AI Sales Agent: What It Actually Does, Where It Fits, and How to Deploy One on Your DMs

An AI sales agent is software that reads an inbound message, decides what to do, and acts: it answers product questions, asks a few qualifying questions, writes what it learned onto the contact record, and hands the conversation to a person when it should not be the one replying. It does not close deals on its own.

That definition is narrower than the one most vendors use, and the narrowness is the point. Almost every disappointing deployment comes from expecting a job an agent cannot do while skipping one it does well. This post is the operating manual: what the agent actually does through a working day, where it fails, how the vendor landscape splits into categories that solve different problems, what the tokens cost at real provider prices, and a step by step deployment on inbound DMs with a human handoff that works.

If you want the product rather than the argument, the AI agents page shows the channels, modes and controls in one screen.

What an AI sales agent is, precisely

Four properties, all four required. Drop one and you have something else wearing the same name.

• It reads. The incoming message is input, not a trigger. A system that matches the word "pricing" and fires a saved paragraph never read anything.
• It decides. Given the message, the thread so far and what it knows about the contact, it picks an action: answer, ask, escalate, or do nothing.
• It acts through tools that write. Not just prose back to the customer. It moves a contact to a stage, adds a tag, writes a note, suggests a score. A system that describes your pipeline but cannot touch it is an articulate read-only reporter.
• It can quit. It has the authority to declare a conversation out of scope and route it to a person with the transcript attached. A system that cannot stop is not autonomous, it is just uninterruptible.

The fastest way to test a vendor against that definition is to ask for the write actions, as verbs. "Sets contact stage, adds tag, creates note, assigns owner, escalates to human" is an answer. A paragraph of adjectives is also an answer, and it tells you the thing is a language model in a text box. We wrote the full taxonomy in AI agent vs chatbot, and the short definition lives in the AI agent glossary entry.

Three things it is not

Not a chatbot with a script. A scripted bot has a finite output space: every string it can produce was written by a person and could be printed on a wall before launch. That is a real property with real value, and for some jobs it is strictly better. It is also a different product. When a scripted bot runs out of branches it says "I didn't get that". When an agent runs out, it either looks something up or hands the conversation over.

Not an email sequencer. A sequencer emits step two on day three whether or not step one landed. Its unit of work is the campaign. An agent's unit of work is the conversation, and it only speaks when spoken to. Teams conflate the two because both get sold as "automation", and then they wire an agent to send the first message to a cold list, which is how you end up with a compliance problem that has a machine learning component.

Not a dialer. A dialer maximizes attempts per hour. An agent is not trying to increase the number of conversations you have, it is trying to make the ones you already have finish properly. If not enough people are messaging you, an agent will not fix that. It will handle your unchanged volume slightly faster.

"AI sales agent" and "AI SDR" are not synonyms

The two terms get used interchangeably. They describe different risk profiles, and the difference is who started it.

An AI SDR generally means outbound: it builds a list, researches each prospect, writes a first-touch message, and keeps going until someone replies. The recipient never asked for it, and everything hard about that product is downstream of that one fact: deliverability, consent, brand damage, and the platform rules that exist precisely because the pattern got abused. An inbound AI sales agent replies to people who messaged you first. The consent question is already settled, the context is already there, and a bad reply annoys one person who was already talking to you. Both get sold under "ai agents for sales". Only one is a reasonable first deployment for a small team.

The jobs an AI sales agent actually does

The whole list, in the order a real conversation hits them. Nothing here is impressive alone. Together they are most of a first reply.

Read the right-hand column again. Two of the six jobs are pure extraction, which is what this technology is genuinely excellent at. Turning "we're like 40 people ish, looking to sort this before our renewal in November" into a headcount band, a deadline and a note is work a human rep will not do at 11pm. The value is not a charming sentence, it is that the conversation arrives already structured.

Six things it should not do

Each has burned somebody, and each has a one-line fix.

