Walk into any large enterprise’s AI program in 2026, and you’ll find a familiar slide. It’s titled something like “Use Case Portfolio” or “Agentic AI Initiatives.” The list runs twenty or thirty items long. Almost every entry describes a way to make an existing workflow faster, cheaper, or more accurate. Automate FNOL triage. Summarize underwriting submissions. Triage tier-one support tickets. Generate first-draft contracts. The work being done by the AI is, in every case, work that already existed before the AI showed up. The AI is doing it in fewer minutes and with fewer people.
Call this door one. Most enterprise AI lives behind it. There are good reasons for that, and we’ll get to them.
Door two is the door less taken. It asks a different question. Not how do we do this work faster? But given what agents can now do, is this still the right work? The answer is sometimes no. Not in the sense that the goal goes away, but in the sense that the steps to reach the goal could be radically different from the ones a human-staffed workflow evolved to use.
These two doors lead to different futures, and later I’ll add a third — a narrow one, for workflows that don’t exist yet. But the choice that matters most is between these two. Most enterprises pick door one first, and most should. The mistake is treating door one as the destination rather than the entry point, and budgeting accordingly.
Door one: agentify the work you have
The instinctive move with a new capability is to apply it to existing work. The instinct is right most of the time.
A claims operation runs FNOL today through a call center plus a digital intake form, with human triage. Door one says: put an agent on the intake form, have it ask the right follow-up questions, classify the loss type, and route to the right team. The workflow doesn’t change. The work changes hands.
A bank’s underwriting team manually reviews two hundred small-business loan applications a week. Door one says: an agent reads each application, surfaces the credit signals, drafts a recommendation, and presents it to the human underwriter who approves or revises. The application form is unchanged. The reviewer’s role is unchanged. The output is the same kind of decision in less time.
This is augmentation. It’s good work. The ROI is legible, the regulatory exposure is bounded, the org chart doesn’t move, and the business case writes itself: we processed N% more cases with the same headcount, our cycle time dropped from X to Y, our error rate held steady or improved. Boards understand it. CFOs underwrite it. Operating teams don’t need to learn a new mental model. They’re doing the same job with a faster tool.
Most enterprises in 2026 should have a large door-one program. Probably the largest line in the AI portfolio.
What door one quietly preserves
The thing door one leaves intact is the workflow itself, and that’s worth thinking about, because every workflow inside an enterprise is a fossil of its era.
The FNOL process exists in roughly its current shape because, in 1985, the cheapest way to capture a loss notice was a phone call to a human operator. The steps that follow (first notice, coverage confirmation, assignment, contact, inspection, estimate, settlement) assumed a sequential, human-mediated flow because that’s what the tools allowed. The form fields, the handoffs, the documents, the timing: all of it is downstream of constraints that have since dissolved.
When you put an agent on FNOL today, you’re scaling that 1985 design. The agent is faster than the human operator. The form is still the form. The sequence is still the sequence. The handoffs are still the handoffs. Nothing structural has changed about how a claim moves through the carrier.
This works fine for a while. The cycle time drops. The cost-per-claim drops. The numbers look good on a quarterly board update. What’s not visible in those numbers is the opportunity accumulating in the next-door category. That’s where door two lives.
Door two: redesign the work given what’s now possible
A different question: if we were standing up this carrier today, in 2026, with agents who can extract structured data from voicemails, photographs, and policy documents in parallel, would we have an FNOL step at all? Or would the moment a customer takes a picture of their damaged car and texts it to a number be the moment a draft claim already exists in the system, fully reserved, with a recommended next action waiting for the customer to confirm?
That’s door two. Same goal (pay the right claim, fast, with appropriate scrutiny), different work to get there.
A few sketches of what door two looks like in practice.
In a credit shop, door two doesn’t make the underwriter faster. It asks whether a twelve-page application form is the right input for a small-business loan when the agent can read three years of bank statements, a year of POS data, and the operator’s public business history, and synthesize a position in seconds. The form may shrink to two questions, or disappear. The “submission” ceases to be an event and becomes a moment in an ongoing relationship.
In customer service, door one is a better triage of incoming tickets. Door two notices that a meaningful share of tier-one tickets is predictable from in-product behavior thirty seconds before the customer files them, and starts intervening before the support form opens. The ticket volume drops because the upstream workflow that generated the tickets has been changed.
In hiring, door one is faster resume screening. Door two asks whether the resume is the right artifact at all when a candidate’s actual work history is partially observable through public artifacts and a structured forty-five-minute interview with an agent produces a higher-signal record than a resume ever did.
In each case, the destination is roughly the same (the claim paid, the loan decided, the support issue resolved, the role filled), but the path to get there is reshaped, sometimes radically. The agent isn’t doing the old work faster. The old work has been replaced.
Why is door two harder
Three reasons, in roughly the order they bite.
First, door two threatens the org chart. Door one slots into existing functions, reporting lines, and budgets. Door two changes what each function does, sometimes who owns it. The FNOL example earlier (where a customer sends a photo and a draft claim already exists) crosses the boundary among the contact center, the claims operation, and the data team. Three functions. Three budget owners. Three sets of incentives. Door two work is unsponsored until someone takes it on, and there are good reasons no individual function wants to take it on.
