Writing

• AI in Marketing
• 1 October 2026

AI in marketing and sales: where the handoff gets fixed, and where not

AI in marketing and sales is most useful at the handoff between the two teams: scoring leads against your own conversion history, routing them to the right person in minutes rather than days, and passing the rep a summary of what the buyer has already done. It does not fix disagreements about what a good lead is, messy CRM data, or the fact that buying decisions are made by groups rather than individuals.

Adoption is no longer the question. Salesforce’s State of Sales 2026 (4,050 sales professionals in 22 countries, including the UK, surveyed August to September 2025) found 87% of sales organisations use some form of AI and 54% of sellers have used AI agents. Gartner, meanwhile, predicts AI agents will outnumber sellers 10 to 1 by 2028, while fewer than 40% of sellers will say those agents improved their productivity. That gap is where the handoff lives.

How is AI used in marketing and sales right now?

On the marketing side, Salesforce’s State of Marketing 2026 (4,450 marketing decision makers, surveyed October to November 2025) puts AI adoption at 75%, and then in the same breath reports that 84% of marketers admit to running generic campaigns. So we have bought the most sophisticated personalisation technology in history and are using it to send the same email to everyone slightly faster. Well done us.

On the sales side, Salesforce’s respondents expected AI to cut prospect research time by 34% and content creation time by 36%, and 55% already use it for prospecting. Gartner’s survey of 227 chief sales officers found that organisations giving sellers AI-enabled next best actions were 2.6 times more likely to achieve commercial growth.

Between those two sits the old relay, where marketing generates interest, qualifies it somehow and throws it over the wall to sales. (For the wider picture of tools and categories, see what AI in marketing actually covers.)

Where does AI finally fix the marketing-to-sales handoff?

Lead scoring that learns from your own conversions

Traditional lead scoring is a points system someone built in a workshop years ago: plus ten for downloading a whitepaper, plus fifteen if the job title contains “Director”, minus twenty for visiting the careers page because they probably want a job rather than your software. Nobody has revisited the weights since, and the person who set them has left.

Predictive scoring flips the direction. Instead of marketing guessing which behaviours matter, the model looks at leads that did and didn’t convert and works out the weights itself. All three major CRMs now do this, and they retrain on a schedule: Microsoft’s Dynamics 365 Sales can retrain automatically every 15 days, and Salesforce says Einstein reanalyses your leads approximately every 10 days. That removes one old argument, because the score is tied to deals sales actually closed rather than to what marketing hoped would happen.

Routing and first response

The second fix is speed. A lead that fills in a form at 6pm on a Friday used to sit in a queue until someone on Monday decided whose territory it was. AI agents can now reply, qualify and book the meeting in the gap. Salesforce’s own internal deployment, reported in the State of Sales announcement, had agents contact 130,000 leads and create 3,200 opportunities in four months, which works out at roughly 2.5% of leads contacted becoming an opportunity. That is a vendor reporting on its own product, so treat it as proof of scale rather than a benchmark.

The routing mechanics are more prosaic than the marketing suggests. Salesforce’s documentation for its Agentforce Engagement agent (the one that nurtures leads) says it works only over email, randomly assigns meetings among available members when a group owns them, and needs the Owner field filled in before a record can be assigned to it. The agent is fast, but it still routes by ownership rules a human set up.

Context that travels with the lead

A rep picking up a lead used to get a name, an email address and a score of 67. Now the CRM can summarise which pages they read, what they asked the chatbot and which webinar they left after nine minutes. HubSpot’s prospecting agent, for example, draws on up to a year of engagement (form submissions, page views, calls, meetings, notes, email opens) plus company news such as funding rounds or product launches, and can run in a “review before sending” mode where a person approves each email. That saves everyone the painful “so, tell me about your business” call.

How does AI lead scoring and routing actually work?

Predictive lead scoring is a classification model: it takes historical leads labelled converted or not converted, finds the attributes and behaviours that separate them, and gives each new lead a probability or grade. It needs enough history to learn from, and the thresholds vary a lot by platform.

