Agents are real: The scaling question is what to solve for

The question about enterprise AI agents quietly changed over the past six months. The debate used to be whether agents were ready for production. Today, 51% of companies have them in production, 78% plan to deploy them, and 90% of nontech companies are on the same timeline as tech. Seventy-nine percent of executives tell PwC they are already rolling them out.

So agents are real. What’s more interesting is what the same data says next.

Performance quality — not cost, not safety — is the single largest barrier to deployment, cited more than twice as often as either. Forrester puts a sharper number on it: 3 in 4 enterprises that try to build their own agentic architecture will fail.

The era of "Can it be done?" is over. The era of "Can it be done reliably, on content we can't afford to get wrong?" has started.

This is the transition shaping every serious enterprise AI conversation in 2026. It is why the next phase of content intelligence has to be engineered around governed outcomes and scaling autonomy, not impressive demos.

Supervision, not human-in-the-loop

In 2023, a New York lawyer filed a motion citing six legal precedents that did not exist. He had asked ChatGPT to draft the brief. Then, when opposing counsel flagged the missing references, he asked ChatGPT to verify them — and the model confirmed, in sequence, that the fictional cases were real. The judge called the resulting analysis "gibberish" and sanctioned the firm $5,000.

The problem was not the model. The only supervision layer between the model and the courtroom was a human reading the same output for plausibility.

This is the scaling problem enterprise AI is running into. The dominant production pattern today is orchestrator-worker agents with a human approving significant actions. That is a good starting point but it is not a scaling answer. If every consequential decision needs a human to review, reason and decide again, productivity stays close to the cost of doing the work yourself. Forrester's Advisory-AI ceiling of 10-50% productivity gains is real, and it's low.

The next move is supervision. Not removing humans but changing what they do: Observe agent decisions at scale, track outcomes and capture human feedback. Use the accumulated decision trace to figure out, per situation, how much experience the agent actually has and how well it has performed. Then use that metric to gate autonomy — agents earn the right to act more autonomously where they have demonstrated success, and autonomy gets pulled back where outcomes slip. Humans stay in the loop for validation, calibration and exception. Not for every action.

This is the architectural core of what we are building toward at Hyland: a governed spectrum where confidence and outcome thresholds unlock autonomy per workflow. Governance does not mean a human in front of every decision. It means a system that measures, audits and adjusts autonomy continuously.

The moat is context, engineered on demand

In November 2021, Zillow shut down its iBuying arm after a single quarter of $420 million in losses and laid off 25% of its workforce. The algorithm that priced home offers had extensive data about houses but lacked situational context. In this case, it’s the ability to recognise when market dynamics had shifted outside the conditions it was calibrated for. Knowing a lot about houses is not the same as understanding the market those houses trade in.

Much of the 2026 agentic conversation still treats content as the substrate: Retrieve the right document, put it in the prompt, let the model answer. That already underestimates what enterprise workflows need. The value is not in content as isolated artefacts; it is in unstructured data plus the context that comes from relating content, enriching it and composing it into the situation-specific skill a decision needs.

I think of this as engineered just-in-time skills. The agent does not arrive at a decision with a generic capability. It arrives equipped with a skill specifically composed from its memory of past decisions, the outcomes of those decisions and the human teachings that corrected or confirmed them. Similarity retrieval decides which memories and teachings apply here. The skill is temporary, traceable and auditable.

The compounding effect is the point. The longer a content-aware agent runs in a workflow, the better its engineered-skill composition becomes. The longer the workflow runs, the better the agent gets. That is the moat.

The agent-content interface is still immature

How agents access and act on enterprise content is the biggest unresolved engineering question of the next twelve months. Wrapping REST APIs in tool descriptions and hoping the agent figures it out works for demos. It does not scale to production workloads where token cost, latency and serialised correctness matter.

Three shifts are reshaping this surface.

1. Model context protocol (MCP) has become the defacto standard for agent-to-tool integration — Claude, ChatGPT, Copilot, Cursor and a long tail of developer tools all support it. Every serious enterprise AI platform now ships an MCP client, an MCP server or both.

