AI AGENTS FOR SALES: WHAT THEY ACTUALLY DO TODAY
OCTOBER 26, 2026
We buy capability. We almost never buy the conditions that make capability work.
It is the oldest pattern in enterprise software, and AI agents are about to repeat it at scale. Every evaluation asks which agent is best. Almost none asks what the agent will be allowed to know. The first question produces a shortlist. The second decides whether anything on that shortlist works.
The evidence for that comes from an unlikely place. Salesforce sells AI sales agents, and Salesforce's own research arm published the study showing that the agents it tested finish roughly a third of multi-step sales and service tasks. CRMArena-Pro put nine frontier models against nineteen expert-validated sales, service and quoting tasks across 4,280 queries. Single-turn success landed around 58 per cent. In multi-turn settings, where an agent has to carry context across several steps the way a real sales task does, success fell to about 35 per cent. One analyst quoted by CIO.com put it bluntly: a 35 per cent success rate in multi-step workflows is a non-starter for enterprises.
Read that as a verdict on the technology and you would conclude that agents do not work. Read it as a verdict on the conditions and you find what the same study also measured, which is success above 83 per cent when the steps were defined in advance. The agents are not failing. They are being asked to decide things nobody has told them how to decide.
So the useful question is not which AI sales agent is best. It is which parts of a sales job can be defined clearly enough for an agent to finish, and what it needs to see in order to finish them. That is answerable stage by stage, and harder than it looks, because three quite different products are sold under this one name and they fail in different ways.
What an AI sales agent actually is
An AI sales agent is software that takes a sales goal, decides the steps itself, and acts on them using your systems and data, with limited human input. The published definitions converge on that. Salesforce describes agents as applications that analyse your sales and customer data to perform tasks with little or no human input. IBM draws the line more usefully: assistants are reactive and act when asked, while agents are proactive, planning and taking action toward a goal using the tools and permissions available to them. Creatio frames it as going beyond passive automation to act autonomously rather than assist.
Those definitions are written by organisations selling into the category, and they describe intent rather than behaviour. One test cuts through all of it.
What can the software do without a human click?
If it drafts an email that a rep then sends, it is an assistant with good automation. If it decides who to contact, writes the message and sends it, it is an agent. The distinction matters because the second one can be wrong in public. The fuller version of that distinction is worth reading before you sit through a demo.
Worth noting that IBM's own phrasing is careful. Agents reduce the need for human intervention at every stage. Reducing is not removing, and no shipping product in this category claims otherwise once you read past the homepage.
Three different products share this name
This is the part that makes comparison shopping so confusing. Ask three platforms for a demo of their AI sales agent and you will see three unrelated things.
Several of Microsoft's agents are still in preview rather than general availability, which is worth confirming for any you are shown.
Which one you need follows from your bottleneck, not from a shortlist. If reps lack pipeline and your ICP data is clean, the outbound group is the relevant one. If inbound sits unworked, follow-up is slow and CRM hygiene is poor, start with the CRM-native group, where grounding is best and permissions are inherited. If reps cannot find content or take too long to ramp, it is the enablement and knowledge group. These three are not substitutes, and a brief sent to all three comes back as three unrelated proposals.
The content layer the third group operates on. An agent is only as good as what it is allowed to see.
What the benchmark actually measured
Most capability claims in this category come from the organisation making them. CRMArena-Pro is the exception worth knowing, and it is worth knowing precisely.
Salesforce AI Research built it with nineteen expert-validated tasks spanning sales, service and configure-price-quote work, then ran nine frontier models across 4,280 queries. Three findings matter for anyone buying.
Single-turn success was around 58 per cent. Ask an agent to do one bounded thing and it succeeds a little over half the time.
Multi-turn success fell to roughly 35 per cent. Real sales work is multi-turn almost by definition. A qualification conversation, a follow-up sequence that adapts to a reply, a quote that depends on three earlier answers.
Workflow execution scored above 83 per cent. This is the number from the opening, and it is the one to build a deployment plan on. When the steps are defined in advance and the agent executes them, it is reliable. When the agent has to decide the steps, it is not.
The researchers also noted that agents demonstrated low confidentiality awareness, meaning they may surface information they should not.
