The History of AI in Sales and Revenue Operations

By Victor Hoang · Updated Oct 24, 2026

In December 1971 a program called CALLPLAN was published. It asked a salesperson to estimate what each extra call on each account would be worth. Then it told them where to spend their time.

Machine intelligence entered selling in the 1940s and 1950s as statistics: credit scorecards, sales-response models and call-planning programs. The first rule-based AI to touch a sales order, John McDermott's R1 at Digital Equipment Corporation, went into use in January 1980. Predictive scoring, forecasting and call analysis came inside CRM platforms from 2016. Generative drafting arrived in 2023, and software that takes sales tasks on its own was announced in 2024. The throughline: each wave got its first real use on the part of selling that already left data behind. Each wave was only as good as the records it was fed.

CALLPLAN is an early, documented case of software advising a seller on judgment calls. The history below is the long argument about how much of that judgment a machine should supply.

Two terms need care. Precursors are earlier practices that pursued the same goal with different tools; the 1940s credit models and 1971 call planner are precursors, not "AI in sales" as the field now uses the phrase. The concept itself, learning or reasoning software embedded in selling and revenue workflows, takes shape with R1 and its planned sales companion in 1980, and becomes a product category with the CRM vendors of 2016. For the wider field see The History of AI. For the sales systems these tools sit inside, see the history of CRM and the history of revenue operations.

Timeline at a Glance

The Scorecard Years (1941-1977)

Statisticians learn to rank customers and plan sales calls

The scene is a 1940s consumer-finance office with a stack of loan files. David Durand's 1941 National Bureau of Economic Research study, Risk Elements in Consumer Instalment Financing, covers credit-rating formulae and applies discriminant functions to samples of good and bad loans. The idea was simple and lasting: past customers whose outcomes are known can tell you something about the next applicant. Fifteen years later, according to FICO's own history page, the company was founded in 1956, and its first scoring system followed in 1958. Reference for Business names the founders as William R. Fair and Earl Isaac, and says that first system, Credit Application Scoring Algorithms, predicted whether credit holders would pay on time, pay late or not pay at all.

Selling got its own models in the same decades. In June 1957 M. L. Vidale and H. B. Wolfe published an operations-research study of sales response to advertising. Their model had three parameters (a sales decay constant, a saturation level and a response constant), and the authors said it had been useful for deciding how to divide advertising budgets. Then, in December 1971, Leonard Lodish published CALLPLAN in Management Science. It was an interactive system that took the salesperson's own estimates of the contribution from each call policy for each account, added travel time and call frequency limits, and solved for the allocation that maximized contribution. Fourteen salesmen used it in six sales situations, and the judgmental inputs implied sales gains of roughly 5% to 25%. Those are projections from the salesmen's own estimates, not measured results.

A measured result came later. Fudge and Lodish reported a 1975 United Airlines experiment in Interfaces: ten salespeople in New York and San Francisco used CALLPLAN, ten matched colleagues did not, and after six months the average CALLPLAN user had 8.1% higher sales than the matched counterpart. The sample is small, and the paper's abstract notes that participants began skeptical and came to see that they controlled the model.

These are precursors. Nobody called them AI, and none learned from new data on its own. What they solved was the question of where effort pays. What they left open was the data: every one depended on someone entering good estimates. And note the design: the human stayed in charge of the inputs, which is the same arrangement the newest tools still negotiate.

For teams today: the oldest lesson is to test a planning aid against a matched control, as Fudge and Lodish did. A tool that "feels" useful and a tool that raises sales are different claims.

The XCON Years (1978-1980s)

Rules configure the order while a sales assistant stays on the drawing board

The scene is Digital Equipment Corporation's order desk around 1980. A customer orders a VAX 11 computer from a list of parts, and someone must check that the parts fit and add the cables, backplanes and boxes that were forgotten. According to McDermott and Barbara Steele's 1981 IJCAI paper, work on R1, a rule-based program that configures VAX 11 systems, began in December 1978. Since January 1980 DEC's manufacturing organization had used it to configure almost all VAX 11 systems shipped. It was written in the OPS5 rule language, held about 850 rules, and DEC had a group of 15 people maintaining it. McDermott's AI Magazine account describes it as producing diagrams of the roughly 90 components that typically make up a system.