Sources for the table: Salesforce Help, Microsoft Learn, HubSpot Knowledge Base.

A worked example

Take an illustrative UK B2B software firm that creates 150 inbound leads a month and converts about 6% of them. Over 200 days that is roughly 1,000 leads and 60 conversions. It clears Salesforce’s lead volume bar but not the 120-conversion bar, so Einstein will score its leads using the global model, which is trained on other companies’ buyers. It clears Dynamics comfortably on a six-month window, and it clears HubSpot’s minimum sample easily.

So the same firm could get a model trained on its own buyers in one CRM and a borrowed one in another, and the sales team would not know which unless someone told them. Before anyone trusts a score, I’d check which model is live and how many conversions it has actually seen.

Routing is a set of rules with a model in front

In most setups the score decides whether a lead goes to a person, an agent or a nurture track, ownership rules (territory, account owner, round robin) decide which person, and an agent handles first contact out of hours. The AI is the triage at the front, so if your territory map is wrong, the agent will send the wrong rep to the right lead very quickly.

Where doesn’t AI fix the handoff?

It can’t agree what a qualified lead is

A predictive model learns from whatever you labelled as “converted”. If marketing counts a booked demo as a conversion and sales counts a signed contract, you have two definitions of success and the model will faithfully optimise for whichever field someone ticked in setup. The argument about MQLs that has run through every quarterly business review since the invention of the MQL does not go away. It just moves into a configuration screen, where fewer people can see it.

It scores people, and people don’t buy alone

Forrester’s 2026 buyer research (published January 2026) found a typical buying decision involves 13 internal stakeholders and nine external influencers. The models in the table score individual leads or contacts, so they can tell you a procurement manager is highly engaged while the finance director who will actually veto the deal has never opened an email. Forrester moved its own demand model from individual leads to buying groups back in 2021; most CRM scoring is still catching up.

It inherits your data problems

In Salesforce’s sales survey, 46% of sales professionals using agents said data quality issues hurt their sales, and high performers were far more likely to prioritise data hygiene (79% versus 54% of underperformers). The marketing survey shows the other half of the problem: only 56% of marketers have complete access to sales data, so the model on the marketing side is scoring leads without knowing what happened to the last lot. We then blame the model, the agency, the attribution software, probably the moon.

It speeds up the process you already have

Masha Shunko (University of Washington Foster School of Business) and Serguei Netessine (Wharton) make this argument in HBR, in “Stop Automating Old Processes. Design New Ones Instead.” (14 September 2026): companies automate individual tasks rather than redesign the workflows that produce value, and they need to design for the “unhappy path”, meaning how a system recognises uncertainty and escalates exceptions. In handoff terms, that is the existing customer who fills in a new-business form, or the student writing a dissertation who looks like a Head of Procurement on paper. If nobody designed what happens to those, the agent will handle them with total confidence.

It doesn’t prove its own value

Gartner’s CSO survey found 31% of chief sales officers cited difficulty proving the ROI of AI tools as a top challenge for 2026. Sandhya Mahadevan, Senior Director Analyst at Gartner, said many leaders still try to prove AI ROI “as if value should appear quickly, clearly, and in familiar financial terms.” For lead scoring, the fair metric is conversion rate by score band over two or three quarters, compared with what you had before. There is more on that in measuring the ROI of AI in marketing.

What does AI in marketing, sales and service look like together?

The handoff most businesses forget runs from sales to service and back again. Salesforce’s State of Service report (6,500 service professionals in 39 countries, surveyed April to June 2025) found AI handles 30% of service cases today, with service teams expecting 50% by 2027. They also projected agentic AI would lift upsell revenue by 15%, which makes service a source of sales leads in its own right.

Marketing mostly can’t see any of this. Only 58% of marketers in the Salesforce marketing survey had complete access to service data. So a customer can spend Tuesday on a support chat about a billing problem and receive a cheerful upgrade email on Wednesday, because the email tool had no idea Tuesday happened. Teams with unified customer data were 42% more likely to respond to customers regularly, according to the same survey.