Read those together and a deployment strategy falls out. Give agents bounded, well-defined work with clear inputs. Do not give them open-ended judgment and expect it back correct.
Stage by stage: what works now
Combining that benchmark with product documentation and public operator accounts gives a realistic map of a sales cycle.
Evidence strength above is an editorial read of the benchmark, published product documentation and public operator accounts taken together, not a formal rating. Two patterns run through that table. Agents are strongest where the task is bounded and the data is clean, and they are weakest exactly where sales gets interesting. And in every row the human job shifts from doing the work to defining the boundaries and checking the output.
Agent washing, and the test for it
Gartner coined the term for what is happening to this category: the rebranding of existing products, including assistants, robotic process automation and chatbots, without substantial agentic capability underneath. It has estimated that of the thousands of organisations claiming agentic AI, only around 130 genuinely qualified. Gartner has not published how it reached that number, so treat the figure as directional, but the pattern it names is real and easy to check.
Three questions separate a real agent from a renamed one:
What happens with no human in the loop at all? Ask them to turn approval off in the demo and show you what runs. Many products cannot, which is useful to know and not necessarily a problem.
Can it handle a case nobody scripted? Workflow execution scores above 83 per cent precisely because the steps were defined. Ask for something outside the script.
Will it run on your data, today? Not a sandbox, not a curated demo set. This is the single most revealing request you can make, and the answer tells you more than any feature list.
Why agents are confidently wrong
Language models produce plausible text. Where they lack a fact, they produce something that reads like one. In a sales context that is not an abstract risk, because the output goes to a buyer.
Salesforce's own description of its agents makes accuracy conditional: they deliver accurate, personalised output only when they fully process organisational systems, customer details and deal specifics. That conditional is the whole ballgame. An agent with an ordinary model and clean, permissioned access to your price book, approved content and CRM will outperform a stronger model working from whatever it can scrape.
Two documented cases show the shape of the failure.
At the developer tool Cursor, an AI support agent told users that being signed out across devices was expected under a new one-device policy. No such policy existed. The cause was a bug, the explanation was invented, and customers cancelled publicly before the company confirmed that it had no such policy and committed to labelling AI responses.
The second is quieter and more common. A team running agents at scale reported one promoting an event that had already taken place, and another giving away free tickets during an A/B test. A model has no built-in sense of today's date, and no sense of what it is not authorised to offer, unless someone writes those checks in.
The pattern in sales is the same with a longer fuse. An agent asked about a price, a term or an integration that is not in its reach will answer anyway, and the buyer has no way to tell.
An answer returned from a governed library with its source attached. The source is the part that makes the answer checkable.
Before you shortlist anything, take the ten questions your team answers most often for buyers and ask each platform to answer them from your own content. We will run that set against your library and you keep the results either way.
Who is liable for what the agent says
One ruling has already tested this, and its reasoning travels further than the case itself.
In Moffatt v. Air Canada, an airline's chatbot gave a customer incorrect information about bereavement fares. Air Canada argued it could not be held responsible for information provided by one of its agents, servants or representatives, including a chatbot. The tribunal called the suggestion that the chatbot was a separate legal entity responsible for its own actions remarkable, and ruled that it should be obvious a company is responsible for all the information on its website. It made no difference whether that information came from a static page or a chatbot.
Two caveats matter. This was a small-claims tribunal, so the decision is persuasive rather than binding precedent, and the award was $812.02. Neither changes the reasoning, which is the part that applies to any agent quoting a price, a term or a commitment to a prospect. Whatever the agent says, you said.
That is why the approval boundary is a commercial decision rather than a technical setting. Where human approval belongs in an agent workflow goes through the specific points worth gating.
The oversight nobody budgets for
The most detailed public operator account of running sales agents at scale comes from a company that documented its own deployment in the open. Its numbers are worth reading with the caveat that it is one company with an unusual motion, not a benchmark.
Setup ran to fifty or sixty hours, with roughly a month of training before the agents were useful. Ongoing supervision ran fifteen to twenty hours a week, and performance tracked human attention closely. Performance tracked human attention closely enough that the team described it as ebbing and flowing with how much supervision it got.
