Blog · AI leadership
The model is not the bottleneck. The organization is.
A new paper in IT Revolution’s Enterprise Technology Leadership Journal asks why AI returns are not visible yet, and answers with a sociotechnical architecture. Here is what it asks of leaders, and where XALT Northstar fits, layer by layer.
A letter to the CEO
The Fall 2026 issue of IT Revolution’s Enterprise Technology Leadership Journal carries a paper by eleven technology leaders, from companies such as SoFi, IBM and Nubank: Sociotechnical Architecture of AI-Enabled Enterprises – Principles, Patterns, and Practices for Human-Governed AI. It opens as a letter to the CEO and asks the question many executive teams avoid: why are the returns from AI not visible yet?
The paper cites a 2025 MIT study of three hundred enterprise AI deployments: 95% showed no measurable effect on profit and loss. Same models, same vendors, very different results. The authors’ answer is that the difference is not the technology but the organization it lands in. The technical systems can be rewired in weeks. Decision rights, escalation paths and accountability change over years, if at all.
The idea is not new. In the late 1940s, researchers at the Tavistock Institute saw productivity fall in British coal mines after mechanization. The machines worked; the social system around them had not been redesigned. For seventy years, people have absorbed that gap with judgment, questions and escalation. Agents don’t. In the authors’ words, Humans have always absorbed ambiguity in these structures. Agents do not.
Before your next leadership meeting
Three questions from the paper
Who is accountable?
For every AI-enabled workflow: the named person, and what the AI may decide without them.
What outcome are we after?
For every significant AI deployment: the business result, and how you will know it worked.
Where are we already exposed?
Where AI is already running without approval, and what that means today.
Layer by layer
The paper’s architecture, and where Northstar fits
The paper builds four layers: a shared language, a handful of principles, the patterns of organizations that get it right, and recommendations for this quarter. On the left is what the paper asks for, summarized; on the right is what XALT Northstar already does about it.
Shared language
Name the kinds of AI, and where people stand to them
- Deterministic, generative and agentic AI need different oversight.
- The three waves and the Copilot-versus-agents comparison give the leadership team one vocabulary.
- Human-in-the-loop versus human-on-the-loop.
- The maturity audit asks where your organization stands, in seven capabilities, before anyone picks a tool.
Principles
What holds when the playbook runs out
- Humans hold accountability; AI holds capability.
- One dedicated agent per person, and nothing runs without that person’s approval.
- AI does not decide on its own about people.
- The agent suggests contacts with a reason, and sends an introduction only after explicit approval.
- Autonomy grows with verifiability.
- Every change gets a risk assessment and a short video; only low-risk changes deploy on their own.
- Shared context comes before delegation.
- The Enterprise Brain: one curated, permissioned knowledge layer, and every answer carries its source.
- Governance keeps pace with the work.
- The agent inherits its user’s permissions, revoked together, and every agent action is logged.
- The line between in-the-loop and on-the-loop is designed.
- Approval before acting is the default. Acting alone is reserved, by design, for low-risk changes.
Patterns
How organizations that get it right are changing
- Context as architecture.
- Curate once, use everywhere: the Brain feeds every agent instead of each one searching every source.
- The agentic paved path.
- From developer self-service to a DevEx platform to agents on the same paths, with principles and decisions kept in the repo.
- Ambiguity by design.
- Coming soon Your agent turns a loose request into a clear brief with the outcome, the owner and the deadline.
Recommendations
What to start this quarter
- An AI governance board that sets guardrails.
- Your job The board is yours to form. We bring role-based access, audit logs and the documentation your Legal and Risk teams ask for.
- Upskilling: from prompts to managing agents.
- Workshop From better Copilot prompts to the Agentic Coding Workshop on your own repository.
- Champions and a center of excellence.
- Workshop Start with your most modern team and make it the example: one-pagers of prompts that work, then a workshop.
- Setting goals the agents can work against.
- Coming soon Set your goals once; your agent connects them to KPIs and keeps a dashboard up to date.
The left column summarizes the paper in our words. “Coming soon” marks features we have announced but not built yet.
What software can’t do for you
The paper is clear that this work cannot be handed to the CTO alone, and it cannot be handed to a vendor either. Northstar makes the right way the easy way: approvals by default, permissions inherited from the user, every action logged, one curated source of context. Agents take whatever route the system makes easiest, so the supported path has to be the easiest one. But some layers are leadership work:
- naming the person accountable for each AI-enabled workflow,
- deciding where an agent may act alone, and revisiting that line as trust grows,
- rewarding teams for the flow of work rather than for individual output,
- and treating goals as something agents can be measured against, not as slogans.
That is where we work with you directly: in a discovery workshop that starts from your bottlenecks, not from a tool.