AI for sales: what it does well, where it fails, and how to test it

This guide is for sales leaders, SDR managers, founders and revenue operations people deciding where artificial intelligence belongs in their sales process. It covers what the technology does, which tasks to give it, how it fails and a test you can run in a week. For autonomous tools that send outreach on their own, see the guides to AI SDRs and AI sales agents go deeper on those.

What AI actually does in a sales team

"AI for sales" covers three different technologies. They fail in different ways, so it helps to know which one a tool uses.

Most products marketed as AI sales tools are a language model wrapped around a data source. The data source matters as much as the model: a model writing about a company it can't look up will fill the gap with plausible guesses.

Where sellers' time goes, and where AI fits

Salesforce's seventh State of Sales survey (4,050 sales professionals in 22 countries, fieldwork August to September 2025) found reps spend about 40% of their week selling and 60% on everything else, with almost a full day a week on prospecting. Nearly half (47%) called cold outreach one of the worst parts of the job. Sellers expected agents to cut prospect research time by 34% and email drafting time by 36% once fully implemented. Those are hopes from a vendor's survey, not measured savings.

The best controlled evidence comes from outside sales. A study of 5,179 customer support agents by Brynjolfsson, Li and Raymond (NBER, 2023) found an AI assistant raised issues resolved per hour by 14% on average, by 34% for novice and low-skilled workers, and had minimal impact on the most experienced. If that pattern carries over to selling, expect the biggest gains for new reps learning your accounts, and expect your best sellers to notice it least.

Tasks to give AI, and tasks to keep

The pattern: AI does well where the output is text or a list that a person can check in less time than it took to make. It does poorly where a mistake is expensive and hard to spot, such as a discount, a commitment or a judgment about a relationship.

Failure modes to check

• Invented facts. Language models produce fluent text whether or not they know the answer. A 2025 paper by Kalai and colleagues argues this is built in: training and evaluation reward guessing over admitting uncertainty. Ask for a source on every fact and open a sample of them.
• The wrong company. Two firms with similar names, or a parent and its subsidiary, get merged into one record. In GTM-Bench, Blackpearl Group's June 2026 benchmark of AI prospecting agents (Blackpearl owns Bebop), identity errors and unsupported claims were among the main causes of failure.
• Too many weak records. GTM-Bench scored each returned company +1 if it was a strong, well-evidenced match, 0 if it was usable, and minus 1 below that. Four of the six generalist AI agents tested finished with a negative net score, often because they returned hundreds or thousands of rows when a few dozen good ones would have scored better. More output is not more pipeline.
• Stale data. A trigger event from two years ago, or a contact who left in the spring, makes an email wrong on arrival.
• Sameness. If every rep uses the same prompt, prospects get the same email from five vendors.
• Compliance at scale. The FTC's CAN-SPAM guidance says the law "makes no exception for business-to-business email", sets penalties of up to $53,088 per violating email, and says you can't contract away your responsibility when a tool or agency sends for you. Google's sender guidelines ask every sender to keep its reported spam rate below 0.3%, and recommend below 0.1%.
• Your data in the wrong place. Pasting customer contracts or call recordings into a consumer AI tool may break your customer agreements. Check where the data goes and whether it's used for training.

Worked example: testing AI account research

Illustrative example. The company, people and numbers are invented to show the method.

The team: Northbeam HR (northbeam.example) sells HR and payroll software to US companies with 50 to 500 employees. Four SDRs each research about 60 new accounts a week before writing to them, at about 12 minutes per account.

The baseline: 60 accounts × 12 minutes = 12 hours per SDR per week, or 48 hours across the team.

The test: the manager draws 100 accounts from next week's list and splits them at random. SDRs research 50 the usual way. An AI tool drafts summaries for the other 50, and SDRs check and correct each one, which takes 5 minutes per account including generation. A team lead who doesn't know which summaries came from where grades all 100: A (accurate, sourced, useful), B (usable with gaps) or below B (wrong company, invented fact or out of scope).

The AI route produced more good research per hour, but 7 of 50 summaries (14%) were still wrong after checking. Reading them showed one cause: five were about a different company with a similar name. The fix was a rule that every summary must match the account's web domain. On a fresh 50, the below-B count fell to 4 (8%), inside the threshold the manager had set in advance (under 10%).

The payoff: at 5 minutes instead of 12, each SDR saves 7 hours a week, 28 hours across the team. At an illustrative loaded cost of $40 an hour, that's $1,120 a week, or about $4,850 a month, to set against the tool's price and the manager's time spent grading.

A sample of 50 is a floor, not proof. If 5 of 50 are wrong, the true error rate could plausibly be anywhere from about 4% to 21% (a 95% Wilson interval). Keep grading a sample every week after rollout.

A one-week test you can run on your own pipeline

1. Pick one task with a checkable output. Account research, list building or call notes. Not "use AI for prospecting".
2. Write the grading rules before you start. What makes a record A, B or below B, in one line each. Decide the pass mark now, such as "below-B rate under 10% and time per item down by a third".
3. Draw 100 items from real upcoming work and split them at random into two groups of 50.
4. Do one group the current way and time it.
5. Run the tool on the other group, then have the person who'll use the output check and correct it. Time the whole thing, checking included.
6. Grade all 100 blind. The grader shouldn't know which items came from the tool.
7. Compare net quality and net per hour, then read every below-B item. Errors usually cluster in one or two causes you can fix with a rule.
8. Decide using the rule you wrote in step 2, and keep grading 20 items a week after rollout.

What to ask a vendor

• Where does the company and contact data come from, and how old is the oldest record you'd show me?
• Does each fact come with a source I can open?
• What happens when the tool can't find something: does it say so, or fill the gap?
• Will you run your tool on 50 of my accounts so I can grade the output before I sign?
• Is my data used to train your models or anyone else's? Can I turn that off in writing?
• What can the tool do without a person approving it, and can I limit that?
• What exactly am I charged for: seats, credits, actions, conversations or results?
• Who is responsible for opt-outs, sending limits and do-not-contact lists?

Be wary of guaranteed results. In August 2025 the FTC sued Air AI, alleging it sold small businesses "conversational AI" pitched as able to replace human customer service representatives, with false earnings and refund promises. These are allegations, not findings.

What it costs

Pricing follows three models, and the unit decides what a pilot really costs.

Illustrative arithmetic: an agent that takes six actions to research and draft for each account, across 1,000 accounts a month, uses 6,000 actions. At $0.10 an action that's $600 a month before anyone checks the output. Then add the hidden costs: checking time, cleaning the CRM data the tool depends on (46% of sales pros with agents told Salesforce that data quality issues hurt their sales) and the sending setup that protects deliverability.

Self-check before you roll out

• You have a baseline for time and quality from before the tool.
• The pass mark was written down before the test.
• Every fact the tool produces carries a source a rep can open.
• Nothing reaches a prospect without a person approving it in the first month.
• Opt-outs and do-not-contact lists flow into the tool automatically.
• Someone grades a weekly sample after rollout.

Next step: see how the main categories of AI sales tools compare, and which ones fit the task you picked.