Tuesday, November 10, 2026

Enterprise Network Modernization, Part 2: Where to Find Evidence That Isn't Just Slides

In Part 1 I argued for drawing your own map before buying anything. In short: the traffic lives in four domains (data center, campus and branch, WAN and secure access, multi-cloud and hybrid). Three planes govern it (operations and intelligence, security and trust, consumption and sourcing). A physical foundation sits under both. That makes 24 topics, each with exactly one home, decided by who owns the buying decision.

This part is about the question that bothers me even more: once the map is on the wall, where do you get the evidence to decide anything?

The conversation usually jumps straight to "so what do the analysts say?", and that quickly shrinks to looking at the top right corner of a chart. I understand the instinct. It is also how a lot of expensive decisions get made.

Not all evidence is equal

Roughly, from weakest to strongest in terms of relevance to your environment:

  1. Vendor material.
  2. Analyst reports.
  3. Peer conversations and references.
  4. Independent or lab testing.
  5. Your own data and your own proof of concept.

Breadth runs the opposite way. Analysts see the whole market and your data sees only you. So you need both, and you need to know which question each one can answer. Analysts can tell you where the market is heading and who the players are. They cannot tell you whether it fits your network.

Reading analyst output without being fooled

A few habits that have saved me time.

Read the text, not just the picture. The quadrant or the wave is a summary of a long document, and the interesting parts are usually the strengths, cautions and inclusion criteria. Check the date and the scope. A vendor can be a "Leader" in a market defined in a way that has nothing to do with what you want to buy. Remember what the chart does not measure: your fit. And remember who often pays for the copy you are reading. A lot of full reports are paywalled, and what is public is frequently a vendor's reprint or press release. I treat a press release as what it is.

Different firms also do different jobs. Gartner, IDC and Forrester are strongest at defining categories and comparing vendors across them, and they give you breadth. Gigamon, KuppingerCole, Frost and Sullivan and similar firms often go narrower and deeper into features and approach. I find it useful to ask how deep any piece of analysis actually goes. Roughly: does it stop at logos, at checkboxes, at features, at approach, at architecture, or does it get to what the architecture means for you? Most public material stops somewhere around the third step. The value is in the last two, and that is where your own architects and your field experience have to fill in.

And label forecasts as forecasts. Take a line like Gartner's earlier prediction, as reported in the trade press, that by 2026 sixty percent of new SD-WAN purchases would be part of a single-vendor SASE offering. That is a forecast, made at a point in time, by someone with a view. Fine to read, risky to quote as adoption. Always ask who counted, and when.

What the coverage looks like on my own map

I tagged each of the 24 topics with whatever named public analyst evaluation I could find. The pattern was clear enough to be useful. Security and operations are well covered. There is a Gartner quadrant for hybrid mesh firewall (September 2026), Forrester Waves covering SASE and zero trust platforms (both Q3 2025) and network analysis and visibility (Q4 2025), and a Gigamon Radar for network observability (version 6, April 2026). Wireless LAN has both a Gartner wired and wireless LAN quadrant (May 2026) and an IDC MarketScape (2025). Carrier WAN services and managed services have Gartner quadrants too, for global WAN services (March 2026) and managed network services (April 2026).

At the edges it thins out quickly. For edge or micro-DC, IoT and OT networking, optical transport, cloud-native networking, cloud interconnect, service mesh overlays and passive infrastructure, I did not find a named public evaluation I could point to. To be fair, I may have missed some, and paywalled research may cover more than I can see.

That is not a verdict on those topics. Often it is the opposite, because these tend to be the early ones. It just means you should lean harder on field evidence and your own testing, and be more careful about confidence. A gap in analyst coverage is a flag, not a conclusion.

The routine I follow: gather, churn, synthesize

It is boring, and it works.

Gather from three lanes. Market definers for the big picture. Specialists for depth. Then the field: lessons learned, post-mortems, honest conversations with people who have done it, plus your own research. Do not skip the third lane, because it is the only one that is about reality instead of positioning.

Churn is the unglamorous middle. Put everything on the map, topic by topic. Date-stamp each finding. Mark each number as forecast or observed. Cross-check anything important against a second independent source before believing it. Write down where sources disagree and where there is nothing at all. This is where most people stop caring, and it is where most of the value is.

Synthesize into one page per topic. Market context. Adoption, dated and labelled. Considerations. Challenges. The vendor landscape, if you want to show one, on axes you define yourself and disclose, rather than something that looks like somebody else's chart. And a recommendation. Then add one more line: where does practice differ from what the analysts say? That is the field check, and it is often the most interesting sentence on the page.

Your own data is the best evidence you have

The good news is that you probably already own more of it than you think.


None of that comes from a slide. All of it is more relevant to your decision than any quadrant.

The field matters just as much, so write down your own lessons learned while they are fresh, and ask peers what went wrong, not only what went well. When you talk to vendors, ask for the reference that did not go well. I wrote some questions along those lines years ago for self-driving networks (Few Questions You Should Ask Your fav Self-Driving Networks Vendor). The technology names have changed, the questions mostly have not.