• Negotiate a discount. Discounts are a commercial decision with a floor that lives in somebody's head. An agent has no floor. Put "never offer a discount, price reduction or free extension" in the persona, and a handoff rule on the word "discount".
• Promise a date. "We can have you live by the 30th" is a commitment in writing to someone deciding whether to pay you. Ban delivery dates, SLAs, uptime figures and roadmap timing outright.
• Invent policy. Refunds, contract terms, data residency, cancellation windows. If it is not written down in the knowledge base, the correct answer is "let me get you the exact terms" plus an escalation.
• Confirm an integration. The most frequent expensive fabrication in B2B sales. A prospect names a tool, the model does not want to disappoint, and you have promised an integration you do not have.
• Agree with a confident prospect. Models resolve a conflict in favor of the person in front of them. If someone asserts a feature confidently, an under-scoped agent confirms it, fluently, with no tell. The mitigation is architectural, not a better model.
• Handle an angry person. Not because it cannot write a calming sentence, but because a human appearing at that exact moment is worth more than anything the agent could say.

An honest operating manual: the agent's day

Vendor pages describe capabilities. Here is what actually happens between a message arriving and a reply going out, in order, because knowing the sequence is what lets you debug it.

A message lands

The first thing that should happen is nothing. A well built agent waits, because people send three messages in a row: "hi", then "quick question", then the actual question. Replying to "hi" burns your first reply on nothing and tells the person they are talking to a machine. A debounce window of a few seconds, merging whatever arrives inside it, changes the whole texture of the conversation.

The deterministic gates run

Before any model is called, a stack of cheap checks decides whether this message reaches the model at all. Is this contact paused because a human took over? Does the trigger scope match? Is it quiet hours? Has the agent hit its cap for the hour? Does the message contain a handoff keyword? Does a rule already have the answer? Every one of those is a decision tree sitting inside an AI agent, and every one of them is free. The cheapest wrong answer is the one you never generated.

The model runs

It gets the persona, the knowledge base entries, the conversation history and the incoming message. It produces either a reply or a signal that it should not be replying. That is the entire scope of the model's authority in a well designed system, and it is much smaller than the marketing implies.

The write actions fire

Stage change, tag, note, score suggestion. This is the half nobody demos and the half that survives. Even if you decide the replies are not good enough for auto-send, the structured record the agent leaves behind is worth keeping on its own.

Then a human reads it

In month one, all of it. In month three, a sample. Never zero. The team that stops reading discovers a four-week-old fabrication when a customer quotes it back during onboarding.

Where it fails, specifically

Four failure modes, ranked by how often they actually bite.

1. The knowledge base gap. Somebody asks a question you never wrote an answer to. The agent does not say "I don't know" unless you deliberately made that the easier option. It produces something plausible, in your brand voice, with no hedge. This is the number one failure, and it is a writing problem rather than a model problem.
2. The confident prospect. Reproducible in ten minutes on your own setup: assert a feature you do not have, confidently, and see whether the agent agrees with you.
3. The stale fact. You changed a policy in March and the knowledge base still holds the January version. The agent is perfectly accurate against a document that is wrong. Put a review date on knowledge base entries the way you would on a price list.
4. The handoff that goes nowhere. The agent escalates correctly, the notification lands in a channel nobody watches, and the customer waits for a person who never arrives. This is the worst outcome in the system, worse than having no agent, because you manufactured an expectation and then broke it.

Where it runs, and why the channel decides everything

The same agent behaves completely differently depending on where you point it, because channels differ on consent, on how long you may reply, and on what a mistake costs.

The window rule that makes DMs urgent

On the WhatsApp Business Platform, Meta's documentation is explicit: "When a WhatsApp user messages you or calls you, a 24-hour timer called a customer service window starts", the timer "resets to 24 hours" if they message again, and "When the window closes, you can only send pre-approved template messages" (Meta developer documentation, checked September 2026). Messenger works the same way: "Businesses have up to 24 hours to respond to a user" (Messenger Platform policy overview, checked September 2026).