2. Agent2Agent (A2A) has emerged as the complementary standard for agent-to-agent communication. Launched by Google in April 2025 and brought under the Linux Foundation that June, v1.0 shipped in March 2026 with 150+ supporting organisations including every hyperscaler. A2A and MCP are not competing — they sit at different layers. Enterprises are being told to adopt both.

3. CLI as an agent interface. Newer, and promising: fewer tokens per interaction, faster execution, serialisable composable work. An agent scripting CLI commands over a content system behaves a lot more like a capable human operator than JSON-over-HTTP ever did.

Beyond transport there is retrieval. Classic REST assumes the caller knows what to ask for; agents often don't. They need semantic search, knowledge graphs, context retrieval and agentic retrieval that reason over the content graph instead of fetching by ID. ECM has to evolve across all four of these surfaces — not just bolt MCP onto yesterday's APIs.

Unstructured data is where agents earn their keep

The top three production use cases for enterprise agents are research and summarisation (58%), personal productivity (53.5%), and customer service (45.8%). All three are anchored in unstructured data — documents, images, recordings, transcripts, conversations. The flows that matter — claims, contracts, compliance filings, clinical records, product catalogs, inspection evidence — are unstructured at their core.

Vertical use cases don't scale on horizontal AI. They scale on decades of understanding how unstructured content resolves specific workflow outcomes in specific industries, engineered into the AI layer. A generic platform can orchestrate a workflow but it can't tell you that a particular clause in a reinsurance treaty changes the loss-adjustment obligation, or that a specific frame in an inspection video indicates pre-existing damage rather than new.

The platforms that earn a seat at the table in 2027 will be the ones that bring industry-specific content fluency, not the ones with the best generic model integration.

Open is becoming a feature request

Two quiet shifts are converging.

First: Hybrid and on-premises AI is growing in absolute terms, not shrinking. IDC's figures put on-premises and hybrid AI platform spend at $29.3B by 2028, a 19.3% CAGR. Sixty-five percent of enterprises will adopt hybrid edge-cloud inferencing by 2026. For content governance on classified, PHI or PII data, cloud-only is a structural sales blocker, not a preference that pricing or trust signals can overcome.

Second: Enterprises now expect the AI layer on top of their content to be open, inspectable and not locked to a single cloud or a single vendor's model. Open standards like MCP and A2A are part of that. Open source is another. In August 2025 Hyland open-sourced its AI-ready Cloud Content Repository — the underlying content substrate — and contributed it to the Hyland Alfresco™ Community. In procurement, openness increasingly reads as a compliance statement, not a commercial preference.

Two signals to watch

Agent memory as a separable layer. A new category of memory infrastructure is emerging — vendors treating long-term agent memory as a platform concern rather than something every agent implementation reinvents. For content intelligence, this is directly on-thesis: Memory,

similarity retrieval and human teachings are the load-bearing elements that compose engineered just-in-time skills. The question is whether enterprises adopt standalone memory layers or prefer ones tied to the content-and-decision substrate.

Agent Passport for evals-gated production deployment. If quality is the primary barrier, evaluation becomes load-bearing infrastructure rather than a QA hygiene task. The pattern I'm most interested in is a certified, machine-readable record carried by every deployed agent — declaring its identity, skills, autonomy bounds, data-access scope and compliance posture — and required before the agent can run in production. It composes naturally with supervision: The passport declares the maximum autonomy; supervision measures actual performance and tightens the ceiling when results slip. Without something like this, every enterprise AI deployment re-solves the same governance problem agent by agent.

Your agentic future

Enterprises have gotten past "Can agents work." They are now on "Can agents work reliably, and at scale, where a system supervises, measures and gates autonomy continuously rather than requiring a human at every step?"

The platforms that will win the next cycle are not the ones with the best model. They are the ones that solve the memory, retrieval, governance, supervision and interoperability problems around the model — on unstructured data that already lives with the enterprise.

That is the problem we are building for.