Sales and Marketing AI Agents: What They Do and What Needs You

Sales and marketing AI agents can research leads, draft outreach, tidy CRM records and compile campaign reports on their own. The owner keeps four decisions: what the business promises, what it charges or discounts, who it contacts and how often, and how much it spends. Anything that commits money, reputation or a customer relationship needs a person's approval.

That line is easier to hold on paper than in a busy week. It survives contact with a real inbox only when it is written down as specific approval points and as guardrails inside the agent's own instructions, then tested on a small batch before anything goes out. The fence matters more than the choice of agent.

Follow me on Instagram@sagnikteaches

What these agents actually do all day

An agent differs from a chatbot in one respect: it takes a goal and chooses its own steps, using tools you connect to it, such as your CRM, inbox, calendar, spreadsheets and ad accounts. A chatbot answers a question. An automation runs the same steps every time. An agent decides what to do next, and may be wrong. The comparison of agents, chatbots and automations helps you pick the lighter tool when one will do.

Connect on LinkedInSagnik Bhattacharya

In sales and marketing, the jobs fall into three families.

Subscribe on YouTube@sagnikteaches

• Finding and qualifying leads. Given a description of your best customers, the agent searches public information and your own records, checks each company against your criteria, and returns a shortlist with reasons. HubSpot's Prospecting Agent is one packaged version; HubSpot's rate sheet charges 100 credits (about $1) to recommend outreach for one lead, and its knowledge base lists the agent on Starter, Professional and Enterprise editions of the Marketing, Sales, Service, Data and Content hubs.
• Drafting and sequencing outreach. The agent writes first emails, follow-ups and replies to common questions, using what it learned about the prospect, and can queue them for sending or leave them as drafts.
• Reporting and housekeeping. It compiles weekly campaign numbers from your ad, email and website tools, spots a drop in click-through or a stalled deal, and updates CRM fields from calls and emails.

General-purpose assistants now offer agent modes too. ChatGPT Work, OpenAI's successor to the retired ChatGPT agent, handles longer multi-step jobs across your apps and files; OpenAI's pricing pages list it on desktop, web and mobile for Plus, Pro, Business and Enterprise, with limited desktop-only access on the Free and Go plans. The point is that the categories overlap: you can buy an agent inside the CRM you already use, or point a general one at your tools. Either way, the same fence applies. If you are weighing where to begin, AI marketing agents for a small business goes deeper on the marketing half.

The owner's line: three tiers of freedom

The most useful way to divide the work is by what the agent is allowed to do with its output. Think of three tiers.

What puts a task in a tier is the cost of a mistake and whether it can be undone. A wrong lead score is a nuisance; a wrong email to a hundred prospects is a reputation cost you cannot recall; a wrong discount is a commitment someone may hold you to. The tiers should be written down and visible to everyone who works with the agent, not just remembered by the person who set it up.

A rule of thumb for edge cases: if the action goes outside your business (an email, a post, a payment, a call), it is at least "draft, you approve" until the drafts have earned trust. If it commits you to a number, it stays "never alone". What can go wrong when agents act for you lists the failure cases behind this rule.

A sports-club shop's week with an outreach agent: the worked example

An illustrative case. A sports club runs a small shop and a corporate-membership scheme: local firms pay for team access and kit. The club wants 15 new corporate members this season and has one part-time person on sales. They use an agent inside their CRM, set to the three tiers above.

What the agent does each Monday:

1. Research (acts alone). It reads a target list of 60 firms the club assembled, checks each against three criteria (between 20 and 200 staff, a site within the club's catchment, no existing membership), and returns the 40 that fit, each with one line on why.
2. Drafts (waits for approval). For the 25 it ranks highest, it drafts a short first email using only facts from the club's fact sheet (membership options, ground facilities, trial day) and public facts about the firm. Drafts land in an approval queue.
3. Report (acts alone). By Friday it lists opens, replies, meetings booked and deals moved, and names the three drafts that got replies.

The part-time salesperson spends 20 minutes on Monday reading 25 drafts and approving 21, editing 3 and deleting 1 that referenced a firm's expansion the agent could not source. Then the agent sends the approved ones through the club's email account at 8 emails an hour so it does not look like a blast. Costs, on illustrative HubSpot-style pricing, where credits are used when the agent writes outreach for a lead: 25 leads at 100 credits each is 2,500 credits, or $25 at $10 per 1,000 credits, before the cost of the seats and plan and less any credits the plan includes. Over a 10-week season that is about $250 in credits, against a time saving of roughly 5 hours a week on research and first drafts.

Results, as an illustration of how to measure rather than a promise: about 240 first emails over 10 weeks, 31 replies, 12 meetings, 5 corporate members. The numbers that matter to the owner are the cost per meeting (credits plus review time) and what happened to the 3 drafts in a week that were edited heavily, because those show where the agent is weakest. That measurement habit is the same one used when piloting any agent; piloting your first agent without risking customers sets out a safe order of steps.