Read that as a sales constraint, not a technical one. Miss the window and you do not get a worse reply, you lose the right to send a free-form reply at all. That is a different argument from the usual speed-to-lead one, which we took apart in lead response time and the five minute rule, and it is the strongest reason to put an agent on inbound DMs first. The window itself is defined in the glossary.

Why DMs are the practical starting point

Three reasons, none about the technology. Consent is settled, because the person messaged you. Messages are short, which keeps both the token cost and the blast radius small. And volume is usually low enough that you can read every reply for the first two weeks, which is the only quality control that actually works. Email fails all three.

The landscape, by category rather than by ranking

Every "top 15 AI sales agents" list ranks tools that do not compete with each other. Here is the split that matters, by the job the vendor built for. Descriptions are quoted from the vendors' own pages, checked September 2026, and no ordering is implied.

1. Outbound AI SDR platforms

These generate pipeline from cold lists. 11x sells "digital workers": Alice, an "Outbound digital worker" that "engages prospects across channels, handles replies, drives qualified meetings and builds pipeline", plus Julian, an inbound digital worker that covers speed to lead, inbound qualification and an AI phone agent. Artisan sells Ava, which "finds leads, enriches them, sends personalized messages, and books meetings on behalf of your reps". Amplemarket sells Duo, "the first AI sales agent that helps sales teams find and connect with their next customers", alongside a lead database, intent signals, multichannel sequences and a dialer.

What you are buying: a prospecting engine. The agent is the interface, the data and the sending infrastructure are the product. Right purchase if your bottleneck is that not enough people know you exist. Wrong purchase if you wanted help with the messages already in your inbox, and every question in the 2026 outreach compliance playbook becomes yours the day you switch it on.

2. CRM-native agents

Salesforce Agentforce describes a "Sales Development Representative (SDR)" agent that "engages with prospects 24/7, answering questions, managing objections, and scheduling meetings based on CRM and external data". HubSpot describes a "Prospecting Agent" that will "monitor buying signals and launch personalized outreach automatically", and a "Customer Agent" to "resolve inquiries automatically across channels".

What you are buying: agents that already sit on your record of truth. If ten years of opportunity history lives in that CRM, an agent with native access to it starts with an advantage nothing bolted on can match. The trade is that you buy deeper into one platform, and the configuration surface assumes a team with an admin in it.

3. Support-first agents

Fin, from Intercom, calls itself "the highest performing customer agent on the market", says it "works natively across all channels, including Voice, Chat, email, Slack, social, and many more", and now also says it "Handles inbound sales end-to-end". Zendesk describes AI agents that "handle complex, multi-step workflows across channels" and "resolve even the most complex issues on any channel autonomously".

What you are buying: a resolution engine with a decade of deflection engineering behind it. These are the most mature products in the category and genuinely good at answering questions. The mismatch for sales is semantic rather than technical: support optimizes for closing the conversation, sales for opening a relationship, and a headline metric like resolution rate rewards the first. If your inbound is 80% questions and 20% buying intent, this fits well. If it is the other way around, watch what the dashboard teaches you to do.

4. Inbox and DM agents attached to a CRM

This is where Pinlyx sits, with a set of tools built around social and messaging inboxes. The premise: the agent and the human work in the same inbox against the same contact record, so a handoff is a change of who is typing rather than a transfer between systems. The strength is handoff quality, a direct function of how much context crosses the boundary, plus a setup cost low enough for a team with no admin. The weakness is equally plain: this is not the record of truth for a 200-person sales organization, and the agent itself answers rather than prospects. Outbound lives in a separate sequence engine on paid plans, and the agent does not drive it.