The design question is which signals flow in which direction. I’d start with these:

• Open service cases suppress marketing promotions to that account until the case closes.
• Product usage or service conversations that suggest expansion create a lead that goes straight to the account owner.
• Closed-lost reasons from sales feed back into the scoring model’s training labels and into marketing’s segment definitions (the AI customer segmentation guide covers how to use them).

None of these need an agent, just three teams agreeing the rules and a CRM holding all three sets of data, which is less exciting to announce on LinkedIn but does more.

How do you set up AI lead scoring and routing without automating a bad process?

Shunko and Netessine suggest choosing, for each step, whether humans assist, approve, audit or leave it to automation, according to the stakes and how reversible a mistake is. That fits the handoff well, and this is the order I’d follow.

1. Agree one definition of conversion, in writing, signed off by the heads of marketing and sales. Pick the outcome closest to revenue that you have enough volume of (an accepted opportunity is usually a better label than a form fill).
2. Check your data against the platform thresholds in the table above. Count conversions in the training window, not leads. If you fall short, you are using a global model or rule-based scoring, so tell the sales team.
3. Clean the fields the model will use: job title, company size, industry, lead source, owner. Duplicates and blank owner fields break routing before they break scoring.
4. Run the model alongside your existing score for a quarter and compare conversion rates by band. Don’t switch the old one off until the new one wins.
5. Design the unhappy paths: existing customers, wrong territory, competitors, students, partners, anyone the model scores high but a human would disqualify in five seconds.
6. Set the human mode for each step. I’d put first-response emails from an agent in “approve” (review mode) for the first month, routing in “audit” (someone reviews a weekly sample), and score calculation in “automate”.
7. Close the loop: every closed-lost reason and every rep rejection of a lead goes back into the data, and someone owns reviewing it monthly.

For step 5, a general-purpose assistant such as ChatGPT or Claude is useful for drafting the exception list before you configure anything. A prompt along these lines works:

You are helping a UK B2B marketing and sales team design lead routing rules.
Our ideal customer: [describe company size, sector, buyer roles].
Our current routing: [territories, round robin, account ownership rules].
List 15 types of inbound lead that a lead scoring model might score highly
but that should NOT go to a new-business sales rep (for example existing
customers, partners, job seekers). For each, say how we could detect it from
CRM fields or form answers, and where it should be routed instead.

Then show the output to a sales manager, who will add three you missed, at least one of them a person who has filled in the demo form eleven times.

There is a fuller method in how to build an AI marketing workflow, and the difference between assistants and agents is covered in generative vs agentic AI in marketing.

Is AI lead scoring allowed under UK data protection law?

For most B2B lead scoring, yes, though scoring people on their behaviour is profiling, so it belongs in your privacy notice and lawful basis assessment. The Data (Use and Access) Act 2025 changed the rules on automated decision-making, and the ICO ran a consultation on draft guidance about automated decision-making and profiling from 31 March to 29 May 2026. The Data & Marketing Association reads the draft as one that “facilitates marketing automation powered by AI”, and notes it distinguishes fully automated decisions from those with meaningful human input. If your scoring decides who gets a price, credit terms or access to a service, rather than just who gets a call first, take proper legal advice and check the final ICO guidance.

What I’d watch for next

Buyers are changing faster than the handoff. Gartner’s survey of 645 B2B buyers found 67% prefer a rep-free buying experience and 45% had used generative AI, mainly to gather information on vendors and products, yet 69% want to validate what AI told them with a sales rep. So the lead that arrives at your form has often already done its research with an AI assistant, and wants a human to confirm or contradict it. That changes what the rep’s first call is for, and I doubt many scoring models are trained to notice it.

My own bias, from leading AI adoption in marketing at ACTAsia, is to agree the process before choosing the tool. So the question I’d take to your next marketing and sales meeting is this: if you had to write down, in one sentence both teams would sign, what counts as a converted lead, could you?

FAQs

What is AI lead scoring?

