The adoption ladder enterprises keep skipping

Note on sourcing: the trigger for this piece is an InfoQ presentation page that could not be reached at the time of writing (repeated DNS failure). The figures attributed to that talk — a specific token count, a four-rung "adoption ladder" — are therefore not repeated here as fact. What follows is independently reasoned from how agent deployments are known to fail, without borrowed statistics.

A presentation on the InfoQ AI, ML and Data Engineering track argued that real-world AI adoption depends on trust rather than benchmark scores, and that the gap between the two shows up in three places: how an agent builds context before it acts, how it is routed across models, and whether anyone is watching what it does once deployed. That framing is worth taking seriously even without the underlying numbers, because the three failure points it names map directly onto problems enterprise security and compliance teams already see.

Why benchmarks and trust diverge

A model's benchmark score describes what it can do under test conditions. It says nothing about what happens when that model is wired into a workflow that calls five tools, inherits a service account's permissions, and runs unattended overnight. The gap between the two is not a model problem. It is a wiring problem, and wiring problems are invisible to whoever is only looking at leaderboard results.

Three symptoms recur across agent deployments, independent of which vendor built the model underneath:

• Context construction breaks silently. An agent that retrieves the wrong document, an outdated policy, or a record it was never meant to see does not throw an error. It answers fluently, using the wrong input, and nothing downstream flags the mismatch.
• Multi-model routing hides which model actually acted. Enterprises increasingly route requests across several models for cost or capability reasons. Unless that routing decision is logged alongside the request, an incident review has no way to establish which model produced a given output.
• Feedback loops rarely exist at all. Most agent deployments have no mechanism for a human to mark an output wrong in a way that changes future behaviour. Without that loop, a flawed pattern repeats indefinitely rather than being corrected once.

None of these are model-quality failures. They are architecture failures, and they are the reason organisations that have adopted capable models still report low trust in what their agents do unattended.

What it means for a regulated enterprise

For an organisation operating under the EU AI Act or sector rules like DORA, the practical exposure is evidentiary. If an agent acted on stale or wrong context, routed a request through an unlogged model, and received no correction afterward, the organisation cannot reconstruct what happened or show that it noticed. Record-keeping obligations assume the system can state what it did and why. A deployment with no routing log and no feedback trail cannot meet that bar, regardless of how capable the underlying model is.

The commercial exposure runs alongside the regulatory one. An agent that quietly answers from the wrong context is not a rare edge case; it is the default failure mode of retrieval-based systems that were never audited for it. Every one of those answers reaches a customer, a regulator, or a colleague before anyone reviews it, because there is no step in most pipelines where review happens before delivery.

What actually addresses it

Closing this gap does not require a better model. It requires three mechanisms that most agent stacks lack by default:

1. Context provenance. Every answer should be traceable to the specific documents or records that produced it, so a reviewer can check whether the agent reasoned from the right input rather than a stale or mismatched one. 2. Routing accountability. When a request is routed across models, the routing decision itself needs to be recorded with the request and the response, so an incident review does not have to guess which model acted. 3. A correction mechanism with memory. A wrong output needs a way to be marked wrong that actually changes what happens next time, not a support ticket that closes without touching the system.

These are logging and governance decisions, not model-selection decisions. An organisation can run the most capable model available and still have none of the three in place.

What to check on Monday

Without buying anything, a reasonable team can answer three questions about its existing agent deployments this week:

• Pick one agent that retrieves information before acting. Ask it to show its sources for its last five answers. If it cannot, context provenance does not exist for that agent today.
• Pick one workflow that routes across more than one model. Ask which model answered the last request it handled. If nobody can say without digging through application logs that were not built for this, routing is unaccountable.
• Ask whether a flagged-wrong output from last month changed anything about how the agent behaves now. If the honest answer is no, there is no feedback loop, only a complaints inbox.

None of these checks require new tooling. They require asking the agent, and the team that built it, questions that a benchmark score was never designed to answer.

How AANCER answers

AANCER treats context provenance and routing as logged facts, not implicit behaviour. Every request that is routed across models carries that routing decision into the append-only audit ledger, alongside what was retrieved and which of AANCER's 725 connectors the request touched, so a reviewer can reconstruct the chain from question to answer without reopening application code.

Model routing itself is a policy decision rather than a cost optimisation applied invisibly: AANCER's routing has measured 48 to 79 percent token savings while keeping the routing choice visible in the ledger, which is the difference between an efficiency gain and an unaccountable one. Agent access is mapped to 27 regulations so that the record-keeping question — what did this agent do, and under what authority — has a documented answer rather than a reconstruction exercise.

Where a deployment needs a correction to take effect immediately rather than at the next release cycle, revocation of an agent's access completes in under 30 seconds, which is the mechanism that turns a flagged failure into a closed one instead of a repeated one.

Sources - https://www.infoq.com/presentations/agentic-ai-developer-trust-adoption/ — trigger URL, unreachable at time of writing (DNS resolution failure on repeated attempts)