5. General agent builders

A fifth category keeps getting counted in the sales rankings incorrectly. Lindy sells an AI teammate that "connects to all your tools, knows everything about your company, and does real work for the whole team", with sales as one departmental use case rather than the product. Horizontal builders will out-do any vertical tool given enough investment, and they hand you the integration work, the guardrails and the evaluation. A fine trade with an engineer, a bad one if you hoped to be live by Thursday.

Three questions that pick your category

Who starts the conversation? "We do" puts you in category one, "they do" puts you in three or four. Where does your customer history live? In a large CRM the whole company runs on, look hard at category two first; in the inbox, three and four pay off fastest. And who configures and reviews it? If nobody has three hours a week for the first month, do not buy anything on this list yet. That is the input every category depends on.

Anatomy of an agent: the ten things you configure

Whatever you buy, the configuration surface is roughly these ten items. Knowing what each is for, and what breaks when you skip it, is most of the skill.

Two of those ten hold all the work. The knowledge base is a writing task, not a configuration task, and it takes a full day done honestly. The handoff list is a judgement task: ask whoever answers your inbox today which questions they would not want a machine answering. That list is your escalation config, and it beats anything you invent alone at a keyboard.

The setting most people get wrong first

Trigger scope. The instinct is "all messages", because that is what automation is supposed to mean. The better first setting is "first contact only": the agent handles the opening message of each new conversation and then steps back. That covers the case with the most value, a fast first reply inside the platform window, and caps your exposure at one message per new person. Widen it in week three, when you have read two hundred replies.

Qualification logic that survives a real chat

Qualification in a DM is not a form with a chat skin on it. People leave. The constraint that shapes everything else: you get four to six questions across a whole conversation, and each has to feel earned.

The rules that make it work

• Answer before you ask. The person opened with a question. Answer it, then ask one back.
• One question per message. Two questions in one bubble reliably gets one answer, and it will be the easier one.
• Only ask what changes your behavior. If the answer would not route the conversation differently, skip it. "What industry are you in" is usually decoration.
• Budget is not a chat question. Asking a stranger about budget in message three reads as a qualification screen, because it is one. Ask about the use case and the timing instead: both correlate better, and neither is offensive.
• Stop at four to six. Then either hand off or send the next step. A seventh question is an interrogation.

The order that works

1. What are you trying to do? The most informative question in sales, and the only one people enjoy answering.
2. What are you using now? Tells you the switching cost, the competitor, and often the budget without asking about budget.
3. How many people, or how much volume? Sizing, phrased as a fact rather than a test.
4. When do you need this working? The answer best correlated with whether anything happens at all.
5. Who else is involved? Only if the first four suggest a real opportunity. A late-conversation question.
6. How should we reach you? Email or a good time to call. The close, not the opener.

Everything above is extraction, which is what the technology is good at. What the agent should not do is convert those answers into a verdict and act on it. Ranking is fine, rejection is not, and the reason is a whole separate argument we made in AI lead qualification: a system that disqualifies destroys the data you would need to find out whether it disqualifies correctly. If you want the fit half of this written down first, that starts with a real ICP definition.

Example scenario: a qualification that stops at the right moment

An illustrative dialogue, not a customer transcript.

Prospect: hey, do you handle Instagram DMs too or just WhatsApp?

Agent: Both, plus Telegram, Facebook and live chat, all in one inbox. Out of interest, what is the setup you are working with right now?

Prospect: two of us answering Instagram on our phones, it's chaos, we lose stuff constantly

Agent: That is the usual failure mode: two phones and no shared record. Everything lands in one thread per contact, so you can both see it and see who replied. Roughly how many messages a day are you getting?

Prospect: maybe 60-80 on a good day. we're launching a second brand in November so it'll go up

Agent: Noted. November is close enough that it is worth setting this up now rather than during the launch. Is the second brand on its own Instagram account, or the same one?

Prospect: own account. and we'd want different people answering each. can you do that, and can we get a discount if we sign for a year?

Agent: Separate accounts with different people on each is exactly how the inbox is meant to be used, so yes. On annual terms I am not the right person to answer that, so I am bringing in someone from the team now, with this conversation attached so you do not have to repeat any of it.

