Agentic AI Workflows
Thirty senior executives, zero votes at five-out-of-five on actual agentic deployment. Why the bottleneck isn't the technology — it's the coalition, the governance, and the maturity stages most enterprises try to leapfrog.
In the room
- Host
- Jesse Hopps
- Room
- 30 senior executives from Abbott, Baxter, Coca-Cola, Constantia Flexibles, Covestro, CSL, Electrolux, ExxonMobil, Galderma, Honeywell, Inova HealthTech, Johnson & Johnson, LANXESS, LinkedIn, Lonza, Lumen Technologies, Mitsubishi Chemical, Mundipharma, Novonesis, PostNord, SABIC, Signify, Solventum, TD Synnex, Zalando, and Demand Metric
Thirty senior executives, and not one at five out of five on agentic deployment
A live poll opened the session and set its tone: few above a three, zero votes at five. What followed was ninety minutes on why the bottleneck is the coalition rather than the technology.
Thirty senior executives across pharma, chemicals, medtech, industrials, telecom, and e-commerce. The session runs under the Chatham House Rule, so contributors below are described by role and industry.
The short version
“Stage one is individual productivity. Stage two is process enhancement. Stage three is real enterprise advantage.”
- Asked on a 0–5 scale how much real agentic AI their team was deploying, the room returned zero votes at five, eleven percent at four, and a third at three.
- The room did not converge on a definition of “agentic” — and that ambiguity is itself the bottleneck.
- The gap is not the tools. Most mid-to-large-cap companies run the AWS, Microsoft, and Salesforce stacks they already own at roughly ten percent of capacity.
- A three-stage maturity model became the frame most of the room adopted — Assist, Process Enhancement, Enterprise Advantage — and most enterprises try to leapfrog the first two.
- Running agents at enterprise scale calls for a platform — governance, sandboxing, token monitoring, access control — rather than a fleet of ad-hoc agents.
The definition problem
The session opened with a poll: on a 0–5 scale, how much actual agentic AI — not chat, not Copilot summaries, but agents executing workflows — is your team deploying right now? Few landed above a three, and five drew zero votes. That says most of what you need to know about where the market sits, against the volume of noise about agents replacing the workforce.
The conversation then tried to pin down what “agentic” means, and the group did not converge.
“The deeper you go into it, the more fuzzy the definitions get.”
Even her own engineers debate whether a given workflow is agentic enough to qualify — too agentic to build on the current platform, or not agentic enough to bother.
“Regular AI is the chatbot you ask, how do I make this recipe? Agentic AI does the cooking, buys the stuff that needs to be bought, looks in the fridge, checks what’s there, and plates the whole thing for you.”
A supply-chain transformation leader at a global appliance manufacturer drew the line at execution and cost: a real agent is generative AI plus execution, and tokens cost money — which makes deterministic automation the right answer whenever an agent is not strictly required.
“Run this on a schedule and put the result here for me, versus run this and take these actions.”
Both readings turned on the same line: where the work stops being a suggestion and starts being an action taken on the company's behalf.
Six use cases from the room
- 1
Compliance-guardrailed role-play simulator — pharmaceuticals
A Salesforce-integrated AI simulator runs sales role-play against compliance-cleared AI personas, with instant coaching feedback on pace and message. The competency data that comes out of it lets managers spot territories set up for growth.
- 2
Tiered automation across the account base — global telecom
Generative AI for meeting notes and ad-hoc enterprise queries, agentic content nurture for the long-tail mid-market, and next-best-action served to enterprise reps inside the CRM — with the human deliberately kept in the loop on the largest accounts to protect the relationship.
- 3
Four interconnected pieces — global IT distributor
Data signals that surface insights, next-best-action on those signals, AI-distilled value-proposition content drawn from internal libraries few people have time to read, and an AI avatar that lets reps practice the pitch before the real conversation.
- 4
Forecasting with write-back agents — global appliance manufacturer
AI-augmented forecasting paired with high-confidence master-data correction agents that write back into the planning systems. Agents are used only where context has to be understood and probabilistic correction earns its keep.
- 5
Tariff circumvention and anti-dumping evidence — specialty chemicals
External trade-flow data combined with internal product data to identify tariff circumvention by competitors and build evidentiary cases for anti-dumping action. A high-leverage use that rarely comes up at a commercial excellence roundtable.
- 6
The deliberate counterweight — healthcare technology
This team started with low-risk summarization and held back from customer-facing decisions on purpose. Their CEO’s illustration: asked about its own subscription pricing, the model got it wrong. In regulated industries hallucination risk is non-trivial, and skill-set discipline matters more than tool count.
The real bottleneck is coalition, not tools
“We have tools, right? What we’re doing is using Copilot to try to produce the kind of stage-three level — it’s kind of like using Google Maps to navigate an F1 race.”
Most mid-to-large-cap companies, in her view, run their AWS, Microsoft, and Salesforce stacks at maybe ten percent of capacity. The gap is not the technology. It is the absence of IT, commercial, supply chain, and finance leaders sitting at one table long enough to define the job to be done before anyone picks a tool.
“We don’t have an AI strategy; we have a business strategy of how we want to be as an organization. And then how does AI or agentic AI become an accelerator of that?”
His company runs three lanes in parallel, each with its own governance: personal-productivity AI, operational-excellence AI, and sales-growth-acceleration AI. Defining the lanes is what turned the CEO directive to be AI fluent into something useful, rather than the thing that had AI popping up like mushrooms across the company.
The maturity model the room adopted
- 1
Assist
Trigger an automated email when a contract hits ninety days from renewal. Low IT governance, individual productivity wins, and it builds the confidence and prompt skill the later stages depend on.
