Most AI sales agents are marketing fluff. This is the honest guide to what works, what doesn't, and the exact order to roll out AI agents in your sales stack.
We provide neutral advice that works for you.
Away from point sellers, pay to play sites and all the noise that only adds to confusion.
A safe space to ask the questions you want to.
Gartner reckons AI agents will outnumber human sellers ten-to-one by 2028. Fewer than 40% of sellers using them today say productivity has actually improved. Both numbers are worth holding in your head at once, because the gap between them is where most sales tech budgets are being lit on fire right now.
This post is the honest version. What an AI sales agent is, what it actually does, where it pays off, and where the marketing has run ahead of the product.
AI sales agents are autonomous software systems that pursue a goal, take multi-step actions across your tools, and make decisions along the way, with or without a human in the loop.
Strip the marketing language and an AI sales agent is three things stacked: a model that reasons, a set of tools it can use (data, CRM, calendar, email, dialler), and a goal you have given it. Tell it "book me a meeting with the CRO at any UK SaaS company that just raised a Series B" and a real AI agent will go find the company, find the person, enrich the contact, draft the outreach, send it, log the activity, and update you on progress. It will also make a hundred small decisions on the way: which list to use, when to follow up, whether the lead is qualified.
That is the textbook version. The market version is messier, and that is what most sales leaders are buying without realising it.
A real agent has four ingredients:
Anything that lacks one of those four is not an agent. It is an automation, an assistant, or a chatbot wearing a costume.
A chatbot answers questions. An agent takes actions. A chatbot is reactive (the user asks). An agent is proactive (the goal triggers). If your "AI sales agent" only replies to inbound messages, you bought a chatbot.
An assistant suggests, an agent acts. A sales assistant tells your rep "this email looks like it needs follow-up." A sales agent sends the follow-up. The distinction matters because the risk model is completely different. Assistants are safe by default. Agents need governance from day one.
An AI sales agent runs a four-stage loop: gather context, reason, act, check with a human if the action is consequential.
The mechanics are not the interesting part for a sales leader. What is interesting is the loop, because it tells you where the agent can break.
The agent reads from your CRM, your data provider, your email, your calendar, sometimes your call recordings. The quality of the output is bounded by the quality of the input. An agent on top of a messy CRM produces messy outputs faster.
This is the model layer (Claude, GPT, Gemini, take your pick). It interprets the goal, plans the steps, evaluates the data, picks the next action. The reasoning is opaque, which is why you cannot skip the next step.
The agent uses the standard plumbing (often MCP, the protocol that lets it reach your stack) to actually do things: create the deal, draft the email, book the meeting, update the contact. Each action is a permission you have given it.
The good agents put a human in the loop on consequential actions. Reading the CRM is read-only. Writing a deal record is write. Sending an email to a prospect is broadcast. The risk goes up at each step, and so should the friction. If your agent sends prospect emails without a rep clicking approve, you have a generic-spam machine on your hands.
Most "AI agent" stories collapse seven different jobs into one buzzword. Separate them out and the buying decision gets easier.
The agent that builds and qualifies lists. Inputs: an ICP, signal sources, your CRM. Output: a list of accounts and contacts ready for a human review. This is where the value is highest and the risk is lowest, because the human still picks who to actually call.
The agent that drafts and sends outreach. This is where the "AI SDR" pitch lives. Reality check: every AI-drafted message we have seen still reads as AI. The "too cleanly written" tell is the kiss of death. Use these to draft and structure, not to send.
Inbound webform fills, MQL passes, calendar bookings. An agent looks at the lead, scores it against your ICP, routes it. This is a legitimate win. The qualification logic is rules-based enough that an agent can be trusted.
Chat on your website, on WhatsApp, in your product. Decent for FAQ and qualification. Less decent for actual selling. Use them to filter, not to close.
The most-hyped, the least-mature. Voice agents that can run an outbound call exist, and they sound impressive in a demo. They do not yet sound impressive in the wild on the third call. Watch this space. Do not bet your pipeline on it in 2026.
Call review, role-play, in-the-moment coaching. This is the quiet winner. Coaching is the function where AI agents actually move the metric, because they scale a thing (good coaching) that was previously bottlenecked on a human (a good manager). Real-time coaching beats Friday call review, and an agent makes real-time coaching possible.
Pull the data, surface the signal, draft the commentary. Useful for RevOps. Mostly invisible to reps. Often the best ROI in the agent category because it removes manual reporting load from the leadership team.
The honest list. We have used or stress-tested every one of these.
