AI's Impact on Marketing
Where AI is landing in marketing and commercial — and why the pilots that stick tend to begin by removing a pain the user already hates, not by mandating a tool. Governance that locks tools down breeds shadow AI; the real frontier is becoming the brand AI engines recommend.
In the room
- Host
- Jesse Hopps
- Room
- Moderated by John Follett. Participants from AstraZeneca, Chemours, Constantia Flexibles, CWS Cleanrooms, Electrolux, Johnson & Johnson, Medtronic, Quaker Houghton, Sabic, Stäubli Group, TD Synnex, TVH Parts, 79 Dev, and Demand Metric
Nine AI deployments running in the field, and what separated the ones that stuck
An open exchange on where AI is landing in commercial and marketing teams — the real blockers, the use cases delivering today, and what separates a pilot that scales from one that stalls.
Fourteen companies, and a session that favored lived experience over theory. It moved from the blockers the room shares into a working tour of what is live in the field, and the adoption tactics behind each one.
The short version
- Adoption follows value, not mandates. The use cases that stuck removed work people already hated.
- Locking AI down created the risk it was meant to prevent — staff routed sensitive work through personal accounts instead.
- AI accelerates whatever process it touches, including a broken one.
- Human-in-the-loop is how trust got built: SME validation, editable output, explicit go/no-go gates.
- The next battleground is being recommended by the machine, and few teams own it yet.
The blocker is governance — and the usual response makes it worse
The dominant blocker, named across the industries in the room, was fear of proprietary IP leaking into the public domain. The common corporate response — hand out a restricted free Copilot, no training, no conversation about what it is for — drew the sharpest line of the session: that is “pretending you’re doing AI.”
- The archetype. A specialty-chemicals marketing leader described a straightforward request for a professional corporate video that sat undone for a year, because the team leaned on policy and an outside agency instead of being equipped with — or trusted to choose — the right tools.
- The paradox. Withholding proper tools tends to drive IP risk underground rather than reduce it. A health-tech leader confirmed staff were already pasting sensitive information into personal AI accounts, precisely because sanctioned tools were absent.
Four problems that show up at scale
- 1
Fix the process first
A commercial-excellence leader at a manufacturing group put it bluntly: integrate AI into a dysfunctional process and it fails. The blunt version from the room put the failure rate at roughly 95% when AI is bolted onto a broken process.
- 2
Define what “AI fluent” means
A leader at a global IT distributor described a strong top-down fluency mandate that, left undefined, risked twenty business units each building their own quoting system with no shared pricing guardrails. Their fix: separate personal productivity — low-code, Copilot-style tooling — from enterprise transformation, which needs far heavier guardrails.
- 3
Watch the token bill
The same leader flagged token and OPEX sprawl, including an employee-built Outlook bot running unnoticed at about €400 a week to do something a free rule could handle. Early savings can invert as usage scales.
- 4
Homegrown tools come with key-person risk
Several teams are shipping impressive agents built on Claude and internal ChatGPT instances. If the person who wrote and tuned the two-page prompts leaves, the tool breaks and the team has little way to repair it — so the open question is whether to build your own as a stepping stone or buy enterprise-grade at higher cost.
Nine use cases running in the field
The session runs under the Chatham House Rule, so each deployment below is described by role and industry rather than by name.
In front of the customer
- 1
Conversational visit-report agent — industrial spare-parts distributor
A voice agent preps reps before a call with open issues and history, then interviews them afterward and writes the report straight into the correct CRM fields. In a thirty-person pilot there were no refusals, a waiting list formed to join, report quality rose sharply, and the lag between visit and log dropped dramatically. It scales to the full field force in September.
- 2
Account-plan agent on CRM history — same distributor
An agent mines years of historical visit reports against a good-account-plan template and drafts roughly 80% of the plan, alongside a model that estimates a customer’s future potential from public information. A task reps disliked became a quick review-and-confirm.
- 3
Image comparison for troubleshooting — specialty lubricants
Customers photograph an imperfect finished metal product, and the company compares it against a decades-deep image database to diagnose the issue and tune the lubricant program. Paired with predictive mining of usage data, it lets them advise customers to use less chemistry and less energy for a better result. Notably, there was no formal CSAT measurement before or after.
- 4
Negotiation role-play — commodity and specialty chemicals
A coaching-tool pilot covering call prep, negotiation role-play, opportunity progression, and account-plan creation. The role-play tore up an experienced negotiator on the first pass — realistic enough to be genuine rehearsal before a live account.
- 5
Editable next-best-action — global pharma
After reps ignored a flood of accept-or-reject suggestions, the team added a modify option: reps keep the recommended customer but change the topic, the visit books itself into their timeline, and related content surfaces instantly. The modification delivered value and fed the internal model at the same time — and that usage is what built rep trust.
Behind the scenes
- 1
Forecasting and budget planning — flexible-packaging manufacturer
Early in the journey, and focused on internal use cases that make life easier: taking non-selling work off account managers’ desks to free up selling time, with an aspiration of a phone app that walks a manager through planning conversationally.
- 2
Competitive-intelligence agent — chemicals company
An agent built on Claude combs decades of SharePoint files plus competitor pages and earnings releases into a standard leadership-ready report. Two further prompts mine the last thirty days of global call reports for trends by product line and region, and surface growth opportunities from clustered inbound leads — the see-the-forest-through-the-trees problem the marketing-automation platform alone left unsolved.
