The short version
In May 2024 I wrote that AI would revolutionize the salesperson's role. The direction was right, and the piece was a tour of tools and statistics I had not checked. Two years of putting AI to work on how I research and work accounts have given me a better argument.
- Sellers still spend most of the week on work that is not selling. Salesforce's 2026 survey of 4,050 sales professionals puts selling at 40 percent of the week.
- AI earns its place by carrying load: research, first drafts, records and reporting. Judgment stays with the seller, including what each buyer believes, which deals to kill, what a customer sees in the seller's name and the number in the forecast.
- Adding agents does not add selling by itself. Gartner predicts agents will outnumber sellers ten to one by 2028, with fewer than 40 percent of sellers saying agents improved their productivity.
- In outbound, the approval loop is the design: nothing goes out without a person's approval, and a human takes over at the first reply.
- The first question enterprise buyers ask of AI is now what it is permitted to do and how they will know what it did. Sales teams that use agents have to answer it about their own tools first.
1. The 2024 piece named the right tools and missed what they were for.
The original walked through predictive analytics, lead scoring, chatbots and personalization, with a figure from one research firm here and another there. Most of those tools arrived roughly as described, and a model now drafts a personalized first note in seconds. I give the piece credit for the direction.
Two things were wrong with it. The statistics were borrowed, and I had not traced a single one back to its source, which is the opposite of what I now ask of any account plan. I have removed them. The bigger problem was the frame. I wrote about AI as a set of tools a sales leader buys for the team, as if the hard question were which tools to buy. The question that matters is where the line sits between the work AI should carry and the work the seller has to keep, and how anyone can see that line once the AI is acting on the seller's behalf.
2. Most of a seller's week still goes to work that is not selling, and that is where AI pays.
Salesforce's State of Sales report for 2026, built on a survey of 4,050 sales professionals in 22 countries during August and September 2025, found that the average seller spends 40 percent of the week selling. The other 60 percent goes to work around the sale, such as hunting for material and chasing approvals. Younger sellers fared worse, with Gen Z reps selling 35 percent of the time.
The same survey asked what sellers expect once AI agents are fully in place. They expect prospect research time to fall by 34 percent and email drafting time by 36 percent. Those are expectations rather than measurements, and I read them that way. Still, they point at the right target. Research and drafting are load. Both take hours and follow a pattern, and a model can do a first pass of each with every source attached.
That lower bar carries the part most AI pitches leave out. Hours that leave admin do not all go to selling. Some of them become review, which means reading what the AI produced and then approving or rejecting it. I count that as a good trade, because review is where the seller's judgment enters the work, and an hour spent checking a sourced draft is cheaper than an hour spent writing one from nothing.
3. AI should carry the load, and the judgment has to stay with the seller.
Here is the line I draw. Load is work that follows a pattern and can be checked: reading an account, drafting a first note, finding accounts from signals, keeping records and building the report. Judgment is work where the answer lives in people or carries consequence, such as what each buyer actually believes about the problem, which deals to pursue and which to kill, what a customer sees in your name, and the number you call in the forecast.
Account research is where I moved first. Before a first meeting I now have AI read an account the way I used to read it by hand: the filings, the earnings calls, the news, what the company's leaders say in public, and the roles it is hiring for. Every figure it brings back carries its link. Anything I could not verify is marked as a hypothesis, and anything that is my judgment is labeled as mine.
The speed is welcome, but checkability is the bigger change. A plan built this way holds up when an executive asks where a number came from, and a seller who walks in with a sourced point of view gets a different meeting from one who walks in with a deck. Here is how a page of that research reads.
| Line in the account plan | Where it came from | Label |
|---|---|---|
| Revenue grew in the last fiscal year, and margins narrowed | Annual report, linked to the page | Sourced |
| The CFO named the cost of serving customers as a priority | Latest earnings call transcript, linked with the timestamp | Sourced |
| The company is posting several data engineering roles | The company's own job board, linked to each posting | Sourced |
| The push to modernize started with an audit finding | No public source | Hypothesis, to test in the first meeting |
| This account deserves a quarter of my time | My read of the lines above | My judgment |
Look at the last two rows. Everything above them came from the AI, and every line carries a link anyone can click. The hypothesis and the call on whether the account deserves the time stay with me, and the labels make that visible to anyone reading the plan. When the hypothesis is wrong, I find out in the first meeting, which is the cheapest place to find out.
4. More agents will not mean more selling unless someone decides where the time goes.
Gartner has put a sharp number on the risk. It predicts that by 2028 AI agents will outnumber human sellers ten to one, and that fewer than 40 percent of sellers will say agents improved their productivity. Its warning is agent sprawl, meaning more digital activity with little change in what sellers achieve. I agree, and I think the cause is plain. An agent that carries load only helps if someone has decided what the freed time is for.
That decision is a reporting job, and it is one I have run for most of my career. At NTT in 2016 and 2017, a Salesforce reporting program we called Data Driven Decisions decided where cross-sell resources went. The region I led was Top Region in the NTT Group in each of 2016, 2017 and 2018. Forecasting worked the same way for me, with accuracy coming from visibility into every deal through Salesforce and into the team through a weekly cadence.
AI makes that reporting cheaper to build and harder to argue with. The report I want shows each seller where their hours went last week and which deals have not moved since the last review. It replaces a debate about effort with a conversation about where to spend the next week. I would go further with the meetings themselves. My weekly team meeting ended in open breakout sessions we called Blue Sky, and today I would run the recordings through AI into one consolidated read-out, so that nothing said in a breakout room gets lost.
