The Apply Year
Here's what's strange about agentic AI in the spring of 2026. The hardest part isn't the technology anymore.
We've spent the last quarter in operator briefings, board rooms, and LP calls, and the through-line has shifted in a way that surprised even us. A year ago, every conversation about agents bottomed out in a model problem: context windows, hallucination rates, tool-use reliability. Today, every conversation bottoms out in an absorption problem. Whose workflow does this fit into? Who owns the decision when the agent gets it wrong? What does the org chart look like when half the team's outputs come from software that didn't exist last quarter?
This is a different kind of problem, and it points to a different kind of investable surface. Capability is no longer the binding constraint on agentic AI. Absorption is. The phrasing one of our LPs used recently captures it neatly: 2025 was the year of the copilot, 2026 is the year of the coworker. A copilot suggests. A coworker takes the action it can defend, escalates the calls it can't, and learns from the human who steps in. The first is a feature. The second is an operating model.
What follows is what we're seeing across the portfolio, in our diligence pipeline, and in the production deployments operators have walked us through. Five patterns, in roughly the order they're shaping our investment posture this year.
One: The Deferral Architecture
The most consequential design pattern in production agentic systems isn't a model choice. It's a routing rule.
Consider a healthcare claims workflow. A member uploads a receipt. The system reads it, scores its own confidence, and either adjudicates immediately (under a minute, reimbursement issued) or drops the claim into a queue for a human to handle. The human's resolution gets written into a vector store the agent reads on the next pass, so the system grows smarter every day without engineering touching it.
What's quietly radical here isn't the AI. It's the architecture. The agent isn't trying to be right. It's trying to know when it's right and behave accordingly. When confidence clears the bar, it acts. When it doesn't, it hands off cleanly. The member doesn't know the difference, and over time the bar moves up on its own as the knowledge base fills.
This is what separates a copilot deployment from a coworker deployment, and it's a pattern we now look for in every Applied AI investment we evaluate. Without confidence-thresholded deferral, an agent is a demo. With it, an agent is a scalable unit of labor.
The plumbing is unglamorous and largely familiar from any well-run enterprise stack: facade APIs that keep each agent doing one thing well, event-driven service buses for handoffs, vector knowledge bases that capture what humans decide, and observability tuned for behavioral drift, not just latency. None of this is novel. What's novel is that 2026 is the year enterprises actually wire it together.
Source: DataPower Capital Research.
Two: Bottom-Up Beats Top-Down
The most interesting talent data point we've seen this year didn't come from a CIO survey. It came from the operations side of one of the firms we work with: roughly nine in ten of the people building useful agents on their teams aren't developers.
They're operations leads, marketers, finance analysts, customer service supervisors. They're using tools like n8n, make.com, and MindStudio to wire together small agents that take two or three weeks of work down to an afternoon. Some of these agents are crude. Most are saving real money. None required an engineering ticket.
This is the inversion that matters. For thirty years, automation was gated by the engineering bench. RPA, the previous great hope, sat largely in finance for a decade because nobody outside finance could deploy it. Agentic AI is the first wave where the people closest to the work can build the tools themselves, the same week they hear about them, on platforms designed for non-technical users.
It also reframes what hiring should look like. Not every team member needs to build agents. The math works as long as a critical mass can. Five operators who can ship a useful agent in an afternoon will quietly outproduce a team of fifteen who can't, and the gap compounds.
Five operators who can ship a useful agent in an afternoon will quietly outproduce a team of fifteen who can't, and the gap compounds.
Here's where we'd push back on consensus, and where we think a meaningful portion of agentic AI capital is currently mispriced. The reflexive assumption is that the buyer for agent infrastructure is the CTO, because that's how every developer-tools wave priced itself for the last decade. That assumption is now wrong. The buyer is increasingly the head of operations, the VP of customer experience, or the CFO's chief of staff. Products that ship with per-seat pricing pegged to engineering headcount, distribution motions that route through DevRel, and onboarding flows that assume CLI fluency are systematically undershooting the people actually deploying agents in production. We expect the absorption layer (no-code agent builders, vertical workflow products, governance and deployment tooling priced for the line of business) to outperform consensus through 2026 and 2027.
Three: The Productive Failure Curve
A figure that gets repeated a lot is that only about a fifth of AI projects make it to production. The implied verdict is failure. We don't read it that way.
Every meaningful enterprise technology shift has produced the same shape on the way up. The early PC era featured machines costing twenty percent of an annual salary, shared across teams, justified by spreadsheets and word processors that nobody had needed two years earlier. RPA produced a decade of expensive pilots before it found a horizontal beyond finance. The shape is a J-curve, and the projects on the descending part of the curve aren't waste. They're tuition.