Rate the evidence, not just the technology

This is a small habit with a big effect. For every topic, rate two things separately. How attractive is the technology for us? And how confident are we in the evidence behind that? A topic can be attractive and poorly evidenced, which usually calls for a small, reversible experiment and not a big commitment. A topic can be well evidenced and unattractive, which usually means leave it alone. Keeping the two scores apart stops a persuasive demo from doing both jobs at once.

A quick worked example: quantum-safe

Take one topic and run it through the routine, quickly.

Market context: Gartner's 2026 trends for enterprise networking name quantum-safe networks, agentic NetOps and network sovereignty among the themes to watch, and its 2026 Hype Cycle points at regulatory, sovereignty and quantum pressures on network security. So the direction is not in doubt.

Adoption: here I would be careful. What I can find publicly is mostly forecasts and vendor announcements, not measured adoption. I would mark it forecast and leave it there.

Considerations: some data has a long shelf life, and some of your equipment has a longer life than your planning cycle. Whether your vendors support what comes next in firmware matters more than what they say on a roadmap slide.

Challenges: before you can protect anything, you need to know where cryptography is used, and in most organizations nobody has a complete answer.

Own data: the cryptography inventory from the table above. Field: ask peers who have started one how long it took and what surprised them.

Recommendation, and this is a judgement call: start with the inventory, and put crypto agility questions into your next network RFPs. Do not start a rip-and-replace Programme on forecasts. Confidence: reasonable on direction, low on timing.

Notice what happened there. Nobody bought anything, and we now know what to do next.

Where I would start on Monday

Draw your own version of the map, and argue about it until the boxes stop moving. Pick four to six topics, not twenty-four. For each one, find out what the market says, find out what your own data says, and write the two side by side. Mark everything as forecast or observed, with a date. Rate the evidence separately from the technology. And before any budget gets approved, ask the person presenting to show you where it failed for someone else.

None of this is clever. That is rather the point. Stay grounded and keep it real.

HTH...

A Tech Artist 🎨

Sources and further reading

Full reports are largely paywalled. What follows is what I could verify publicly, by title and date, including vendor reprints and press coverage where that was all that was visible.

Posted by at No comments:
Labels: , , , ,

The Agreement Trap: Fixing Group Paralysis in the AI-Era Operating Model




Most enterprises do not have an AI technology problem. They have a collective-action problem, and AI makes it more expensive.

Only 5% of companies capture AI value at scale, while 60% report minimal gains despite substantial spend (BCG, 2025). Just 39% see any enterprise-level EBIT impact from AI (McKinsey, 2025). The models are not the bottleneck. The way groups of people decide is.

Three classic pieces of organizational research explain this better than any technology roadmap I have seen. In this post I connect them, turn them into a simple framework, and tie it back to the operating model every CXO is now redesigning for AI.

Executive summary

  • Three theories explain the gap. Smith and Berg show why groups freeze. Harvey shows why they agree on what nobody wants. Pfeffer and Sutton show why they know what to do but do not do it.
  • Together they form one failure chain. A group can fail to decide (Stuck), decide wrongly through false consensus (Silent), or decide rightly and not act (Stalled). There is no fourth way to miss the right action.
  • AI raises the cost of each failure. Pilots approved by polite consensus, agent programmes nobody owns, and value nobody measures are the same three failures at machine speed.
  • The fix sits in the operating model. Five levers matter: decision rights, team design, workflow, metrics and incentives, and leadership behavior.
  • CXOs can start in 90 days. Diagnose the top 10 AI decisions, install dissent and decision-latency mechanisms, and move execution authority to outcome-owning human-agent teams.

Three diagnoses, one failure chain

1. Stuck: why groups freeze (Smith and Berg)

Kenwyn Smith and David Berg's Paradoxes of Group Life (1987) argues that group life is made of opposing pulls that cannot be resolved. They group 12 paradoxes into three clusters:

  • Belonging: identity, involvement, individuality and boundaries.
  • Engaging: disclosure, trust, intimacy and regression.
  • Speaking: authority, dependency, creativity and courage.

Their companion paper in Human Relations makes the key point: these conflicts are "released" rather than resolved once a group sees them as essential to its life. A group that treats a paradox as a problem to settle, picking one side to win, oscillates or freezes.

Where it fails: before the decision. No decision is made.

2. Silent: why groups agree on what nobody wants (Harvey)

Jerry Harvey's The Abilene Paradox tells the story of a family that makes a hot, 53-mile drive to Abilene that nobody wanted. Each member assumed the others wanted to go. Harvey's insight is that the hard organizational problem is managing agreement, not conflict.

He describes a recognizable pattern:

  • Members privately agree on the problem and on the fix.
  • They fail to say so, and send misleading signals instead.
  • The group acts against what its members actually want.
  • Frustration, blame and subgroups follow, and the cycle repeats.

The deepest driver is fear of separation from the group. Mark Murphy in Forbes recently framed the same issue as a team-design problem: silence gets read as agreement.

Where it fails: at the decision. The wrong decision is made.