The same paper describes the sales side. XSEL, then under development at Carnegie Mellon, was described as "a computer salesperson's assistant": it would interact with a salesperson or customer, select a CPU, memory, software and devices, and pass that skeletal order to R1 to be completed. That is among the earliest documented plans for software that sits in the middle of a sales conversation. R1 itself belongs to the wider rise and fall covered in the history of expert systems, and the learning methods that later replaced hand-written rules are in the history of machine learning.

The 1981 paper also shows the catch. Rules had to be written and kept current by people. R1 could not accept customer-specific constraints until its authors added rules for them. That cost, not the idea, is what limited this wave. Rules fit best where the constraints are crisp, as they are when components must fit together.

For teams today: if a process needs a full-time group just to keep its rules true, count that group in the cost. The same warning applies to every routing and scoring rule in your CRM.

The Response Model Years (1990s-2014)

Mailing lists and customer records become inputs for prediction

Direct mail posed a sales problem with a clean answer: a list of households, and a budget that can't reach all of them. In 1995 Jan Roelf Bult and Tom Wansbeek published Optimal Selection for Direct Mail in Marketing Science. It proposed picking whom to mail by comparing marginal costs with marginal returns, so that selection maximizes expected profit, and it reported returns higher than traditional selection approaches. Choosing the best prospects by predicted response, instead of mailing everyone, is the logic that lead scoring still uses.

The raw material accumulated in the same years. Sales force automation grew into enterprise CRM, and browser-based CRM spread it to smaller firms; that story, with the dates and filings, is in the history of CRM. The function that owned the resulting funnel data is traced in the history of revenue operations. What this period solved was volume: for the first time a company had its own history of leads, deals, wins and losses in one place. What it left unsolved was use: having the records and turning them into predictions are separate jobs.

For teams today: a model needs labeled outcomes. If your CRM cannot say reliably which leads became customers, no vendor can predict the next one for you. Start with the records, as the guide to AI in revenue operations argues.

The Einstein Layer (2015-2021)

Prediction moves inside the CRM, and calls become data

The turning point is a platform vendor adding machine learning to the product every rep already opens. On September 19, 2016 Salesforce introduced Einstein across its clouds. For Sales Cloud it listed Predictive Lead Scoring, Opportunity Insights and Automated Activity Capture. Marc Benioff said "Einstein is now every customer's data scientist." That is the vendor's description of its own launch, not independent evidence that the features worked. A year later, on September 19, 2017, it announced Einstein Forecasting, Opportunity Scoring and Email Insights in pilot. The release says self-learning algorithms use a company's historical CRM data to produce forecasts, and it states that "less than half of deals close as forecasted," a figure it does not source.

A second branch grew from a different data source: the sales call itself. According to a First Round Review account, Gong's founders Amit Bendov and Eilon Reshef began in 2015 after Bendov's earlier team had used recorded sales calls to work out why a quarter went badly. They raised pre-seed funding in October 2015 and finished a beta in January 2016 with 12 design partners. The first prototype was deliberately plain: screen recording, third-party transcription, comments and basic talk-ratio analysis. The article presents the founders' own recollection, so treat the specifics as their account. The idea behind it, that CRM fields miss why deals stall and the conversation holds the answer, is the basis of what is now called conversation intelligence.

Independent evidence arrived in 2020. A Journal of Marketing study by Xueming Luo, Shaojun Qin, Zheng Fang and Zhe Qu ran randomized field experiments at two fintech firms. An AI coach that analyzed agents' sales conversations beat human coaching for middle-ranked agents, while bottom-ranked agents were overwhelmed by feedback and top-ranked agents showed the strongest aversion to AI. A combined AI and human coach did better than either alone. The first experiment had 429 agents, and the agents worked at fintech firms, so the result does not automatically carry to enterprise field sellers.

What the era solved: scoring, forecasting and coaching moved from consultants' projects into default features. What it created: a new dependence on the quality of CRM data, since the forecasts and scores were learned from it.

For teams today: ask for the study, not the slogan. The one peer-reviewed result here says AI coaching works unevenly and best alongside a human manager.

