The Agentic Productivity Paradox: Intelligence Is Scaling Faster Than Accountability

AI agents are getting better at doing work. That part of the story is no longer controversial. What receives less attention is the gap that opens when the capacity to execute grows faster than an organisation's ability to say who authorised the work, who accepted it and who answers for it when it goes wrong. Call it the agentic productivity paradox: more intelligence, applied faster, does not automatically produce proportionally more organisational value.

What the investors are observing

In its Atlas newsletter "Breaking through the ceiling on agentic productivity" (16 October 2026), Bessemer Venture Partners offers an illustrative framing: an individual working with agents might become something like ten times more productive, while the organisation around that individual captures a gain of less than 50%. This is Bessemer's illustrative framing, not a measured result or a general law, and it should be read that way. It is useful because it points to where value may leak: less in the model than in the handoffs, approvals and checks around it.

The 2026 Healthcare AI ROI Scorecard from Bessemer Venture Partners and Bain & Company (published 30 September 2026) gives more grounded data, though still data to read carefully. It draws on 226 surveyed healthcare executives across 65 use cases. Respondents reported average ROI of 3.5x, and reported reaching returns in roughly 12 months against an expected 24. Provider revenue cycle management stood out, with reported ROI of 4.0x and 67% of its solutions described as semi- or fully autonomous agents, compared with around 4% for provider clinical use cases. The scorecard also reports that 42% of organisations have completed or are actively running AI vendor consolidation. This describes reported activity, not a prediction.

Two distinctions matter. These are self-reported survey findings: they describe what executives believe and report, not independently audited outcomes. And anything in either publication about where adoption is heading is prediction rather than finding. Both are valuable signals; neither is proof.

Why the back office may move first

The gap between revenue cycle and clinical autonomy is the most revealing detail, but the scorecard reports the pattern, not its cause. What follows is my interpretation.

Billing, coding and scheduling are not simpler than clinical work in every respect, but they tend to be more accountable by design. Inputs are structured, rules are written down, many errors can be corrected after the fact, and there is usually a clear owner for the result. Clinical decisions carry personal professional liability, regulatory scrutiny and outcomes that cannot be undone. Risk, regulation, data quality and vendor maturity may also contribute.

If that reading is right, autonomy advances where accountability is already legible. That would suggest the ceiling on agentic productivity is set less by model capability than by whether an organisation can answer five questions for every delegated action.

The five conditions for trustworthy autonomy

Identity. Which agent acted, on whose behalf, running which configuration? When an agent operates under a human's credentials, its actions are hard to tell apart from that person's in logs and approvals, which weakens attribution and least-privilege control.

Delegated authority. What was the agent permitted to do, within which limits, and who granted that permission? Authority needs scope, expiry and a way to revoke it.

Evidence. What did the agent see, decide and produce? Without a durable record, every review becomes reconstruction.

Acceptance. Who, or which rule, confirmed the output was fit for purpose before it took effect? Acceptance is the moment work becomes the organisation's work.

Accountability. When the outcome is wrong, which named person or role owns the consequence and the correction?

My working hypothesis is that organisations are more willing to let agents run where all five are clear, and that where any one is vague, work drops back to human speed. That would be one plausible explanation for the leakage Bessemer's framing describes, not a tested one.

What existing tools already cover

Much of the required control already exists. Identity and access management platforms handle authentication, roles and non-human identities. Workflow tools encode approvals and routing. E-signature platforms provide recognised forms of acceptance. Audit logging and governance, risk and compliance systems capture evidence and map it to controls.

For much of the work inside a single organisation, these tools — configured deliberately and extended to treat agents as distinct actors — may well be enough. For many leaders the first step is not a new platform but an inventory: which agent actions today run on shared human credentials, unlogged approvals or implicit acceptance?

Where the control plane runs out

Consider agents acting across organisational boundaries: a purchasing agent negotiating with a supplier's sales agent; a provider's documentation agent exchanging records with a payer's adjudication agent. Each side may have strong internal controls, yet no shared answer to basic questions. Is this counterparty agent who it claims to be? Was it authorised to make this commitment? Whose record is authoritative? When was the result accepted, and by which party?

Today these gaps are usually bridged by contracts, human sign-off and reconciliation after the fact. That works, but it brings human speed back at the boundary. Whether a coherent cross-organisation control plane emerges — through standards, identity federation extended to agents, or new intermediaries — is an open question. Specific approaches may address parts of it; that is not evidence of a universally adopted shared control plane. Leaders should examine the scope and evidence behind any claim of a complete solution.

The investor and operator lens

For investors, the paradox reframes where durable value may sit: not only in more capable agents, but in the infrastructure that lets organisations safely accept their output. For operators, it reframes the ROI conversation. If gains stall between the individual and the organisation, the constraint may well be governance design rather than model choice. Spending on more capability before settling authority and acceptance risks buying speed the organisation cannot absorb.

Three questions for executives

1. For the agent actions already running in your organisation, can you name — for each — the identity, the granted authority, the evidence trail, the acceptance point and the accountable owner?
2. Where are individual productivity gains failing to reach the organisation, and is the blocker capability or an unanswered accountability question?
3. Which planned agent use cases cross an organisational boundary, and what currently stands in for a shared control plane there?

Measured next steps

• Inventory agent actions and the credentials they use.
• Extend existing identity, workflow and audit tooling to treat agents as distinct, scoped actors before buying new platforms.
• Pilot autonomy first where accountability is already legible, as the healthcare revenue-cycle pattern may suggest.
• Define explicit acceptance criteria before granting broader authority.
• Treat cross-organisation agent work as something to watch and pilot cautiously, with contractual and human safeguards, rather than something to scale on assumption.

Intelligence will keep scaling. The organisations that capture its value are likely to be those that let accountability scale with it.

The next competitive advantage in artificial intelligence will not necessarily belong to those who automate the most. It may belong to those who can demonstrate that every consequential action was authorized, every outcome can be verified, and accountability remains intact. — Ronny Boesing