipeline.
• Letting the model invent personalisation. Personal lines must come from facts.
• Skipping the suppression list. Opt-outs must be honoured, whoever drafted the mail.
• Giving write access too early. Start with read-only CRM access.
• Measuring opens instead of replies. Opens are noisy; replies and meetings are the signal.
• Forgetting the buyer's side. Many buyers can tell, and some dislike it. Keep the voice human and specific.

Privacy and data in sales tools

Sales data is personal data: names, work emails, job titles and call recordings. Before you connect a CRM to a model, decide which fields are needed for the job. A brief needs the role and the history, not the home address. Send as little as possible, check how your model provider handles retained data, and log what was sent. If your buyers are in the EU or UK, check the lawful basis for processing before you load their details into any new tool. The regulator's guidance in your target markets sets the lawful basis for B2B outreach.

Who should do what

For leads before the sale, see AI lead generation. For the service side of the same customer, see AI customer support. For business-wide tool choices, see AI tools for business.

Our take

AI belongs in the preparation half of sales, where it saves hours and the stakes are low. Keep the rep as sender, give the agent only facts you trust and put a compliance check between the draft and the outbox. The model cost is tiny and the review cost is not, so measure the time it saves. We built this page from our build guide, the FTC guidance and vendor pages, and have not run a sales team with it. CRM and email product claims are marked for you to confirm with the vendor.