Most AI Isn’t AI. That’s About To Matter.

The Place Where AI Has Nowhere To Hide

A colorectal cancer screening result arrives at a clinic via fax, where 88% of healthcare practitioners say fax-related delays negatively affect patient care. It’s marked positive. It lives inside a PDF. No one sees it unless someone manually opens the document, reads it and decides what to do next. In the meantime, the patient waits. The follow-up doesn’t happen. The care gap widens.

This isn’t a technology failure in the conventional sense. The fax was received. The document was stored. By most definitions, the system worked. What failed was everything that needed to happen next, and no AI tool in the workflow was responsible for making it happen.

I’ve spent more than a decade in health tech, and this is what makes healthcare the clearest lens for evaluating AI: every bottleneck has a human cost, and that cost is impossible to ignore.

Healthcare is where the real standard for AI gets set first. And what’s visible here is coming for every industry.

By 2027, the organizations that have moved fastest will be running a different architecture entirely. Winners will be those with true integration, as Gartner projects that over 40% of agentic AI projects might be scrapped due to lack of real value. Instead of stacking SaaS point solutions and relying on staff to connect them, they’ll have autonomous layers sitting directly on top of systems of record, executing workflows end-to-end without human coordination at each handoff. SaaS products that can’t be absorbed into that model or justify their place alongside it face something worse than a hard renewal conversation. They face obsolescence.

Not All “AI” Is AI

SaaS spent two years calling three very different things by the same name. That conflation has consequences. While some products are truly agentic, others are simply layers of AI on old tech, a distinction increasingly recognized as AI agents reshape SaaS markets.

The first category is AI in name only. These are existing SaaS products with a thin layer added: a summarization feature here, a classification tool there. The roadmap hasn’t changed. The product hasn’t changed. The only thing that changed is what it says on the website. The AI is a marketing frame. The development direction hasn’t moved.

The second is AI as an umbrella term. Vendors bundle machine learning, GenAI, set workflows and genuinely agentic capabilities together and call the whole thing AI. The question worth asking is whether AI is driving the product vision or primarily framing the sales conversation. In most cases, if pressed hard enough, it’s the latter.

The third is genuinely agentic AI. These systems act flexibly within parameters, interpret inputs and take action without waiting for a person to decide what happens next. This is the closest thing to autonomous AI available today, and it represents a real capability leap. But capability isn’t the same as completion. Even the best agentic systems on the market stop well short of owning a workflow end-to-end.

The adoption arc of the past few years tells the same story on a loop. 2022 was run by vaccine scheduling and digital front doors. 2024 was dominated by revenue cycle management (RCM) and coding automation. 2025 brought AI scribes and ambient listening tools. 2026 is the year of agentic voice, when healthcare companies exhibit ROI and agentic capabilities at industry events like HIMSS. Each cycle, the industry finds a new category, declares it transformative and moves on. The underlying problem stays exactly where it was: work still depends on people to carry it forward.

Call it what it is: most AI investments aren’t changing how work gets done. Buyers end up selecting from the wrong category entirely, convinced they’re buying transformation when they’re buying an upgraded inbox. Enterprise AI adoption has expanded massively, yet most systems still haven’t integrated AI into operational workflows at scale. The problem lives in the roadmap. The messaging just obscures it.

For buyers, misreading the taxonomy has a specific cost. It locks in coordination expenses that compound over time. Organizations that invest in the first two categories are paying for AI while still staffing for the same operational burden. The headcount required to hold the workflow together doesn’t shrink. It just gets more expensive to justify.

Good AI, Wrong Definition of “Done”

Take the fax workflow. It’s a useful illustration of where the standard is today and how fast it’s moving.

The worst systems parse the fax, surface the content and leave it for staff to act on. That’s a better inbox. The good systems go further: they extract actionable next steps and execute them, routing the referral, triggering patient outreach, scheduling the appointment. That’s genuine task execution, and a few years ago it would have felt like a significant leap.

The next phase makes even that look incomplete. The standard that’s emerging is whether a system can own the entire downstream journey: following up with the patient, following up with the referring clinician, persisting until the visit actually happens, closing the loop, flagging exceptions without being asked. No handoffs. No one managing it in between.

Most current systems, even genuinely agentic ones, are operating at the middle phase. They execute a task. They still depend on humans to coordinate between tasks. That’s the definition of “done” the market is about to move past.

Execution of individual tasks where people coordinate between them won’t be the standard. Full autonomous coordination across a workflow, from first input to confirmed outcome, is where expectations are heading within the next year.