What separates the projects that climb back up is becoming legible to us across our portfolio. Four things show up reliably.
First, expectations get set wrong on quality. A team commits to ninety-nine percent accuracy on a workflow the model can deliver eighty-five on, and when the production numbers come in, the project loses the political capital it needs to keep iterating. Better to ship at eighty-five with a clean handoff to a human queue, and let the deferral architecture do the rest.
Second, there's no deterministic perimeter around the probabilistic core. Pre- and post-processing, contract-tested APIs, explicit schema validation. None of it is sexy. All of it is what gives a risk officer something to sign.
Third, drift goes unmonitored. Models receive silent updates and production behavior shifts underneath the team. Catching this requires daily evaluation against a held-out reference set and an alerting layer most enterprises haven't staffed for. This is data-science work, not engineering work, and the org chart usually doesn't have the headcount.
Fourth, projects get built top-down without bottom-up signal. The strategy deck calls for agentic transformation, the engineering team builds something, and nobody in operations was in the room to say what would actually save them time. The result ships and dies on the vine.
Source: DataPower Capital Research.
Four: Cybersecurity Becomes Symmetric
The risk story we're watching most carefully in 2026 isn't hallucination. It's that agents now write agents.
The offensive side has changed shape. A capable adversary no longer needs to hand-craft exploitation. They can point an agent at a target, let it crawl for vulnerabilities, hit barriers, write its own follow-on programs to get past those barriers, and report back only when something useful turns up. The traditional defensive posture, which combined passive perimeter tools with periodic human-led pen testing, was already strained against well-resourced threat actors. Against autonomous offense, it's structurally outmatched.
The defensive response has to be symmetric. Agentic security stitches together monitoring, active testing, and dynamic defense generation into a single autonomous loop. Threats get identified, defenses get authored, perimeters reshape themselves in real time. This isn't an incremental category. It's a category reset, and we expect public market security comparables to reflect a re-rating over the next six to eighteen months.
For enterprise risk officers reading this, three things are non-negotiable in 2026. Programmatic kill switches on every production agent. Observability instrumented at the agent boundary, not just the model boundary. Gated approvals on any irreversible action. And one quieter risk worth keeping in front of the board: enterprises will cross the line from "agent that suggests" to "agent that acts" without realizing they have done so. The accountability boundary moves a feedback loop at a time, and almost nobody is auditing the moment it moves.
The accountability boundary moves a feedback loop at a time, and almost nobody is auditing the moment it moves.
Five: Capability Was Never Going to Be the Constraint
The most cited timeline revision in the AI discourse this quarter walks back an earlier prediction that the next twelve to twenty-four months would automate roughly half of entry-level white-collar and blue-collar work. The new framing stretches that horizon to one to five years and brings back a vocabulary of patience that had largely vanished from the discourse.
We read the revision as honest. The people closest to frontier capability are now saying out loud what enterprise practitioners have been telling us in private for two years: capability isn't the bottleneck. Absorption is. Governance, change management, trust, the human work of teaching organizations how to use the tools without breaking on them.
Capability isn't the bottleneck. Absorption is.
That diagnosis is what's shaping how we allocate. The investable surface in 2026 isn't whether agents work, that question is settled. It's where absorption compounds: vertical agents in regulated industries, where getting confidence thresholding right is expensive and the moat builds with every human-validated decision; inference optimization that bends the unit economics of agent-heavy workloads; and the intersections where agentic AI becomes the operating layer for physical-economy products.
Source: DataPower Capital Research. Snapshot of Q1 2026.
What This Means for DataPower Capital
Our positioning across the three pillars of the firm follows directly. In Applied AI, we're concentrated in companies whose product surface embeds the deferral architecture and is priced for non-engineering buyers. In Inference Infrastructure, we're positioned around the unit-economics shift, with voice optimization producing some of the cleanest demos we're seeing this cycle. In Frontier Tech, the agentic operating layer is the connective tissue between robotics, defense and space, and the physical economy, and the most interesting deals of 2026 will live at the intersections of those categories.
For our LPs, the through-line is simple. 2026 is the year capital starts being rewarded for picking application over experimentation, deferral over autonomy theater, and absorption-aware founders over capability-only ones.
For founders, the call is shorter still. Apply the AI.
This note draws on DataPower Capital's Q1 2026 portfolio diligence, LP conversations, and a recent Tech in Motion panel on Agentic AI in Tech hosted by General Partner David Yakobovitch. All views are DataPower Capital's.
For investor inquiries: ir@datapower.vc | datapower.vc
