AI agents for sales: automate prospecting, follow-up and quoting

Almost every sales team I audit has the same hole in it, and it is not in the pitch: it sits between the moment someone fills out a form and the moment a human replies. That is where prospects you already paid advertising money for fall through. An AI agent does not sell more because it is more persuasive than your best rep; it sells more because it answers in seconds, never forgets a follow-up, and keeps the CRM clean without anyone typing anything in by hand.

The first three numbers are the sales pitch; the fourth is why this article spends as much time on limits as on use cases. None of them comes from a controlled experiment inside a Mexican sales team: they are an observational study from almost twenty years ago, a global self-reported survey, a consultancy's experience with its own clients, and an experiment run in strategy consulting. They are useful for calibrating, not for promising.

An AI sales agent receives a goal (respond to this lead, prepare this quote, revive this dormant account) and executes a sequence of steps using real tools: it reads and writes to the CRM, queries the price catalog, sends a WhatsApp message, books time on a rep's calendar, escalates to a human when it should. A chatbot returns text. An agent changes the state of your systems.

The distinction is not semantic, it is architectural. You test a chatbot with a conversation; you test an agent with permissions, with idempotency (so it does not create three opportunities from the same form), with a log of every action, and with a plan for when the CRM goes down while the operation is running. Everything interesting about a sales agent happens in the integration, not in the prompt.

If you could only automate one thing, it would be the first reply. The classic reference is a lead response study published in 2007 by InsideSales.com with Dr. James Oldroyd, then a Faculty Fellow at MIT Sloan, covering three years of data from six companies, more than 15,000 web leads and more than 100,000 call attempts. The finding: the odds of contacting a prospect drop 100 times when you call at 30 minutes instead of 5, and the odds of qualifying them drop 21 times; within the first hour the odds of contact are already more than 10 times lower ².

With all those asterisks attached, the direction of the effect matches what you see in any operation: the prospect filled out your form because they are evaluating right now, they probably filled out two competitors' forms as well, and their attention lasts minutes. The same study reports that after the first 20 hours, each additional call attempt starts to reduce the probability of qualifying the prospect ². Persisting later does not make up for arriving late.

Speed-to-lead breaks for boring reasons: the form lands in a shared inbox nobody checks after 6 p.m., the WhatsApp message goes to a phone carried by someone who is out on a site visit today, or the round robin assigns the lead to a rep on vacation. An agent fixes that: it acknowledges within seconds, asks two or three qualifying questions, offers real slots from the calendar and keeps the conversation warm until a human picks it up.

This is the order worth attacking it in. Each stage is a different system even when it shares a model, and each one launches and gets measured separately. Do not attempt all six at once.

1. 1

   1. Prospecting and enrichment

   The agent takes a list (content downloads, webinar attendees, companies from a directory) and enriches it: normalizes the legal name, identifies industry and size, finds relevant public signals and discards duplicates against your CRM. The output is not more leads: it is a list with the fields your qualification criteria need, already filled in. The heavy lifting here is data work, not AI.

2. 2

   2. First reply in minutes, not hours

   A form, a WhatsApp message, an email to sales or a missed call comes in, and the agent responds immediately with a message that actually says something about the specific case, not a generic acknowledgment. This is the stage where the evidence on response speed hits hardest ² and the easiest one to measure: median time from lead arrival to first useful reply, before and after.

3. 3

   3. Qualification against explicit criteria

   The agent asks the three or four questions your team always asks (volume, timeline, approximate budget, who decides) and maps them to a score with rules you can read. The escalation threshold is what matters: large account, non-standard terms or a technical question jump to a human. An agent that over-qualifies saves you time; one that over-disqualifies costs you the quarter.

4. 4

   4. Follow-up that does not collapse

   The most underrated stage and the simplest to build. The agent keeps the sequence alive, detects replies that are really objections and escalates them, and stops when the prospect says no or when a meeting is already booked. The value is not in writing better messages: it is that no account is ever left without a next step.

