At 9 a.m., the night’s work is already in the queue
At 9 a.m., a data product owner opens her review queue. Overnight, agents enriched the requirements of a new data product, checked contracts and classifications, ran tests and prepared a decision context for the business. Every run carries its receipts: what was done, by which agent, under which authorization, with which result. Nothing was published; exceptions are flagged with evidence attached. During the day the team reviews outputs, resolves ambiguities and refines the limits within which agents operate.
McKinsey reports that organizations that redesigned software delivery around agents already see productivity gains of three to five times, with teams about 60% smaller. The point is the cause, not the number: gains do not come from deploying agents, but from reconfiguring the operating model so that people and agents collaborate around the clock. Preparing data, qualifying products, composing context, deciding, acting and reconciling results is the same kind of work — human intent translated into execution, with judgment concentrated where it is indispensable.
The other side of the number
The gains are real — and rare. Deloitte observes that access to sanctioned AI tools rose about 50% in a year while use in the daily workflow stalled. Widely cited readings point the same way: 95% of generative AI pilots without measurable return (MIT NANDA, 2025), more than 80% of AI projects failing (RAND, estimate), about 6% of companies extracting real profit impact (McKinsey State of AI, 2025). These are external research readings, not Arcogi measurements.
The cause sits in the budget, not the technology: roughly 93% of AI spend goes to the tool and about 7% to people, processes and the redesign of work — the line that turns purchased capability into installed capability. The Klarna case is a cautionary illustration: measuring speed and average cost without redesigning the service is a shortcut, not a strategy. The honest conclusion is not to spend less on technology, but to stop spending only on it.
1. Design the operation for continuous flow
Most organizations still run at the pace of meetings. In the agentic model the rhythm inverts: during the human shift, the team reviews outputs, resolves what needs judgment, strengthens operating limits and prepares structured inputs; during the agents’ shift, executable work happens at scale — enriching requirements, validating architecture and contracts, generating and testing artifacts, packaging results with evidence for review.
This cycle depends on practical foundations: a clear definition of what must be produced, a consistent environment with modular conventions, a standardized path from intent to execution, and the same stakeholders engaged throughout. Without that consistency, agent output is fragmented and hard to trust.
2. Extend automation to remove manual handoffs
Traditional automation focused on the end of the pipeline; most of the effort, from requirements to decisions, stayed manual. When requirements, policies, contracts and specifications exist in standardized, machine-readable formats, the pipeline runs end to end in hours and people intervene at defined review gates — not as intermediaries.
In this design, engineering details become governance mechanisms. An execution plan can be bound by a hash to the authorization that approved it: change any field and the earlier authorization no longer holds. Every relevant change produces a receipt; a replay does not re-execute effects; publishing requires review by a person other than the submitter. Segregation of duties stops being a policy in a PDF and becomes an invariant on the execution path.
3. Build knowledge infrastructure to unlock autonomy
Agents produce correct results when they have context and organizational memory. Today that context is scattered across tools, tickets, incident reports, summarized policies and the tacit knowledge of people who have been in the operation for years. Organizations at the frontier are building a connected knowledge layer that relates data products, past decisions, policies, evidence and owners within each domain, in a form agents can navigate — and answer with sources.
Every decision becomes traceable to its origin. This does not start with a large top-down ontology effort: the graph grows organically around priority domains and live programs. At scale, knowledge stops being static documentation and becomes production infrastructure.
4. Capture value: redesign roles, authority and evidence
Productivity only becomes value when the organization changes with it. That takes reskilling people for judgment, review and supervision; bringing risk, legal, compliance and procurement into the design instead of leaving them as gatekeepers at the end, with policy as code and multi-step approvals built into the flow; and consciously reinvesting the capacity that is freed up.
There is a fourth, decisive element: the discipline of truth. Where agents execute, trust is demonstrated, not declared. The system must keep the distinction between what was specified, simulated, implemented, tested and approved for operation. An operating model that cannot say, at every moment, what is real and what is intent does not deserve autonomy.
The business model changes too: services as software
When the operating model is codified — practices, contracts, gates, evidence — the service that delivers it stops being a project and becomes a continuous capability. We call it services as software: not a volume of expert hours, but a codified practice in an operating environment that stays with the client. Knowledge no longer leaves with the team; decisions, criteria, exceptions and evidence accumulate value with each cycle.
Delivery is modular and sovereign: the operations center can be deployed in parts in the client’s own environment, starting with organization and agents and adding data and governed decisions in the same workspace. Authority does not transfer — configurations, approvals and mandates stay under the organization’s control. The economics are described qualitatively; concrete gains depend on each client’s scope, program and conditions.
Where to start
Start where the impact is greatest, with a small first move by design: one prioritized domain, one live program and one named owner. First organize who does the work — roles, agents, knowledge and limits of authority — and bring data and decisions into the same environment as trust is demonstrated. Scale comes from repeating verifiable cycles, not from a big-bang rollout.
As agents take on execution at scale, human roles concentrate on architecture, product judgment and systems design, and institutional knowledge and technical coherence become the decisive differentiators. Organizations that build these capabilities as part of reconfiguring their operating model will not just move faster — they will redefine how data creates value.