What was on the table
Three questions, one evening.
Everyone has a few AI super-users saving hours a week, and the conversation turned on one question: what separates the organisations making that the norm across a whole team from those running pilots that never reach the bottom line. The dinner ran over three courses, and each course carried its own question.
Champions don't replicate alone
How do you get a whole team AI-fluent, not just a few stars?
Every organisation now has super-users. Almost none has made that the norm. The first course dug into why the jump from the few to the many is so hard.
Beyond the pilot
Why does most AI never reach the bottom line?
Pilots multiply, agents get built, and the P&L does not move. The table dug into measurement, incentives, and enabling a team versus making it proficient.
Redesign, don't digitise
Rebuild the work around AI, or just speed up the old way?
Most AI effort makes the existing process a little faster. The teams pulling ahead are rebuilding the process itself, a different order of change.
The four conclusions
Where the room landed.
The two tables never heard each other's conversations, yet both landed on the same four conclusions. That independent convergence is the strongest signal of the night. Tap each conclusion for what was actually said around the table.
Champions don't replicate alone +
A few people get good and become proof it works. Then it plateaus, because the skill lives in their heads, not the workflow. One leader likened handing someone AI to putting them in an F1 car: getting in is easy, driving it well is not.
The champion trap was confirmed around the table without prompting: teams get inspired by one or two examples of what a champion does, and that is all they ever use AI for. People are being asked to deliver more without being given the time to learn to drive, and the only model attendees had seen work is someone embedded in the team who takes workload off while enabling people at the same time.
The gap is rarely ability. It is permission (people unsure what they are allowed to use), incentive (nobody rewarded for changing how they work), and design (the AI-assisted route takes more effort than the old way, so the old way wins). Adoption spreads when the AI-native route is the easier route.
Training doesn't transmit it +
Tick-box webinars and engineer-built courses do not move behaviour. People pick this up by working alongside someone who already does it. Every table had a version of the same story.
One attendee described their enterprise L&D catalogue as AI made for engineers, not for marketers: a marketer-shaped hole in enterprise training that everyone recognised. Another described their organisation's training as a ticking-the-box exercise led by HR rather than anything hands-on. One global business had dismissed its entire dedicated training academy just weeks before the dinner.
The alternative the room kept returning to: fluency doesn't transfer through documents or training decks, it transfers through someone sitting inside the team, doing the work alongside them, and being measured on how many colleagues adopt rather than on their own output.
Fear is the real blocker +
When people fear for their jobs, no mandate makes them believers. What works is honest, repeated communication about what they will do with the time AI gives back.
Fear operates at two levels: fear inside the organisation (job security) and a deeper fear of not being able to keep up with working life without these skills. Secret use is real: one attendee admitted closing their AI assistant whenever their boss walks past, despite having seen that same boss using it. Adoption is happening in the shadows because organisations have not made it legitimate.
The unblocking tool is communication that paints the actual picture. Telling people to use AI "so you can do more strategic work" was called out as meaningless. Leaders need to describe the future work concretely, and repeat the why, the where, and the what-it-means-for-you relentlessly, because workload erases it.
Redesign beats digitisation +
The teams pulling ahead stopped speeding up the old way and rebuilt it. Efficiency-only AI strategies looked self-limiting from every seat at the table, whatever the sector.
Because experienced people carry twenty years of workflow assumptions, the job may be unlearning before learning. Taking a legacy workflow and asking "how would I do this with AI" fails; the workflow has to be redesigned. As one attendee put it, you have to think of the workflow before you can think of the agent.
The most advanced organisations pair redesign with innovation, pursuing new products, services and experiences rather than pure efficiency. Doing the same thing slightly worse but a lot cheaper is not a strategy, it is a ceiling.
The most popular business strategy in the world right now is to do the same thing that you used to do, slightly worse, but a lot cheaper.
Redesign in practice
Speeding up the old way, or rebuilding it.
Replace one step with AI
Swap a task for a model and call it done. The gains are real but shallow, and the workflow underneath stays exactly as it was designed for people, not for AI.
Rebuild the workflow around AI
Map the work first, then ask what the process would look like if you built it today. As one attendee put it, you have to think of the workflow before you can think of the agent. The teams doing this describe an end-state where routine steps disappear behind a single human approval.
What we took from the night
The organisations making real progress were not the ones with the most tools or the biggest training budgets. They moved capability from the few to the many, and rebuilt the work rather than bolting AI onto it. A people problem before a technology one, and the reason most teams feel they do not have time to save time.
The deep dive
How established enterprises are actually making the journey.