5. 5

   5. Quote generation

   The agent assembles the draft: it cross-references the requirements from the conversation against the catalog and the pricing rules currently in force, applies the discounts that fall inside policy and generates the document in your format. Hard rule: prices are read from a system, never generated. The model writes the copy; the number comes from a query. And it goes to human review before it is sent.

6. 6

   6. CRM updates

   After each interaction the agent writes the activity log, updates the stage, records the loss reason using a closed taxonomy and schedules the next activity. It sounds minor and it holds up everything else: the forecast is only useful if the CRM reflects reality, and that cannot depend on a tired rep typing things in at 8 p.m.

This is where it pays to turn the volume down. Most of the figures circulating about AI in sales come from consultancies reporting their own experience with clients, with no published methodology and no control group. That does not make them useless, but it does make them something other than causal evidence:

The honest summary: I do not know of a published controlled trial on AI agents operating inside a sales team. The closest thing is an NBER field study in customer support: at a single Fortune 500 business process software company, with 5,179 chat-based support agents, the staggered rollout of a conversational assistant increased issues resolved per hour by 14% on average ¹. That is support, not sales, and it is one company, not an industry average. The interesting part is how that gain was distributed.

The best evidence on the limits comes from a preregistered field experiment with 758 knowledge workers, run in collaboration with Boston Consulting Group. On 18 tasks placed inside the model's capability frontier, those who used GPT-4 completed 12.2% more tasks and did them 25.1% faster, with significantly higher quality. On a complex managerial task chosen deliberately to sit outside that frontier, those who used AI were 19% less likely to reach the correct solution ⁷. The setting is strategy consulting, not sales: what transfers is not the percentages but the concept of a jagged frontier. There are tasks where AI helps a lot, and adjacent tasks that look just like them where it makes the result worse without warning you.

The second finding takes us back to the NBER study cited above: that 14% average improvement was 34% among novice and lower-skilled agents, with minimal effect on the productivity of the most experienced ones, and the authors find indications that the tool may reduce the quality of conversations for the most skilled agents ¹. Translated into a commercial operation: your new rep will probably gain far more than your best rep, and forcing the tool on the latter can backfire. Design the rollout with that in mind instead of imposing it uniformly. These are the failure modes that show up again and again:

The order matters more than the tools. These are the prerequisites I check before writing a line of code:

• A single source of truth for prospect status. If the pipeline lives in the CRM and also in a spreadsheet, fix that first.
• Written qualification criteria. If your team cannot explain them in five lines, the agent will not be able to apply them.
• Catalog and pricing rules queryable via API or database, with effective dates. A price list PDF does not work.
• An explicit escalation threshold: deal size, non-standard terms or a risk signal jump to a human immediately.
• Scoped permissions: the agent writes only to the fields it needs, and cannot delete or close opportunities.
• A log of every action and every message, readable by the sales manager without asking IT for help.
• A baseline metric measured before you turn anything on. Without last week's number, any result will be arguable.

On timelines: first reply and follow-up are usually in production in three to five weeks if the integrations exist. Quoting takes longer, because you almost always discover that the pricing rules have exceptions that only lived in two people's heads. That discovery is often worth more than the agent.

Four metrics are enough, and all four have to exist in your CRM before launch so you have something to compare against:

• Median time to first useful reply, from the moment the lead arrives. Median, not average: one lead forgotten for three days hides the real improvement.
• Follow-up coverage: what percentage of open opportunities have a next activity scheduled. This is the one that moves fastest.
• Effective contact rate by weekly cohort, not cumulative. If the agent is doing its job, this rises before conversion does.
• Hours returned to conversations: how many hours a week stopped going into data entry, document assembly and writing follow-ups.

What I do not recommend measuring at the start is total funnel conversion: it moves late, it depends on season, price and competition, and it starts an argument you will not win with three months of data. Begin with what the agent actually controls.