Before you buy any of the platforms above, build a monitoring and signal layer first. Point it at your TAM and the accounts already in your CRM, and have it watch for the things that actually predict a buying window: job changes into the roles you sell to, funding rounds, specific hires with keywords in their profiles (a new VP Sales, a Head of RevOps), expansion into a new geo, headcount jumps. Let the agent source the contact and the reason to call. Then hand the list to a rep to actually reach out. That is the workflow that pays back fastest in 2026, and it lets you learn how far you can trust an agent before you give it write access to a prospect's inbox.
Once the signal layer is running, then the platform question:
If you are HubSpot-first, start with Breeze. It is in your tenant already.
If you are Salesforce-first, start with Agentforce, but only if your data model is genuinely clean. An agent on dirty data is worse than no agent.
If you are neither, do not buy an "AI SDR" platform first. Buy a coaching agent first (SecondBody, Hyperbound, Second Nature, take your pick). Coaching is the function where AI agents move the metric reliably. The "AI SDR" category will catch up. It is not there yet.
The principle we keep coming back to: the AI-SDR category is the loudest, the AI-coaching category is the most valuable. Do not let the volume confuse the value.
Only 13% of organisations feel prepared to manage AI agents at scale. The other 87% are buying anyway. Here is how to be in the 13%.
Pick something that happens 100+ times a week. Lead routing, list-building, call summarisation, follow-up drafting. The agent's ROI is volume-bound. Low-volume tasks do not pay back the setup cost.
The goal is the agent's spine. Vague goals produce vague agents. Replace "improve outbound" with "double the number of qualified meetings booked from inbound MQLs that arrived in the last 24 hours." Specific, measurable, time-bound. Same rules as for a human rep.
Agents are bounded by the data they can see. Connect the CRM. Connect the calendar. Connect the call recordings. Connect the enrichment provider. Do not connect everything you have access to. Connect the minimum needed for the goal.
Read-only by default. Write access by exception. Send-to-prospect access by very rare exception, with a human approve step. Build the loop, then loosen it slowly. The teams that get burned by AI agents are the teams that loosened first and checked second.
Track three things: actions completed, actions approved by the human, downstream outcome (meeting booked, deal moved, ticket resolved). If approval rate is below 70%, your agent is too aggressive. If actions-to-outcome is below your manual baseline, the agent is not earning its keep yet.
AI agents for sales are real, useful, and overhyped at the same time. The hype says they replace reps. The reality says they replace the worst hour of a rep's day, the one spent on admin, research, list-building, and CRM hygiene. That hour is worth real money to a sales team. Pretending it is the whole day is what gets companies into trouble.
Start with one workflow. Pick a coaching or research agent first. Earn trust before you give the agent your outbound. Keep a human on every send. The teams who treat AI agents this way win. The teams who buy the "AI SDR" pitch on day one tend to come back to us a quarter later, asking why the meetings dried up.
The unsexy answer is sometimes the right one: dial more, coach better, then layer AI on top.
Autonomous software that pursues a sales goal across multiple tools, makes decisions along the way, and reports back. Distinct from a chatbot (which only answers questions) or an assistant (which only suggests).
Trusted to do what? Trusted to read CRM data and summarise a call, yes. Trusted to send an unreviewed message to a real prospect, no. Trust is per-action, not per-tool.
They run a loop: read context, reason, take action, check with a human if the action is high-risk. The work happens across your existing stack via standard APIs and MCPs.
No, not in 2026. They can take 30 to 50% of the admin load off an SDR's day, which is a real lift. They cannot replace the conversation, the rapport, the cold-call resilience, or the follow-up discipline.
Either with a workflow platform (Clay for RevOps, Sendr for outbound-with-a-human-on-send) or natively inside your CRM (Salesforce Agentforce, HubSpot Breeze). Start with the goal, the data sources, and the approval loop, then build.
Through native integrations (Agentforce in Salesforce, Breeze in HubSpot) or through MCP servers (a growing list). The native route is faster to set up. The MCP route is more flexible across multiple tools.
High-volume, repeatable sales motions benefit most: SaaS, fintech, B2B services. Anywhere with a long, low-touch top of funnel. Complex enterprise sales benefit less, because the value is in the rep relationship, not the volume.
A chatbot answers what you ask. An agent goes and does what you said. Same underlying model, completely different job.
What Is an MCP? The Sales Leader's Guide
AI in B2B Sales
James Donaldson
Founder, Stakki
📧 james@stakki.io