The Generative Turn (2022-2023)

Software starts to write the email, the summary and the next step

On November 30, 2022 OpenAI released ChatGPT as a free preview, and generative text reached ordinary workers; the background is in the history of generative AI and the history of large language models. Salesforce reacted within three months. Its March 7, 2023 release announced Einstein GPT, headlined by the company as "the World's First Generative AI for CRM." That claim is Salesforce's own. The release says Einstein GPT for Sales would automate routine tasks including drafting emails, scheduling meetings and preparing reps for their next customer interaction. The same page now notes that Einstein GPT evolved into Agentforce.

The move to drafting changed what sales AI could be wrong about. Earlier tools ranked or predicted, and a bad score was a wasted call. A drafting tool writes in the company's name, and a bad draft goes to a buyer. The wider pattern of assistants embedded in work tools, and the way one brand name spread across vendors, is told in the history of AI copilots.

For teams today: outbound text is a brand decision. Set who reviews AI-drafted messages before they send, and keep that rule even when volume rises.

The Agent Age (2024-present)

From assisting the rep to attempting the task

The turning point is the shift from suggestion to action. On September 12, 2024, per Salesforce's Dreamforce recap, the company announced Agentforce, including a Sales Development Representative agent it described as engaging prospects around the clock, answering questions, handling objections and scheduling meetings from CRM and external data. On October 29, 2024 it announced general availability. The release describes agents that reason, make decisions and complete tasks, and says Agentforce can qualify sales leads. All of this is the vendor's description of what its product is designed to do.

Compare this with R1 and XSEL in 1981. A program once planned to take a customer's needs and hand a skeletal order to a configuration engine now has a descendant that is said to open the conversation. The same concern runs through both: the quality of the underlying records and rules decides the outcome. The history of agents as a field is in the history of AI agents, the definition is in what an AI agent is, and a practical view of the role is in what an AI sales operator is.

What this wave has not yet produced is independent, peer-reviewed evidence that agents raise revenue in B2B selling. The evidence in this history is of three kinds: models tested on small samples (CALLPLAN), vendor announcements (Einstein, Agentforce) and one randomized coaching study. Treat agent claims accordingly. The rules around them are newer still: see the history of AI governance for the principles and frameworks companies adopted, and the history of AI regulation for the laws on automated decisions.

For teams today: before an agent talks to a buyer, define what it may promise, when it hands off to a person, and who owns the data it reads.

Key People

Key Facts

• Lodish's CALLPLAN appeared in December 1971; a 1975 field test with 10 users and 10 matched controls found 8.1% higher sales for users after six months.
• R1 began in December 1978 and has been used since January 1980 to configure almost all VAX 11 systems DEC shipped, with about 850 rules and a 15-person maintenance group.
• Bult and Wansbeek (1995) proposed choosing mail recipients by marginal cost against marginal return.
• Salesforce introduced Einstein with Predictive Lead Scoring on September 19, 2016 and Einstein Forecasting in pilot on September 19, 2017.
• A randomized study with 429 agents in its first experiment found AI coaching helped middle-ranked agents most, and AI plus human coaching beat either alone.
• Salesforce announced Agentforce general availability on October 29, 2024.

The Pattern Behind the History

Across eight decades the same sequence repeats. We call it the Record-First Pattern. It is our synthesis of the history above, not a research finding.

1. The data exists before the tool. Loan files, call logs, CRM histories and recorded calls each sat unused until someone modeled them.
2. The first win is narrow. Credit scores, call-time allocation and VAX configuration each solved one bounded problem.
3. The human stays in the loop. CALLPLAN needed salespeople's estimates, R1 needed people to maintain its rules, and the coaching study found a human manager still helped.
4. Distribution beats invention. Each wave reached most sellers only when a platform they already used shipped it by default.
5. Claims outrun evidence. Vendor announcements arrived years before independent tests, if any came.

What It Teaches Teams Today

• Fix the record first. Every wave depended on the quality of the data behind it, from loan files to CRM fields.
• Ask who was measured. The best sales-AI evidence is a handful of small experiments, so check them against your own team.
• Budget the upkeep. R1's 15-person maintenance group is the early example of a cost every model and rule set still carries.
• Design the handoff. Decide what software may decide, what it may send, and when a person takes over.
• Separate the label from the capability. "AI-powered" has described a regression, an LLM draft and an agent in the same decade.

The question under every tool here is the one the 1971 sales manager faced: which parts of the job should a machine take, and which should a person keep? Ask it again of each new product, and ask for the evidence before you answer.