- 2
Process enhancement
Same trigger, but retrieval pulls customer-service transcripts, billing history, and account context into the decision before the email goes out. More IT involvement, and document quality starts to matter.
- 3
Enterprise advantage
Event-condition-action workflows on unified data: the renewal trigger checks margin tier and supply-chain status, then routes to the right human with the right play. This is where a unified data platform earns its keep — which is why it has to be built during stages one and two.
Building for scale
Running production AI at enterprise scale calls for a platform rather than a fleet of ad-hoc agents. One team is building exactly that — LLM-agnostic across several model providers, with an internal surface described as an app store but for agents, and connectors into the CRM, marketing automation, data layer, and email systems. The reason is governance, sandboxing, token monitoring, access control, and being able to maintain agents over time.
“How are you all talking to your leadership teams about what we’re actually trying to accomplish? Because it can go a number of different ways, and you can roll out another massive transformation project where the whole organization is already overwhelmed.”
The room did not fully answer her. The implicit answer was the medtech lead’s: get the right coalition into the room first, then pick the technology.
Pulling solutions out rather than pushing them down
The specialty-chemicals CFO offered the most concrete pattern: a Commercial Excellence Academy, with fifty-two Growth Projects launched over the past year. Bottom-up problem definition, external operator-coaches assigned, cross-business-unit teams, and leadership acting as a blocker-removal layer rather than the initiator. Some are AI-flavored; most are commercial. Several have since been scaled into corporate solutions and folded back into regular business processes.
Personal AI comes before enterprise AI
The session closed on a different scale of the same problem. The host described running the Consortium on Claude Code with roughly seventy skills, twenty integrations, and forty to fifty agents — doing the job of ten people as one person part time. Then he polled the room by show of hands.
- Who has started using Claude Code or something like it personally? Most hands went up.
- Who feels ready to coach someone else through it? Not a hand.
- Would a structured cohort program help — themed weekly format, in a group, with a coach? Interest was unanimous.
Practices to apply immediately
- Settle the definition before the design debate. Until business and IT agree what agentic means for your organization, scope arguments burn time without producing scope.
- Reach for deterministic automation whenever an agent is not strictly required. Agents consume tokens, and tokens are an operating cost that grows with use.
- Start at Assist. Individual-productivity wins build the confidence and the prompt skill that the later stages depend on.
- Get IT, commercial, supply chain, and finance at one table to define the job to be done before anyone picks a tool.
- Audit what you already own. Most mid-to-large-cap companies run their existing stacks at roughly ten percent of capacity.
- Build the platform before the fleet: sandboxing, token monitoring, access control, connectors, and a way to maintain agents over time.
- Stay model-agnostic. Different providers lead on different tasks, and the leader changes often.
- Keep a human in the loop on your largest accounts, whatever the automation can do across the long tail.
Questions the room worked through
- What is agentic AI, and how is it different from a chatbot?
- The analogy the room liked: regular AI is the chatbot you ask how to make a recipe; agentic AI does the cooking, buys what needs buying, checks the fridge, and plates it. The sharper technical line drawn in the session was generative AI plus execution — and because execution consumes tokens, an agent is only the right answer where deterministic automation would not do the job.
- How much agentic AI are large enterprises running today?
- A live poll of thirty senior executives asked exactly that on a 0–5 scale. Zero votes at five, eleven percent at four, a third at three, and the rest at one or two. This was a room of advanced practitioners, so the honest answer is that very little real agentic execution is in production, whatever the volume of noise suggests.
- Why do agentic AI projects stall?
- The room’s answer was the coalition rather than the technology. Most mid-to-large-cap companies already run AWS, Microsoft, and Salesforce stacks at roughly ten percent of capacity. What is missing is IT, commercial, supply chain, and finance leaders at one table long enough to define the job to be done before anyone picks a tool.
- What are the stages of agentic AI maturity?
- Three. Assist: trigger an automated email when a contract hits ninety days from renewal — low governance, individual productivity, builds confidence. Process enhancement: the same trigger, with retrieval pulling service transcripts, billing history, and account context into the decision first. Enterprise advantage: event-condition-action workflows on unified data, where the trigger checks margin tier and supply-chain status and routes to the right human with the right play. Most enterprises try to leapfrog the first two and fail at the third.
- Should you use an agent or ordinary automation?
- Agents earn their keep where context has to be understood and probabilistic correction is required — master-data correction at high confidence, for example. Everything else is cheaper and more predictable as deterministic automation. Treating agents as a general-purpose replacement for automation is expensive, and the cost scales with adoption.
- Why build an agent platform rather than individual agents?
- Governance, sandboxing, token monitoring, access control, and the ability to maintain agents over time. One team is building an LLM-agnostic platform with connectors to CRM, marketing automation, the data layer, and email, plus an internal surface they describe as an app store but for agents. Without that scaffolding you get agents appearing across the company that few teams can maintain.
- How do you talk to a leadership team about AI without launching another transformation program?
- Lead with the job to be done rather than the technology, and frame AI as an accelerator of the business strategy rather than a strategy of its own. As one leader put it, his company has no AI strategy — it has a business strategy, and asks how agentic AI accelerates it. Defining separate lanes with their own governance, for personal productivity, operational excellence, and sales growth, is what keeps the ambition from turning into sprawl.
- How do you get bottom-up momentum rather than another top-down rollout?
- The most concrete pattern in the room was fifty-two Growth Projects run over a year: bottom-up problem definition, external operator-coaches assigned, cross-business-unit teams, ninety-day scope with a defined end goal, and leadership acting as a blocker-removal layer rather than the initiator. Several have since been scaled into corporate solutions.
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