Independent research points the same direction from a different angle. A Gartner survey of sales leaders published in May 2026 found AI saves sellers an average of 4.8 hours per week, and that 72 per cent of sales organisations report low reinvestment of that reclaimed time into high-value selling. Organisations that did reinvest were materially more likely to beat their growth and conversion goals. The Gartner analyst's summary is the line to take into a budget conversation: AI is not the hero of this story, AI is the accelerant.
The spread underneath those averages is the number worth carrying into a business case. Reporting on the same survey, 25 per cent of sales organisations said their AI investment returned more than 50 per cent, while 20 per cent reported negative returns worse than 50 per cent. Roughly as many lost heavily as won heavily. That variance, rather than any average, is what a pilot is actually testing.
Gartner has also predicted that by 2027, 40 per cent of enterprises will demote or decommission autonomous AI agents because of governance gaps found only after something goes wrong in production, and separately that over 40 per cent of agentic AI projects will be cancelled by the end of 2027. Both are forecasts rather than measurements. Both point at the same missing line item, which is the cost of supervision.
Supervision needs somewhere to look. A program view is what turns oversight into something a manager can do weekly.
The consolidation nobody mentions
Worth factoring into any evaluation that runs past this quarter.
Three of the four best-known enablement platforms have lost independence. Showpad and Bigtincan merged. Highspot and Seismic signed a merger agreement in February 2026, and Highspot became a product line inside Seismic in August 2026. Each of those platforms had shipped its own agent lineup shortly before.
None of that makes the products worse. It does mean that agent roadmaps across the merged stacks are genuinely uncertain through 2027, and that a roadmap commitment made during a sales cycle may belong to an organisation that no longer sets the roadmap. Ask which entity owns the agent roadmap, and ask what happens to it at renewal.
How to evaluate one
Drawn from the benchmark, the analyst guidance and the case law rather than from the product pages.
Ask what runs with no human in the loop. The answer defines what you are buying and what you are liable for.
Ask for multi-turn evidence, not a demo. Demos are single-turn and scripted. Ask how it performs when a buyer replies with something unexpected three steps in.
Ask what it can see and what it may say. Which content, which CRM fields, which price book, and what is explicitly out of reach. Confidentiality awareness was a named weakness in the benchmark.
Ask what the controls are. Draft-and-review mode, per-action approval for anything customer-facing, caps on what the agent may offer, and validation of dates and facts.
Ask what governance looks like after go-live. Monitoring, rollback, and a named owner for agent behaviour.
Verify references directly, and ask about renewal. Reported customer counts and logo walls have both been contested in this category. Speak to customers in your own segment, in your own motion, and ask what they renewed rather than whether they liked it.
Ask what human time it needs. Setup hours and weekly supervision, in writing.
The broader version of this, covering the whole category rather than agents specifically, is in how to evaluate revenue enablement software.
Where pf1 fits
pf1, from the house of Paperflite, is an agentic growth team: six role-based agents working on one body of knowledge, grounded in what your business knows to be true. Grounding is the whole design premise rather than a feature, which is the direct answer to the failure mode in this article.
Run it against the questions in the section above, which is the only fair way to read any platform on this page.
The rest of the team works off the same graph. Coaching diagnoses each rep's gaps and builds practice from your own deal history, with roleplay drawn from the objections that stalled real deals. Deal reads every channel on every live deal and scores it against how deals win and lose at your company. Lead captures qualified leads from traffic you already have, building pages from approved knowledge. Experience turns a prompt into a finished interactive experience built from approved content. Every agent reads from the graph and writes back into it, so the ground they stand on improves with use.
Most platforms will answer that table in a deck. Ask us to answer it live, on your content, in one session.
What to do first
Start where the evidence is strongest and the blast radius is smallest.
Pick a bounded job, not a role. Inbound lead response, account research and meeting prep are where agents demonstrably work. Replacing a seller is not a first project.
Fix what the agent can see before you turn it on. Every failure mode in this article traces back to an agent answering from something other than your validated knowledge. That work is worth doing whether or not you buy an agent, which makes it the safest money in the project.