AI lead scoring, also called predictive lead scoring, uses a machine learning model trained on your past leads to estimate how likely each new lead is to convert. Instead of marketing assigning points manually, the model finds which attributes and behaviours separated converted from unconverted leads. Salesforce, Microsoft Dynamics 365 and HubSpot all offer it, each with a minimum amount of historical data before it will train on yours.

How much data do you need for predictive lead scoring?

It depends on the platform. Salesforce Einstein needs at least 1,000 leads created in the last 200 days with at least 120 converted, otherwise it uses a global model. Dynamics 365 needs at least 40 qualified and 40 disqualified leads closed in the training window. HubSpot’s AI scores need a sample of at least 50 contacts, 25 converted and 25 not, on Marketing Hub Enterprise.

Can AI agents replace SDRs for inbound leads?

They can handle first contact, follow-up and meeting booking at volume and out of hours, and Salesforce reports its own agents contacted 130,000 leads and created 3,200 opportunities in four months. But they route according to ownership rules people set, and Gartner predicts fewer than 40% of sellers will say agents improved their productivity by 2028. Most teams will use agents for first response and keep humans on qualification calls.

Why do marketing and sales still disagree about lead quality with AI?

Because the model learns from whatever outcome you label as a conversion. If marketing counts a booked demo and sales counts a signed contract, the model optimises for one definition and the other team will distrust it. Poor CRM data makes it worse: 46% of sales professionals using agents in Salesforce’s 2026 survey said data quality issues hurt their sales.

How does AI connect marketing, sales and service?

By sharing signals across the three teams in one CRM. Open service cases can pause marketing promotions, expansion signals from service can create leads for account owners, and closed-lost reasons can retrain scoring models. Salesforce’s State of Marketing 2026 found only 58% of marketers had complete access to service data, so this connection is still missing in many organisations.

Sources

1. Salesforce, “Salesforce Announces State of Sales Report for 2026”, 3 February 2026
2. Salesforce, “State of Sales Report” (2026 edition, PDF), February 2026
3. Salesforce, “75% of Marketers Have Adopted AI Yet Still Use It To Send One-Way, Generic Campaigns”, 19 February 2026
4. Salesforce, “Salesforce 2025 State of Service Report”, 13 November 2025
5. Gartner, “Gartner Predicts AI Agents Will Outnumber Sellers 10 to 1 by 2028, Yet Fewer Than 40% of Sellers Will Say Agents Improved Productivity”, 28 July 2026
6. Gartner, “Gartner Survey Finds Sales Organizations That Provide AI-Enabled Next Best Actions Are 2.6x More Likely to Achieve Commercial Growth”, 20 May 2026
7. Gartner, “Gartner Survey Shows 31% of Chief Sales Officers Cited Difficulty Proving ROI of AI-driven Tools as a Top Challenge for Sales Objectives in 2026”, 19 May 2026
8. Gartner, “Gartner Survey Finds 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights”, 20 May 2026
9. Forrester, “Forrester’s 2026 Buyer Insights: GenAI Is Upending B2B Buying As Leaders Face Mounting Pressure To Justify Every Dollar Spent”, 21 January 2026
10. Forrester, “Forrester Debuts Next-Generation B2B Revenue Waterfall To Help Firms Accelerate Revenue Growth”, 4 May 2021
11. Harvard Business Review, Masha Shunko and Serguei Netessine, “Stop Automating Old Processes. Design New Ones Instead.”, 14 September 2026
12. Salesforce Help, “Considerations for Setting Up Einstein Lead Scoring”, accessed October 2026
13. Salesforce Help, “Considerations for Using Agentforce Engagement”, accessed October 2026
14. Microsoft Learn, “Configure predictive lead scoring”, accessed October 2026
15. HubSpot Knowledge Base, “Overview of the lead scoring tool”, accessed October 2026
16. HubSpot Knowledge Base, “Set up and use the prospecting agent”, accessed October 2026
17. ICO, “ICO consultation on the draft guidance about automated decision-making, including profiling”, 31 March 2026
18. Data & Marketing Association, “ICO Seeks Views on Draft Automated Decision-Making Guidance”, April 2026