5. In outbound, the approval loop is the design.
Outbound is where AI can damage a brand fastest, so it is where the design has to be strictest. SMARTHAUS built Prospect Forge as an outbound tool for revenue teams that sell AI, and a company can put it to work today. It finds accounts from hiring signals on public job boards, researches each one, drafts an opener grounded in what it found, and runs every draft through an automatic check for verified addresses, required compliance lines and sending limits.
Two rules sit on top of that, and they matter more than everything beneath them. Nothing is sent without a person's approval, and the moment a real person replies, a human takes over the conversation. The engine is built, and this article makes no claim about what it has produced.
A hypothetical shows why the approval step exists. Say the engine finds a company posting several roles for a new AI platform team and drafts an opener tying those roles to a problem worth a conversation. The draft passes the automatic check. On review, though, the posting it cites turns out to be three months old, and the company has since announced a reorganization that the draft does not mention. Nothing in the draft is false, so no automatic check could catch it. The reviewer rejects it and sends the account back for fresh research, and the prospect never sees a note that made the sender look as if nobody had done the homework.
The handoff at the first reply follows the same logic. A reply is the start of a relationship, and a relationship runs on judgment from the first sentence. An AI can prepare the seller for that moment with a summary of everything the research found, but the answer itself comes from the seller.
6. I watched AI change selling from inside AWS before I wrote the 2024 piece.
The 2024 article read as a forecast, though part of it was already behind me. At AWS in 2023 I co-led the initiative behind integrating MongoDB into Amazon CodeWhisperer, an AI coding assistant. The two companies announced it on 6 November 2023. MongoDB supplied curated training data and helped evaluate the output, so that CodeWhisperer would give better suggestions to developers building on MongoDB in C#, Go, Java, JavaScript and Python.
I saw the integration as a way to make MongoDB easier to build on, which is a selling argument any developer understands, and it is the version of AI in sales I trust most: AI pointed at the customer's problem, with a person deciding which problem matters. Around the same time, at t-emgee, the company I founded, I built LAMP, a procurement-as-a-service platform with a generative AI pre-sales assistant built in, and sold the platform design to a Northeast regional systems integrator. Both taught me the same lesson from different sides. AI earns its place in a sale when it makes the buyer's work easier, and it loses the sale when the buyer cannot tell what it will do.
7. The first question enterprise buyers ask of AI is about trust, and sales teams have to answer it about their own tools first.
The enterprise buyers I talk to ask two questions first: what is this agent permitted to do, and how will we know what it did? Capability comes later. A policy wrapped around a model after the fact will not answer either question in front of a regulator or a board.
The research explains why the questions come first. McKinsey's 2025 State of AI survey found that 88 percent of organizations use AI regularly in at least one function, and 23 percent are already scaling an agentic AI system somewhere in the enterprise. More than half reported at least one negative consequence from AI in the past year, with inaccuracy the most common. Gartner, for its part, predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Adoption has run ahead of trust, and buyers know it.
I founded SMARTHAUS, an IP company, on exactly that question in December 2023, and it has shaped how I think about selling with AI as much as selling AI. A sales team that uses agents is now on the receiving end of the same question. When the seller's own agent writes to a customer, the customer's security team is entitled to ask what that agent was allowed to do and who approved what it sent. With an approval loop and a record, the team can answer in a sentence. Without them, it has a trust problem sitting in the middle of its own pipeline.
For a seller whose product is AI, this changes the first meeting. Leading with what the system can do puts capability in front of a buyer who is still worried about control. Opening with what it is permitted to do, and how every action it takes is recorded, answers the question the buyer walked in with.
8. What this means for sales leaders
- Measure the week before you buy anything. Find out where your sellers' hours actually go. If the answer looks like Salesforce's 40 percent, the first job for AI is the other 60, and the tool list follows from that.
- Write down the line between load and judgment. Name the work AI may do on its own and the work it may only draft for a seller to approve. Everything else stays with the seller, and sellers adopt AI faster once they know which decisions remain theirs.
- Put a person on every action a customer sees. Require approval before anything is sent, and hand the conversation to a human at the first reply. The approval step is where judgment enters the work, and it costs minutes.
- Make every figure carry its source. Ask for the link behind every number in an account plan, and label judgment as judgment and hypotheses as hypotheses. A plan built that way survives the executive who asks where a number came from.
- Report on capacity and outcomes, and stop counting activity. The report should show each seller where the time went and where it should go next. More agents producing more activity is the sprawl Gartner warns about.
- Answer the trust question about your own stack before a buyer asks it. Know what every agent your team uses is permitted to do, and keep a record of what it did.
If I wrote the 2024 piece again today it would be much shorter: give the seller back the hours, and keep a person on every decision that touches a customer.
Sources
- Salesforce, "State of Sales" 2026 announcement, 3 February 2026, survey of 4,050 sales professionals in 22 countries, August and September 2025: salesforce.com, State of Sales 2026.
- Gartner, "Gartner Predicts by 2028 AI Agents Will Outnumber Sellers by 10x, Yet Fewer Than 40 Percent of Sellers Will Report AI Agents Improved Productivity", 18 November 2025, repeated in a 28 July 2026 release: gartner.com newsroom. Also reported by Insurance Canada (December 2025) and destinationCRM (September 2026).
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025: gartner.com newsroom.
- McKinsey, "The state of AI in 2025: Agents, innovation, and transformation", November 2025: mckinsey.com. Also reported by IT Brief, 11 November 2025.
- MongoDB and AWS, "MongoDB and AWS Collaborate to Optimize Amazon CodeWhisperer Suggestions for Developers Building Applications on MongoDB", 6 November 2023: mongodb.com press release.
- Philip Siniscalchi, "Transforming Sales: How AI Will Revolutionize the Salesperson's Role", LinkedIn, 28 May 2024. The original, whose statistics this version removes.