Four questions, one objection answered, one clean stop. The agent handed over on the word "discount", not because the question was hard but because the answer is commercial. The human inherits a contact with a channel, a volume, a deadline, a second brand, a requirement and a live question, rather than a notification that says "customer needs help".

How to deploy an AI sales agent in Pinlyx

The concrete version, with screen names, in the order you will click them. AI agents are on the paid plans: Pro includes three agents, Business unlimited agents plus learning from what your operators send. The free plan is Telegram and the CRM core and does not include agents. Plan details are on the pricing page.

1. Add your own model key. Go to Settings, AI Assistant, pick a provider and paste an API key: OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Groq, OpenRouter, Cohere, Perplexity, Together, Moonshot, Qwen and Meta Llama are all on the list. There is no platform-supplied key. Model usage is billed by your provider, on your account, at their rates.
2. Create the agent. In the sidebar, open AI Agents and click New AI agent. On the Basics step, choose Learn from your chats, which reads the real conversations on your connected accounts and drafts a persona and starting knowledge from them, or start from a template: Sales SDR, Customer Support or Lead Qualifier.
3. Set persona and channels. Write the identity, the goal and the never-do list. Then pick channels: Telegram, live chat, and the social inbox, which covers WhatsApp, Instagram, Facebook, Bluesky and Reddit. Two limits to plan around: on X the agent only produces suggestions and never sends, and email is not an agent channel, it runs through the inbox rule engine's auto-reply action.
4. Choose triggers and mode. Trigger: Every incoming message, Only when keyword matches, First contact only, or Use rules. Reply mode: Auto-send, Suggest, or Draft only. Start on First contact only plus Suggest, whatever the demo made you want to do.
5. Write the knowledge base. The Knowledge tab takes titled plain-text entries. There is no URL crawling and no PDF or DOCX import, so those twenty answers have to be written by a person. That is the real work of the setup, and the difference between a useful agent and a fluent liar.
6. Add rules. The Rules tab maps a condition (keyword, regex, contact stage, platform) to an action: reply, set stage, add tag, enroll in a sequence, hand off to a human, or block. Put the handoff rules above the reply rules, so an escalation is never overridden by a canned answer.
7. Configure the handoff. Escalation fires on model uncertainty, on your handoff keywords, when the customer asks for a person, and optionally on negative sentiment. Two behaviors come for free: when an operator types a manual reply the agent goes quiet for that contact, and you can pause a contact permanently.
8. Set limits and quiet hours. Defaults are 20 replies an hour, 200 a day, and a 5 second delay before sending, with quiet hours in UTC. The debounce is 8 seconds (4 on live chat) and merges up to 12 messages, so the three-messages-in-a-row problem is handled before you touch anything.
9. Test it. The Test playground tab runs conversations against the live configuration with no customer on the other end. Replay twenty real threads from your own inbox, including the five where the customer was wrong about something. Those five tell you whether it agrees with confident people.
10. Watch it. The Activity and Alerts tabs are where the first month lives. Contacts the agent handles land on your board at Pipeline, the score it suggests is visible and editable in lead scoring, and every escalation arrives in the unified inbox with the full thread attached. For live chat, the widget is a single script line and carries its own AI disclosure text, handoff message and business hours.

The deeper settings walkthrough is the deploy AI agents guide, and the persona writing is in the agent training guide. To run this against your own messages, create a free account, connect a channel, and keep the agent in suggest mode for a week before anything is sent on its own.

Guardrails, consent and disclosure

None of this is legal advice and jurisdictions differ. Four things every deployment needs an answer for.

Consent and who started it

Inbound is the easy case: the person messaged you, and replying is what they asked for. Outbound is where the rules bite, identically whether a human or a model wrote the message. If you are considering an agent that sends first, read the compliance playbook before any vendor page.