What the club never delegated: the price of a corporate package, any discount for a bigger team, and any reply that mentioned a complaint. When a prospect asked, "can you do it cheaper for 40 staff?", the agent flagged the reply and the owner answered.

Guardrails you can paste into the agent's instructions

Most agents let you set standing instructions, a system prompt or a "guidelines" field. Whatever it is called, the limits go there, and they should be specific. An illustrative set for an outreach agent:

You help [business name] find and contact potential customers.

Facts you may state about us: only those on the attached fact sheet. If a prospect asks about anything not on it, do not answer; flag the message for a person.

You must never:
- state or imply a price, discount, delivery date or stock level
- promise results, guarantees or "exclusive" offers
- contact anyone on the suppression list, or anyone who has replied "stop", "unsubscribe" or similar
- write about a prospect's personal circumstances, health, family or finances
- invent or guess a fact about a prospect: use only what you can cite from a public page or our CRM, and keep the source URL with the draft
- send anything; place every draft in the approval queue

Every draft must include our business name and a clear way to opt out.
Limit: 25 drafts per day.
When unsure, flag it and stop.

Notice the structure. It lists what the agent may say, what it may never do, how it handles uncertainty and a daily cap. "When unsure, flag it and stop" is the line that does the most work, because it turns a risky guess into an item in your inbox. Have a person read the first 50 drafts and keep a list of every fault; put each recurring one into the instructions.

One note on law and good practice. Rules for marketing emails and messages differ by market, and usually require that the sender is identifiable, the message is honest and recipients can opt out. If you contact people in the EU or other regulated markets, check consent and data-protection rules with a qualified adviser before an agent sends anything cold. The agent follows instructions; it does not know the law.

Approval points: where a person must look, and how fast

Approvals fail when they are too slow to use or too many to read. Design them so that a human can actually do them. A workable pattern:

• Batch queue, once or twice a day. Drafts wait in one list, sorted by risk, with the source for each claim visible next to it.
• Spot-check after trust. After a month of good drafts, approve a random sample of one in five and let the rest go, while keeping the "never alone" list fixed.
• Escalate replies. Any prospect reply that mentions price, a complaint, a legal term, a competitor or a bigger deal than usual goes straight to the owner.
• Kill switch. Know where to pause the agent in under a minute, and who is allowed to. A short test run each quarter makes sure it works.
• Spend caps. Where an agent can touch ad budgets or paid tools, set a hard monthly limit in the platform itself, not only in the instructions.

Make the approver's job easy. A draft that shows the email, the facts it relies on and their sources lets a person approve in 20 seconds. A draft with no sources takes minutes and gets approved without reading.

The invented detail that nearly went out

An illustrative failure. A small supplier of garden furniture set up an agent to draft outreach to landscaping firms. One draft read: "Congratulations on opening your new showroom last month." The firm had not opened a showroom; the agent had combined a news item about a similarly named firm with the prospect's details. The draft was caught only because the approver knew the firm.

What went wrong was not the agent's intent. It had been asked to "personalise" and was never told to limit itself to facts it could cite. What fixed it was a rule that every specific claim about a prospect carries a source link in the draft, and that a claim without a source is deleted, together with a short list of allowed personalisation hooks (their trade, their size, a service they list on their own site). The result is less impressive-sounding mail, and none of it false.

Marketing agents: campaigns, content and reports

The same tiers apply on the marketing side, with a few changes. A marketing agent is often given more autonomy because a post or an ad is reversible, but check that. A social post can be deleted; the screenshot cannot.

For reporting in particular, ask the agent to show its working: which accounts it read, which date range, and which figures it combined. Totals built from several tools are the commonest place for a quiet error, such as a date range that differs between two sources. The report should say "Spend $1,240 (ad account, 1 to 7 November)" rather than "Spend was up". A person reads the three numbers that drive a decision, and ignores the rest.

Reading an agent's draft the way an approver should

A concrete before-and-after shows what approval means in practice. Say a gift shop that sells to local businesses for staff presents has an agent drafting first emails. This is an illustrative draft as it first arrived:

Subject: Gifts your team will love, 20% off this week

Hi [first name],

I saw your team recently won an award, congratulations! We'd love to supply your end-of-year gifts, and we can guarantee delivery before 20 December. Order today and we'll give you 20% off.

Best wishes,
[sender]

An approver should catch four problems in under a minute. The award has no source. "Guarantee delivery" is a promise the shop has not made. "20% off" is a price the agent was never given. And there is no opt-out line. The edited version keeps the useful part:

Subject: Staff gifts for the end of the year

Hi [first name],

We supply small gifts for staff and clients to local businesses, in batches from 10 to 200. If you are planning end-of-year gifts, I can send a short list with sample prices and order deadlines.

Would a quick call this week suit?

[sender]
[business name] | Reply "stop" and we will not contact you again.