- 3
Market-entry intelligence — engineering and construction
Custom skills and GPTs research new markets — data centers and adjacencies beyond the chemical and refining base — pulling not only 10-Ks but government land, environmental, and permitting filings to flag competitor expansion moves. Everything is stage-gated through SME review meetings and a formal go/no-go.
- 4
Agents on top of dashboards — global medical devices
Facing 300+ systems inherited through acquisition, the team is deploying an agent layer over existing dashboards to surface leading and lagging indicators, plus agents that talk to other agents — streamlining rather than ripping out the systems underneath. Leadership is responding to the build-on-top framing.
What made adoption stick
- Start from the business outcome, not the tool. The pilots that landed began with a defined objective and a specific user pain.
- Give the user the win first. Reps queue up for a tool that erases work they hate, and ignore one that adds a step.
- Put a small squad next to the data-science team. One manufacturer stood up an AI-savvy spearhead group tightly coupled to data science, so experimenters knew which tools and data were fair game. Formal training followed the pilots rather than preceding them.
- Keep a human in the loop. SME validation sessions, engineer-led go/no-go gates, and editable output convert skeptics — and when the edits feed the model, they improve it over time.
- Frame it as a bounded project. Contained, customer-serving use cases drew more pull than blanket tool provisioning, though leaders flagged the reverse risk: subject-matter experts resisting digitization out of fear for their jobs.
The emerging frontier: being recommended by the machine
A growth consultant reframed the stakes as high ground versus low ground. Most AI effort today sits on the low ground — CRM, research, content production. The high ground is being the brand an AI model recommends when a buyer, or a procurement team weighing a multi-million-dollar tender, asks who is best in the category.
- Run it in parallel with the AI work already underway, rather than after it.
- Engineer content for the machine layer — clean HTML, indexable, structured.
- Build authority signals through co-published research and credible media.
- Expect a six-to-eighteen-month horizon before the brand is the visible answer.
Practices to apply immediately
- Lead with the user's pain, not the tool. The pilots that stuck removed work people already hated — writing visit reports, building forecasts, combing competitor filings.
- Provide sanctioned tools and the training to use them. A locked-down free tool with no guidance pushes sensitive work into personal accounts, which puts more IP at risk rather than less.
- Draw the line between forward-facing and proprietary up front. That single distinction decides whether a project runs fast or crawls.
- Fix the process before you automate it. AI accelerates a broken process as readily as a good one.
- Separate personal productivity from enterprise transformation. The first needs low-code tooling; the second needs guardrails and a cost model.
- Keep a human in the loop. SME validation, editable output, and explicit go/no-go gates are what turn skeptics into adopters.
- Start answer-engine optimization now. Being the brand an AI model recommends is a six-to-eighteen-month build, and few teams own it yet.
Questions the room worked through
- How do you get past the fear of proprietary IP leaking into a public model?
- Draw the line up front between what is forward-facing and what is proprietary. Several leaders said that single distinction is what decides whether a project runs fast or crawls. The alternative — a restricted free tool handed out with no training — pushes people to use their own accounts, which puts more IP at risk rather than less.
- Who should run the early AI experiments?
- One manufacturer stood up a small spearhead group of AI-savvy people positioned close to the data-science team, so experimenters knew which tools were sanctioned and which data was fair game. Formal training for managers followed about three months later, once there were real use cases to train on.
- What does an “AI fluent” mandate get wrong?
- Left undefined, it invites twenty business units to build twenty versions of the same thing — twenty quoting systems with no shared pricing guardrails. The fix the room landed on was to separate personal productivity, which suits low-code Copilot-style tooling, from enterprise transformation, which needs guardrails, shared logic, and a scalability plan.
- Why do AI projects fail?
- The room's answer was process, not technology. Integrate AI into a dysfunctional process and it accelerates the dysfunction. One leader put the failure rate at roughly 95% when AI is bolted onto a broken process.
- How do you get sales reps to adopt an AI tool?
- Give them the win first. A visit-report agent that interviews reps after a call and writes the report into CRM drew no refusals in a thirty-person pilot, and a waiting list formed to join it. By contrast, a next-best-action engine offering only accept-or-reject was ignored until a modify option was added — the edit delivered value to the rep and fed the model at the same time.
- What does AI cost once it scales?
- More than the pilot suggests. One leader described an employee-built Outlook bot running at about €400 a week to do something a free mail rule could handle. Token consumption is an OPEX line, and it grows with adoption rather than shrinking.
- What is the risk with homegrown agents?
- Key-person risk. Several teams have capable agents built on two-page prompts refined by one person. If that person leaves, the tool breaks and the team has little way to repair it. The open question is whether to build your own as a stepping stone or buy enterprise-grade at higher cost.
- What is answer-engine optimization, and why should commercial excellence care?
- Most AI effort today sits on the low ground — CRM, customer research, content production. The high ground is being the brand a model recommends when a buyer, or a procurement team weighing a multi-million-dollar tender, asks who is best in the category. The play is to run it in parallel with internal AI work: engineer content for the machine layer so it is clean and indexable, and build authority through co-published research and credible media. Only a handful of the companies in the room had a team on it.
Want to be in the next session?
Consortium membership is invite-only and reviewed weekly. Directors and VPs of Commercial Excellence, Strategy, and Transformation at $1B+ enterprises.
No consultants, SaaS vendors, or service providers.
Request to Join