Platform windows

Sourced above: 24 hours on the WhatsApp Business Platform, resetting on each inbound message, with pre-approved templates the only option once it closes, and 24 hours on Messenger. Two consequences for an agent. Part of its job inside the window is keeping the window open, and one relevant reply does that. And do not design a follow-up sequence that assumes you can message someone tomorrow, because on these platforms you frequently cannot.

Disclosure

Two rules worth knowing by name. In the EU, the AI Act's Article 50 requires that "designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious", and it applies from 2 August 2026 (Article 50 text, checked September 2026). In California, Business and Professions Code section 17941 makes it unlawful to use a bot to mislead someone about its artificial identity "in order to incentivize a purchase or sale of goods or services in a commercial transaction", with a safe harbor: "A person using a bot shall not be liable under this section if the person discloses that it is a bot", and that disclosure must be "clear, conspicuous, and reasonably designed to inform" the people it talks to (California legislative text, checked September 2026).

The practical version is one sentence at the top of the conversation. Teams worry that labeling the agent will destroy conversion. The drop is smaller than the fear, and the cost of being caught is not a number you get to choose.

Escalation SLA

Write down how long a handed-off conversation may wait, and who owns it in each hour of the week. If you cannot answer that, you are not ready for auto-send, and the correct configuration is suggest mode, where nobody is ever promised a human who does not arrive. For planning, Salesforce's own Agentic Enterprise Index reported escalations to humans rising from 22% in the first quarter of 2025 to 32% in the second, and framed the rise as agents getting better at recognizing their limits (checked September 2026). Staff for a third of conversations reaching a person, not a tenth.

What to log

Per message: which agent handled it, which gate stopped it if any, what the model was given, what it produced, the latency, the cost, and the outcome. If you cannot reconstruct why a sentence was sent six weeks later, you cannot debug it, audit it, or defend it to whoever asks.

What it costs

Three separate lines, and the one people budget for is the smallest.

Line one: model tokens

Provider pricing is public and you pay it directly when you bring your own key. As a sizing rule, the OpenAI documentation states that "1 token is approximately 4 characters or 0.75 words for English text" (OpenAI API concepts, checked September 2026).

Example scenario, not a customer result. Take a conversation of six agent replies. Each reply is one model call carrying the persona, the knowledge base and the history so far, averaging roughly 3,000 input and 150 output tokens per call as the thread grows. That is about 18,000 input and 900 output tokens per conversation, so a thousand conversations is roughly 18M input and 0.9M output. Against published list prices:

Sources, all checked September 2026. The OpenAI API pricing page lists gpt-5 at $1.25 input and $10.00 output per million tokens, and gpt-5-mini at $0.25 and $2.00. The Anthropic pricing page lists Haiku 4.5 at $1 and $5 per million, Sonnet 5 at $2 and $10. The Gemini API pricing page lists Gemini 2.5 Flash-Lite at $0.10 and $0.40 on the paid tier, with a free tier under rate limits. Cheaper options sit further down every one of those lists, and prices change often enough that you should open the three pages rather than trust this table six months from now.

Three things move that number more than the model choice does. Prompt caching cuts the repeated part of the input substantially, and every major provider now offers it. A knowledge base you keep adding to multiplies across every call, so a 40-entry base is not twice the cost once, it is twice the cost on every message forever. And long threads get expensive faster than you expect, because each turn resends the history.

Line two: the platform

Whatever tool runs the agent has its own subscription, usually the larger of the two software costs at low volume and the smaller at high volume. Ours is on the pricing page, with no numbers quoted here, because a blog post is the worst place to learn a price.

Line three: human review, which is the real cost

In month one you read everything. Fifty conversations a week at two minutes each is about 100 minutes, call it 1.7 hours. The US Bureau of Labor Statistics publishes no "sales development representative" occupation, so the nearest series is sales representatives for wholesale and manufacturing, at a May 2025 median of $76,460 a year and $36.76 an hour (BLS Occupational Outlook Handbook, checked September 2026). At that rate, 1.7 hours of review a week is roughly $62, an order of magnitude more than the tokens above.