Output is not execution. Task execution is not workflow ownership. The industry is only beginning to reckon with the second half of that distinction.

For builders, this is where competitive advantage is already starting to separate. Systems that stop at outputs are becoming easier to compare, easier to commoditize and harder to defend at renewal. Systems that execute tasks without owning the full outcome are only one phase ahead. The defensible position belongs to systems that can run a workflow from start to confirmed finish, autonomously across systems over time.

What the Next Phase Actually Looks Like

A few years ago, I would have described the agentic AI category and called it the destination. It isn’t. It’s a waypoint. The next phase will be defined by systems that take responsibility for completing work across systems of record. Improving a step within one system is the old bar.

The old model is straightforward: software performs a function, and people connect the steps. A system extracts information. A person decides what to do with it. Another system schedules the appointment. Another person follows up. The coordination burden lives with the staff. It has always lived with the staff. That’s the model AI has been layered onto, not the model it’s replacing.

The emerging model shifts that responsibility. Systems carry the workflow forward. People step in when something breaks, requires judgment or falls outside policy. The coordination is built into the system, not absorbed by the team.

By 2027, the systems that matter will be those connected directly to core systems of record, executing everything needed to reach an outcome without waiting for human intervention at each handoff. Owning the result from end-to-end.

Building at this level has specific requirements: integration across fragmented systems that weren’t designed to work together, reliability in production environments where edge cases are constant, compliance with regulatory constraints that vary by context and continuous ownership of outcomes across the entire workflow. Most organizations aren’t structured to build this internally. The vendors who can will define the next era of enterprise software.

The vendors who can’t will find themselves in the same position as the SaaS point solutions they replaced: useful for a while, then in the way.

What This Means if You’re Building, Buying or Betting on AI

For Builders

The uncomfortable truth is that most AI systems are architected to assist, and assistance doesn’t compound. It doesn’t expand. It doesn’t become harder to displace over time. Execution does.

The question worth reorienting entire roadmaps around: do systems eliminate the need for human coordination across the workflow? Whether they qualify as agentic is a lower bar, and the gap between the two is where most builders are currently sitting.

Getting to execution requires owning integration first. Builders can’t trigger downstream actions in systems they’re not embedded in. Surfacing recommendations into a workflow is a fundamentally different capability from operating within it. The handoffs in product, the moments where a person must pick up what the system put down, are the clearest signal of where builders are on this spectrum. Audit them. Everyone is a liability in a market consolidating around systems that eliminate them.

If a builder’s roadmap is still oriented around making outputs smarter, they are optimizing around a problem that will catch up with them. The window to reposition toward execution is open now. It won’t stay open indefinitely.

For Buyers

If pilots are delivering value, keep them running. The question is whether they’re leading somewhere. Evaluate whether the vendor can expand beyond improving a single step.

That starts with the roadmap. Ask every vendor where they’re headed and apply the taxonomy to what’s heard. Are they solving discrete problems with AI, or building toward end-to-end execution? A vendor still oriented around point solutions will keep selling buyers point solutions. The roadmap tells buyers which kind they’re dealing with before the contract does.

Once buyers know where they’re headed, they must pressure-test the metrics. A vendor advertising 80% automation rates or high containment scores will provide output data. Push past it: what percentage of workflows complete without human intervention end-to-end? How does that number hold up as volume and complexity increase? Run the math on what genuine coordination reduction would mean for operational costs.

Then get the contracts right before scaling. As AI embeds across scheduling, intake, follow-up and referrals, costs are calculated differently than traditional SaaS. Model cost per AI-assisted interaction the way one would model cloud computing. Negotiate accordingly, with spend caps, tiered commitments and price protections built in. How buyers negotiate their Azure spend is going to start to look like how they negotiate all their spend. The organizations that skip this step when volumes are low regret it when volumes aren’t.

For Investors and Analysts

Feature depth and interface quality are no longer reliable proxies for defensibility. The companies that will expand their operational footprint are those taking responsibility for outcomes across systems of record. Assist-layer products face margin pressure and consolidation risk as buyers get more sophisticated about what they’re actually purchasing. When evaluating AI vendors, ask where the human handoffs are. The answer tells investors and analysts more about long-term competitive position than the demo will.

The Standard Is Being Set Now

Healthcare will show the rest of enterprise software where this shift lands first. The stakes are too high and the inefficiencies too costly for the gap between output and execution to stay invisible. The organizations that move now will have built what everyone else is trying to catch up to. The next phase belongs to systems that complete work. Everything else is overhead.