Budget the supervision. Fifty to sixty hours of setup and a standing weekly commitment, with a named owner.
Decide the approval boundary in writing. What the agent may send, offer and commit to, before anything reaches a buyer.
The teams getting value here are not the ones who bought the most capable agent. They are the ones who gave an ordinary agent clean ground to stand on, and then reinvested the hours it gave back.
What is an AI sales agent?
Software that takes a sales goal, decides the steps itself and acts on them using your systems and data, with limited human input. The practical test is what it does without a human click: software that drafts a message for a rep to send is an assistant, while software that decides who to contact and sends the message is an agent.
Can AI agents do sales?
They can do parts of it reliably and other parts poorly. Salesforce's CRMArena-Pro benchmark found the agents tested succeeded around 58 per cent of the time on single-turn tasks and about 35 per cent on multi-turn ones, while scoring above 83 per cent on defined workflow execution. In practice that means bounded work such as inbound response, account research and meeting prep works today, while open-ended discovery, negotiation and forecast calls do not.
Which AI agent is best for sales?
The question hides a category problem, because three different products share the name. Outbound agents run outreach to people who have not met you. CRM-native agents work inside your system of record on records and inbound leads. Enablement and knowledge agents handle content, coaching and deal risk without contacting prospects. Decide which of those three jobs is your bottleneck first, because they are not substitutes and are rarely compared against each other.
What are the big 4 AI agents?
There is no established big four in sales specifically. The agents most often named sit inside the major CRM platforms, including Salesforce Agentforce, Microsoft's Dynamics 365 Sales agents, HubSpot Breeze and Creatio, alongside enablement lineups such as Seismic's Aura Agents and Mindtickle's ElevateOS. Any ranked list in this category is worth checking against the three product types before treating it as a shortlist.
How much do AI sales agents cost?
Almost nothing in this category publishes a price, so the comparable number has to be requested. What differs structurally is the unit being charged for. Outbound agents tend to meter capacity such as contacts, sends or call minutes. CRM-native agents increasingly meter agent work itself, charging per action or per outcome on top of existing licences. Enablement platforms tend to sell seats plus a platform fee on an annual contract. Ask for the annualised total including set-up, and ask what happens when usage runs above the plan.
What is agent washing?
Gartner's term for rebranding existing products, including assistants, robotic process automation and chatbots, as agents without substantial agentic capability underneath. It has estimated that only around 130 of the thousands of organisations claiming agentic AI genuinely qualified. The quickest test is to ask what the software does with approval switched off, and to ask it to handle a case nobody scripted.
Do AI sales agents need human approval?
Shipping products generally offer both modes, and the safe pattern is draft-and-review for anything customer-facing, loosened gradually as evidence accumulates. HubSpot's prospecting agent documents both a review-before-send mode and an automatic send mode. Treat autonomy as a dial you set per action rather than a property of the product.
Who is responsible if an AI agent gives a customer wrong information?
You are. In Moffatt v. Air Canada, a tribunal rejected the argument that a company's chatbot was a separate legal entity responsible for its own actions, and ruled that it makes no difference whether information comes from a static page or a chatbot. The principle applies to any agent quoting a price, a term or a commitment to a prospect.
Why do AI sales agents give wrong answers?
Because a language model that lacks a fact will produce something that reads like one. Accuracy depends on what the agent can actually see: your approved content, current pricing, product detail and CRM state. An ordinary model with clean, permissioned access to validated company knowledge will outperform a stronger model working from whatever it can find, which is why data access rather than model choice usually decides the outcome.
How much human oversight does an AI sales agent need?
More than most plans allow for. The most detailed public operator account reports fifty to sixty hours of set-up, around a month of training before the agents were useful, and fifteen to twenty hours a week of ongoing supervision, with performance tracking how much attention they got. Budget a named owner and a standing weekly commitment rather than treating deployment as a project that finishes.
What should you ask when evaluating an AI sales agent?
What runs with no human in the loop, how it performs on multi-turn tasks rather than scripted demos, what data it can see and what it may say, what approval controls and offer caps exist, what governance and rollback look like after go-live, how much human time it needs, and who owns the agent roadmap. Ask to run it on your own content rather than a demo set.