The most valuable thing about the room was who was in it: leaders from large, complex organisations that were not born as technology or AI-first companies. Banks, pharma, consumer goods, financial services, established platforms. The question they all carried, in one form or another:
Here is how the room answered, drawn from both tables.
Part one · The starting point
Three adoption cultures, three failure modes.
Every established enterprise in the room fell somewhere on the same spectrum. None of the three archetypes has solved diffusion, but each fails differently, and knowing which one you are is the first honest step of the journey.
What it looks like
AI usage is tied to performance expectations, scope, and promotion. Managers are personally accountable for their team's adoption. Communication is relentless: the why, the where, and the what-it-means-for-you, repeated monthly because workload erases it. In the most advanced version at the table, adoption passed 70% and doubled inside three months.
The failure mode
Measuring usage instead of outcomes. When usage itself is the metric, people ask AI for the weather every morning to hit the number. Fast uptake, hollow diffusion, and internal agents that average one or two users each: enormous building energy, near-zero spread.
What it looks like
Tools are bought and permitted, and that is where it ends. Regulated environments in particular default here: agency contracts signed with AI clauses unresolved because internal guidance doesn't exist yet, licences issued with no expectation attached, and adoption left to individual initiative.
The failure mode
Stall. Permission without incentive produces exactly what you would expect: the champions carry on, everyone else waits. Nobody is rewarded for changing how they work, so nobody does, and the gap between the few and the many quietly widens.
What it looks like
Relationship-driven, physical-product, or heritage businesses where automation can feel like it goes against the culture itself. The pattern one leader described: lots of pilots, lots of noise pushed by the hype, but nothing at scale, no KPIs, no incentives. Even headline AI moments read internally as PR rather than transformation.
The failure mode
Pilots and noise, nothing at scale. The evolution these organisations described runs fancy toy, then fear (copyright, hallucination, responsible AI), then structure. Slower, but when structure arrives it tends to be smarter.
Part two · Where agents are landing
Where complex organisations are embedding agents first.
A clear pattern emerged: the organisations getting past pilots start with universal pain, encode senior judgement into the agent, prove confidence in one workflow, then expand to adjacent ones. These were the beachheads described around the table.
Part three · Pilots to scale
The playbook for getting past the pilot.
Most organisations at the table admitted they introduced tools with no baseline, so they are now hunting for incremental gains against a "before" they never measured. The ones escaping that trap follow a recognisable sequence.
Baseline before you build
Map the workflow end to end and time it, down to tasks as small as booking a call. Nobody can prove "after" without a "before", and the missing baseline was the single most universal confession of the night.
In practiceAny organisation that maps and times its workflows before automating is ahead of nearly everyone.
Prioritise by jobs, not by demos
The most disciplined model in the room: define every function's key jobs to be done, get them approved at VP level, estimate time spent empirically, and prioritise automation against the biggest jobs rather than the loudest pilot.
In practiceRank workflows by volume, handoffs, and hidden pattern-matching. Low-frequency, high-stakes work usually is not worth rebuilding yet.
Decentralise ideas, centralise what works
Let ideas generate at the edges, where the work is. Then add a deliberate second layer of prioritisation and editorial that surveys what all teams want, spots the three things six teams all need, and builds the common things first. Without that layer you get agents that only their maker uses.
In practiceIf everybody builds their own agents and nobody learns from each other, the building energy never becomes scale.
Aim for proficiency, not just enablement
Enablement hands teams processes to consume: someone builds an AI-assisted process and everyone uses it. Proficiency means people can build and change their own. Most rollouts stall at the first because the driver is pressure to show return on spend, not a considered view of the workforce.
In practiceAsk who could rebuild this workflow if the tool changed tomorrow. If the answer is one central team, you have enablement.
Govern usage, not just access
Cost discipline is arriving. With hundreds of people on the tools, indiscriminate usage gets expensive fast, and token economics will force the prioritisation discipline most organisations currently lack. The mature question is no longer "who can use it" but "what is worth using it for".
In practiceThe honest test of any pilot: can anyone name the metric it moved, and would the team fight to keep it if switched off tomorrow?
Part four · Build vs partner
What to own in-house, and where an outside partner earns its place.
The room's view was consistent and a little counterintuitive: the goal of any partnership should be your own self-sufficiency. Dependency on any external team is a risk when the technology changes this fast. But two capabilities kept surfacing as the ones enterprises struggle to grow alone.
The work and the fluency
The people who are the users should be the ones who own and iterate the business processes, because when things change, and they will, waiting on an external team is the bottleneck.