Second, door two has worse near-term ROI math. Door one tells you, in advance, how many cases will move, by how much, and what the savings are. Door two is a redesign exercise with a longer payback period, less certainty about the endpoint, and a non-trivial chance that the first attempt is wrong. The first six months of door two often produce no measurable cost reduction. The first twelve sometimes produce a regression: you’ve started building something new and haven’t fully retired the old.
Third, door two requires you to know what work you’re actually doing. Door one can be done from process documentation, even when that documentation is partially fictional. Door two cannot. You can’t redesign a workflow if you don’t know how it actually runs: the variants, the workarounds, the exceptions, the moments where senior staff override the script. Most enterprises haven’t observed their own work at that resolution. Door two stalls in the discovery phase.
None of these reasons is an argument against door two. They’re arguments for treating it differently: as a different kind of program, with different budgets, different sponsors, and different success metrics.
The 80/15/5 rule
A working allocation, drawn from the carriers and banks that have done this well, looks roughly like this.
Eighty percent of AI spend goes to door one: augmenting existing workflows. This is the bulk of the program. It’s where the cycle time drops, the cost-per-case improvements show up, and the year-one ROI lands. It’s also where the operational learning happens. The team finds out what agents are reliable at, what fails in production, and what the operational overhead of running an agent fleet actually looks like. Don’t underweight it. Most of the answer is here.
Fifteen percent goes to door two: selective reimagining. Not a whole-business redesign. One or two workflows, chosen because either (a) the existing design is visibly creaking, (b) a competitor is moving and you can see what they’re doing, or (c) the unit economics of the current process are bad enough that incremental improvement won’t fix them. The 15% is enough to fund a small team, a long time horizon, and one workflow taken seriously. It isn’t enough to fund six. Constraint is part of the design.
Five percent goes to door three: unbuilt workflows. The option-value bucket. It funds two or three speculative bets a year on workflows that don’t currently exist in the enterprise. Still, it might: proactive customer interventions, agent-to-agent commerce, predictive operational interventions, whatever the speculative idea of the season is. Most won’t ship. The ones that do will define the next decade for the firms that funded them.
The ratios aren’t arbitrary. A 100% door-one program creates the appearance of progress while leaving the firm exposed to a competitor who makes the door-two move. A 100% door-two program produces no near-term wins and burns down credibility before the bets mature. A program with zero door-three loses option value on a step-change opportunity someone else will discover.
The trick isn’t picking one door. It’s running all three at the right ratios, with the right governance, and with honest expectations about which ones produce wins on which timelines.
Why most programs are 95/5/0, and what it costs
In 2026, the typical Global 2000 AI portfolio is roughly 95% door one and 5% what the slide deck calls “innovation,” which mostly funds proofs of concept for door-one ideas in adjacent functions. Door two is rare. Door three is nearly absent.
This is the rational result of the incentives I described above. It’s also expensive.
The cost shows up in two ways.
First, in the process debt. Every door-one investment hardens the underlying workflow. Once you’ve integrated an agent into FNOL, FNOL becomes harder to redesign. The agent is part of the production system, and any rethink has to account for it. After three years of door-one work, the firm has an AI-enabled version of its 1985 operating model. And the 1985 operating model is what the AI is now defending.
Second, in competitive exposure. The firms that ran 80/15/5 instead of 95/5/0 spent the same total dollars and produced roughly the same year-one ROI. After three years, they’ve also stood up two or three workflows that don’t exist at all in the 95/5/0 firm. The 95/5/0 firm doesn’t know it’s behind, because the comparison is invisible. Then a competitor’s small-business lending unit closes loans in forty-eight hours from initial inquiry with no application form, and the firm with the door-one program finds it can’t match without a year of redesign work it never started.
This is the structural reason agentic AI rewards a different kind of strategic patience than the cloud era did. The cloud era largely rewarded execution: pick the right vendor, move fast, capture the productivity gain. The agent era rewards execution plus a small, sustained, well-budgeted investment in the work that doesn’t pay back this quarter. Programs that run only the first half will look fine for two years and behind for the next five.
The transition, not the choice
What we are discussing is the premise of “from workflow automation to workflow reimagining,” and the preposition matters. The argument isn’t that automation is wrong and reimagining is right. The argument is that automation is the entry point, reimagining is the destination, and most programs stop at the entry.
Door one is where you build operational competence with agents. You learn how to deploy them, monitor them, fail safely, handle exceptions, and govern them. You can’t skip this. The 80% augmentation work informs the 15% reimagining work: it draws on the same context, observability layer, and trust architecture.
But once the operational competence exists, the reimagining work becomes possible in a way it isn’t on day one. The firms that move from 95/5/0 to 80/15/5 around year two of a serious agentic program (once they know what their agents can actually do in production, once they’ve seen the failure modes, once the platform is stable) are the firms that will have something interesting to show at year five.
The trap is staying at 95/5/0 forever, because the door-one program never stops generating defensible wins, and the case for door two never quite makes it to the top of the slide.
The model is ready. The platform is ready. The reimagining is the part nobody’s been forced to do yet. That changes soon.
What do you think? Please share your thoughts.