The second draft makes no claim that needs checking, names no price, promises nothing and offers an opt-out. That is the standard to hold the agent to, and each fix can be added to the guardrails so the next batch arrives closer to it. After three weeks of corrections the rate of drafts needing a heavy edit usually falls, and that falling rate is the evidence for loosening an approval point, not a feeling that the agent is doing fine.

A coworking space and its monthly campaign report

A reporting agent shows the lower-risk end of the scale. A coworking space runs small paid campaigns for desks and meeting rooms, plus a monthly email to past enquirers. The owner asks an agent to produce a one-page report on the first working day of each month from three sources: the ad account, the email platform and the booking sheet.

An illustrative month: ad spend $860, 41 desk-trial bookings from all sources, 9 of them from ads. The agent reports a cost of about $96 per ad-sourced trial ($860 divided by 9). The owner checks the division and finds it right, but notices the booking sheet counted trials by booking date and the ad account by click date, so two late-month clicks fell in the next month. The report is useful and slightly misleading, which is the normal state of agent reporting. The fix is a line in the report format: "state the date basis for every source". This is the sort of check the owner keeps, because the decision that follows (spend more on ads, or on email) depends on the figure being comparable.

Notice what the owner did not hand over: the decision to raise the ad budget. The report said the cost per trial was lower on email. The owner chose what to do about it.

Choosing where to start, and what to leave for later

Not every business needs an agent, and not every job in the business suits one. Start with the job that is repetitive, has a clear definition of "good", and costs little to get wrong. For most small businesses that is lead research or the weekly report, because both are internal.

1. Pick one job and write down what a good result looks like in numbers: 25 qualified leads a week, a report ready by 8 on Friday.
2. Write the three tiers for that job and the guardrails.
3. Run it in shadow mode for two weeks: the agent produces output, nothing is sent, and a person compares it to what they would have done.
4. Switch on "draft, you approve" and measure approval rate, edits per draft and time spent.
5. Review monthly. Loosen a limit only when the numbers justify it.

Before switching on anything that touches your CRM, make sure the data the agent reads is in shape. Duplicate contacts and out-of-date notes make an agent confident and wrong; preparing your business for AI agents covers the groundwork. And check the cost model first: some tools charge per action or per lead, which is fine at 40 leads a week and expensive at 4,000. What an AI agent costs a small business helps you do that sum.

What to tell your team and your customers

Two short messages avoid awkward moments. To staff: what the agent does, what it may never do, who approves, and who to tell if it behaves oddly. A one-page "agent rules" sheet pinned in the shared drive does the job. To customers and prospects: nothing in most cases, but if the agent answers enquiries directly, say so, and give a way to reach a person. Where rules or platform terms require disclosure of automated contact, follow them; if you are not sure, ask a qualified adviser.

A closing check for the owner. If you were asked tomorrow who decides what the business promises, charges and spends, would the answer still be a named human? If it is, the agent is working inside your business. If the answer has become "the agent, mostly", pull the limits back in until it is not.

Questions owners ask before switching an agent on

Is an AI agent different from a chatbot or an automation?

Yes. A chatbot answers, an automation follows a fixed script, and an agent chooses its own steps towards a goal, using tools such as your CRM and email. That freedom is the benefit and the risk, which is why an agent needs limits on what it may send, spend or promise.

Should an agent send emails without my approval?

Not at the start. Let it draft and queue, and approve each batch for the first few weeks. Move to approving only exceptions once the drafts have been right for a month, and keep a person in the loop for new audiences, any price or promise, and replies from angry or high-value contacts.

What does a sales agent cost?

It depends on the product. Some CRMs charge per unit of work: HubSpot's credits are $10 per 1,000, and its Prospecting Agent was listed at about 100 credits, roughly $1, per lead recommended for outreach. General assistants with agent modes sit inside plan prices. Test on a small batch, then multiply by your real volume.

What is the biggest risk with sales agents?

Saying something untrue on your behalf: an invented detail about a prospect, a discount you never offered, a delivery date nobody confirmed. Guardrails that limit claims to a fixed fact sheet, plus a person reading the first drafts, remove most of that risk.

Further reads

• What Is an AI Agent? A Plain-English Guide for Business Owners — Get the plain-English definition behind the word.
• Free vs Paid AI Marketing Tools: What Is Worth Paying For — Learn which marketing tools are worth paying for.
• Can ChatGPT Agent Handle Business Admin Tasks Unsupervised? — See how far a general agent can run unsupervised on admin.
• HubSpot AI Agents and Breeze: What Small Teams Should Switch On — Check which HubSpot agents are worth switching on.
• Is AI Marketing Automation Worth It for a Small Business? — Test whether the marketing side earns its cost.
• AI Lead Generation vs Hiring a Salesperson: A Cost Comparison — Compare agent-led lead generation with a hire.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: HubSpot's product and services catalogue (credit rates) and its Prospecting Agent knowledge-base article, OpenAI's ChatGPT and ChatGPT Business pricing pages (ChatGPT Work availability), checked in November 2026. Agent features, plans and prices change often, so confirm on the vendor's own page before you buy.