- Ownership and iteration of everyday workflows by the teams that run them
- The judgement layer: what good looks like, encoded from your own leaders' feedback
- The diffusion rhythm: show-and-tells, AI challenges, peer sessions with an owner
- Governance of usage, cost, and where AI is deliberately not used
The catalyst and the build
Operational teams do not have time to save time. As Martin put it at the table, you need an outsider who is an insider: someone embedded who creates impact directly while converting the team around them, not through training but through working alongside them.
- Embedded experts who take workload off the team while enabling it, measured on colleagues' adoption, not their own output
- Engineering the rebuilt workflows, so your team is not asked to become builders overnight
- Centre of excellence design: proving use cases through embedded experimentation, then cascading what works globally
- The unlearning work: an outside eye on workflow assumptions your experience has encoded
You need an outsider who's an insider: to come in and create impact themselves, but also to infect the rest of the organisation, not through training, but through working with them, redesigning workflows, achieving impact.
Part five · The destination
What "great" looks like.
Drawn from the most advanced practices described at either table. Tick off what is already true of your organisation; the gaps are your redesign agenda.
Routine steps disappear behind a single approval
The redesign benchmark is not a faster old step; it is the step vanishing. Zero seconds of routine human work, one human decision.
AI to 80%, human polish to 100%
An honest heuristic that concedes imperfection: AI carries the volume, humans carry the judgement, and the work still ships twice as fast.
Workflow before agent, always
No agent is commissioned without a mapped, timed workflow behind it. Asking for agentic AI without one is asking for nothing.
Adoption is the KPI of your best people
Embedded practitioners are measured by how many colleagues change how they work, not by their own output. Fluency spreads person to person until it is self-sustaining.
Innovation, not efficiency alone
The most advanced AI organisations pursue new products, services and experiences. Efficiency-only strategies are self-limiting, whatever the sector.
Human imagination is budgeted, not squeezed
As output levels out, judgement and imagination become the differentiator. Great organisations list the work that resists prompting, naming, judgement calls, stakeholder nuance, and protect human time for it.
Tick the ones that are true for you today.
The ideas worth keeping
Six threads from the discussion.
Beneath the four conclusions, six ideas did the real work. Each comes with a way to put it to work.
Enablement is not proficiency
+Enablement hands teams processes to consume. Proficiency means people can build and change their own. Most rollouts stall at the first, and the difference is the gap between using AI and making with AI.
Put it to workAsk who could rebuild this if the tool changed tomorrow. One central team means enablement.
A mission, not a memo
+Change management does not shift how people work. What does: someone doing the same job, embedded in the team, converting people beside them. Conversion spreads person to person until it is self-sustaining.
Put it to workEmbed your strongest practitioner in one team for a quarter. Measure colleagues' adoption, not output.
Three cultures, three failure modes
+Mandates drive fast usage but shallow diffusion. Permission without incentive stalls. Culture-conflict produces pilots and noise, nothing at scale.
Put it to workName your culture honestly, then fix its failure mode rather than adoption in general.
Unlearn before you learn
+Workflow habit makes redesign harder, not easier. Experienced teams can have a harder adoption path than juniors, because expertise encodes the old way.
Put it to workTake one workflow you know by heart and ask which steps would exist if you designed it today.
Imagination, not intelligence
+As output levels out, judgement and imagination become the differentiator. AI builds from what exists. People bring the thing that was not there before.
Put it to workList the work that resists prompting, like naming and judgement calls, and budget human time for it.
Autopilot erodes credibility
+Lean on AI too heavily and you cannot defend your own work. The room spotted it instantly in people unable to back their own thinking.
Put it to workOne team rule: nothing ships that its author cannot defend line by line.
Take it back to your team
Three tests worth running this quarter.
Each conclusion collapses into a question you can put to your own organisation. Silence is a finding too.
What have we stopped doing because it never scaled past the champion?
If the answer is nothing, the plateau is being worked around rather than solved.
Which metric did this pilot move, and would the team fight to keep it?
Hesitation on either half tells you whether it is earning its keep or quietly burning budget.
Which workflow would we rebuild from scratch tomorrow if we had the mandate?
Whatever comes to mind first is where redesign should start.
Where Algomarketing comes in
Ran the tests? Bring us one workflow.
If any of the three tests came back silent, that silence is worth an hour. Pick the workflow you would rebuild first, and one of our embedded experts will map it with you live: where the time actually goes, which steps disappear behind a single approval, and what the rebuilt version looks like. You leave with the redesign case, whether or not we ever work together.