That is the honest shape of the economics, and not the shape the category advertises. The agent does not remove the salary line, it shifts a slice of it from writing replies to reading them. The saving is real, because reading a good reply takes ten seconds and writing one takes two minutes, but it arrives as reclaimed attention rather than reduced headcount.

Metrics that matter, and a weekly ritual

Five numbers. Each can go down and make you feel bad, which is the property that makes a metric worth having.

One metric is deliberately absent: containment, or resolution rate. It counts customers who gave up as successes, for reasons we documented using a vendor's own published definition in the chatbot comparison. On an inbound sales channel it is worse than useless, because silence after a bad reply and silence after a good one look identical.

The Friday half hour

Do this weekly for the first quarter. Thirty minutes, and it is the entire quality program.

1. Read twenty conversations from the activity log, at random and specifically not from the escalation queue. That queue is where the agent already knew it was struggling. The interesting failures are in the conversations it thought went fine.
2. Check handoff recall. Of those twenty, how many should have gone to a person and did not? That is the number that matters, and almost nobody measures it.
3. Look at exactly one metric from the table, the one connected to whatever you changed last week.
4. Make one change. One knowledge base entry, one rule, or one line of the persona. Not three, or you will not know which worked.

One change a week compounds faster than a quarterly overhaul, and it keeps the agent attached to what your inbox actually receives rather than what you imagined in January.

When not to deploy an AI sales agent

We sell one of these. Here is when the answer is no.

When the deal is complex and enterprise. Multi-threaded, six months long, custom scope, a security review, procurement. The agent cannot carry that, and the way it fails is expensive: a confident, slightly wrong answer about something consequential, after which your rep opens the relationship by correcting their own company.

When the advice is regulated. Financial, medical, legal or tax. If a compliance team approved a sentence, that exact sentence gets sent, not a paraphrase of it. Use fixed replies and a rules engine. That is not a downgrade, it is the correct architecture for the job.

When you have no knowledge base yet. An agent pointed at nothing does not fail loudly, it fills the gaps plausibly in your brand voice. If you cannot list twenty questions with written answers, that is this week's project and the agent is next month's.

When nobody can take a handoff. The absolute disqualifier. An agent whose escalations land in a queue nobody watches is worse than no agent, because it manufactures customers who have been told a human is coming and are now waiting. If you cannot staff it, run in suggest mode and take the speed benefit without the abandonment.

When your volume is genuinely low. Under roughly fifty inbound messages a week you have a notification problem, not an automation problem. Fix the notification.

When another tool fits better. If your bottleneck is cold pipeline generation at scale, an outbound platform from category one is the right purchase and we are not it. If your whole company runs on a large CRM and the value is an agent reading ten years of opportunity history, look hard at category two first. If your inbound is overwhelmingly support tickets with exact correct answers, a support-first agent or a plain decision tree beats a sales agent on both cost and accuracy. And two limits on our own side: WhatsApp here runs through the WhatsApp Business Cloud API and does no template broadcasts or bulk sends of any kind, and email is not an agent channel at all, only an inbox rule action.

Where to start

Pick the channel where most of your inbound already arrives. Write twenty answers. Set the agent to first contact only, in suggest mode. Read every suggestion for a week. Then decide, with evidence rather than a demo, whether to let it send.

To try it: the free plan is Telegram plus the CRM core and needs no credit card, so you can get your contacts and conversations in place first. AI agents are on the paid plans and run on your own provider key, which means the model usage is billed to you by OpenAI, Anthropic, Google or whoever you pick, rather than by us. Create a free account, connect a channel, and start in Suggest mode. No credit card required.

Frequently asked questions

What is an AI sales agent?

An AI sales agent is software that reads an inbound sales message, decides what to do, and acts: answering product questions from a knowledge base you wrote, asking a few qualifying questions, writing what it learns onto the contact record, and handing the conversation to a person when it should not be the one replying. The four properties that separate it from a chatbot: it reads, decides, writes to your systems, and has the authority to stop.

How is an AI sales agent different from a chatbot?

A chatbot's output space is finite: every string it can produce was written in advance, and when it runs out of script it loops back to a menu. An agent generates a reply, can look something up mid-conversation, can change a record, and can escalate to a human. That is an architectural difference, not a model-quality one. A frontier model in a text box with no tools is still a chatbot. The full taxonomy is in AI agent vs chatbot.

Can an AI sales agent close deals?

Not the ones worth closing carefully. It can carry a simple, already-decided purchase over the line. It cannot negotiate, build trust with a skeptical buyer, or handle a decision involving more than one person, and it should be explicitly banned from discounts, delivery dates and any commitment your business would have to honor. The realistic job is to keep the conversation alive inside the platform window, answer the answerable, capture the details, and hand a structured thread to a person.

What does an AI sales agent cost?

Three lines. Model tokens, which you pay directly to your provider when the tool uses your own key: on published list prices, a thousand six-reply conversations works out at roughly $6 on gpt-5-mini or roughly $31 on gpt-5 (OpenAI pricing, checked September 2026). The platform subscription, which is on the pricing page. And human review time, the largest of the three in month one and the one nobody budgets for.

Which channels can an AI sales agent work on?

In Pinlyx, agents run on Telegram, the live chat widget, and the social inbox, which covers WhatsApp, Instagram, Facebook, Bluesky and Reddit. On X the agent produces suggestions rather than sending. Email is not an agent channel here; automatic email replies come from the inbox rule engine instead. Across the wider category, voice agents are a separate product with different latency requirements, and outbound email agents carry a much larger compliance burden.

Do I need my own OpenAI key?

For Pinlyx, yes: you bring a key from OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Groq, OpenRouter, Cohere, Perplexity, Together, Moonshot, Qwen or Meta Llama, and that provider bills the model usage to your account. There is no platform-supplied key. The trade is deliberate: one more setup step, in exchange for seeing exactly what the model costs and paying no markup on tokens. Other vendors bundle the model and charge per resolution instead. Both are defensible, and you should know which you are buying.

How do I stop the agent from making things up?

Four things, in order of how much they help. Write a real knowledge base, because plausible invention is what fills a gap. Tell the agent in the persona that admitting ignorance beats an unsupported claim, and give it an escalation path easier than guessing. Use rules for anything with an exact answer, so the model is never asked what a lookup could have handled. And read the activity log, sampling conversations the agent thought went fine rather than the escalation queue. You are not choosing between fabrication and no fabrication, you are choosing a rate and a blast radius.

Is it legal to use AI agents for sales messages?

Replying to inbound messages is generally uncontroversial. The rules tighten around outbound, and around pretending to be human. Two to know by name: the EU AI Act's Article 50 requires people to be informed they are interacting with an AI system and applies from 2 August 2026 (Article 50, checked September 2026), and California's Business and Professions Code section 17941 makes it unlawful to use an undisclosed bot to mislead someone "in order to incentivize a purchase or sale of goods or services", with disclosure as the safe harbor (statutory text, checked September 2026). Platform rules apply on top: the 24-hour customer service window on WhatsApp and Messenger limits when you may send a free-form message at all. This is not legal advice, and your jurisdiction may add more.

Related reading

• AI agent vs chatbot: the four categories vendors blur into one word, and why containment rate hides your failures.
• AI lead qualification: why an automated qualifier should rank rather than reject, and how the feedback loop eats your best segment.
• Lead response time: where the five minute rule came from, and what actually decays on a DM channel.
• How to build a WhatsApp chatbot: the no-code setup, the AI layer, and the rules that get bots blocked.
• Instagram auto reply: automatic DM and comment responses, and the point where you need more than an autoresponder.