3. Stalled: why knowing is not doing (Pfeffer and Sutton)

Jeffrey Pfeffer and Robert Sutton's The Knowing-Doing Gap (2000) shows that firms usually know what to do. Five things stop them doing it (chapter summary):

  • Talk substitutes for action.
  • Memory substitutes for thinking.
  • Fear prevents acting on knowledge.
  • Measurement obstructs good judgment.
  • Internal competition turns colleagues into rivals.

Stanford GSB's excerpt stresses the last one: zero-sum internal contests destroy the cooperation that interdependent work needs.

Where it fails: after the decision. The right decision is not executed.

Why AI raises the stakes

AI adoption is broad, but value is narrow. Every major 2025 study finds a small leading group pulling away from a large stalled majority.



The three failure modes map directly onto these numbers:

  • Stuck shows up as decision latency. AI shifts authority, identity and dependency all at once. Those are exactly Smith and Berg's paradoxes of speaking and belonging. When leaders are unsure whether a human or an agent owns an outcome, they default to committees. Gartner expects at least 15% of day-to-day work decisions to be made autonomously by 2028 (W.Media). Decision rights written for a human-only firm will not hold.
  • Silent shows up as pilot theatre. Many AI programmes get approved because nobody wants to look like the sceptic in the room. The approvals look like consensus. The cancellations that follow are the drive home from Abilene.
  • Stalled shows up as the value gap. Most firms know AI needs workflow redesign, yet McKinsey's high performers are almost three times more likely to actually do it. MIT NANDA blames a "learning gap" and brittle integration, not model quality. That is the knowing-doing gap in its purest form.

The human signal is moving the wrong way. Manager engagement fell from 30% to 27% in 2024 (Gallup). Positive views of GenAI rise from 15% to 55% where leadership support is strong, yet only about a quarter of frontline staff say they get that support (BCG). Managers are the layer through which all three failures flow.

My framework: the Movement Stack

A group misses the right action in only three ways, and each is fixed through the same five operating-model levers. The failure modes are mutually exclusive by symptom: no decision, a wrong decision, or a right decision left undone. Together they cover the whole path from tension to action.

 


The five levers are the parts of an operating model that shape how groups decide and act:

  1. Decision rights: who decides, and by when.
  2. Team design: how humans and AI agents share the work.
  3. Workflow: how work moves end to end.
  4. Metrics and incentives: what gets counted and paid.
  5. Leadership behavior: what leaders model every day.

Diagnose first, then apply the lever. Adding more dissent, a Silent fix, to a Stuck team only adds paralysis.

If you are Stuck: create decisiveness

  • Decision rights: one named owner per decision, with a decide-by date.
  • Team design: define what humans own above the loop and what agents do inside it.
  • Workflow: sort decisions into reversible and irreversible, and give the reversible ones a fast lane.
  • Metrics and incentives: track decision latency as a board KPI.
  • Leadership behavior: name the paradox out loud and hold both sides instead of forcing a choice.

If you are Silent: create real voice

  • Decision rights: collect independent written positions before any group discussion.
  • Team design: rotate a formal challenger role, and add an AI red-team agent that is told to disagree.
  • Workflow: run a pre-mortem and set kill criteria before any AI investment is approved.
  • Metrics and incentives: measure voice, such as recorded dissent and "my opinion counts" scores.
  • Leadership behavior: the most senior person speaks last and rewards bad news.

If you are Stalled: create motion

  • Decision rights: execution authority sits with the team that owns the outcome.
  • Team design: small, cross-functional human-agent teams replace functional committees.
  • Workflow: redesign end to end; every pilot scales or stops within 90 days.
  • Metrics and incentives: a few P&L-linked value measures, and no rankings that pit teams against each other.
  • Leadership behavior: treat reasonable failure as learning, and build the system rather than making every decision.

This is consistent with where operating models are heading anyway. McKinsey's work on the agentic organization describes small, outcome-focused teams of humans who supervise AI workflows, with governance that is real time rather than periodic. Those teams will only work if the three failure modes are designed out from the start.

What CXOs should do in the next 90 days

Treat collective action as an operating-model workstream with an owner, a baseline and a deadline, not as a culture campaign.

Days 0 to 30: Diagnose

  • Take the 10 largest open AI decisions and investments, and classify each as Stuck, Silent, Stalled or healthy.
  • Baseline three numbers: decision latency, pilot-to-production rate, and share of AI spend with a tracked P&L outcome.
  • Run Harvey's test anonymously with each leadership team: would members privately reverse a decision they publicly backed?

Days 31 to 60: Install the mechanisms

  • Give every AI decision a single owner and a decide-by date.
  • Make pre-mortems and kill criteria a gate for AI investments above an agreed threshold.
  • Collect written positions before every steering committee, and have the most senior person speak last.
  • Redraw decision rights for three end-to-end workflows, stating where humans sit in the loop, above it, or out of it.

Days 61 to 90: Move authority and money

  • Hand two or three workflows to small human-agent teams that own the outcome end to end.
  • Shift part of leadership incentives from functional KPIs to shared value measures.
  • Scale or stop every pilot older than 90 days.
  • Report the three baseline numbers to the board quarterly, next to the financials.

The test of success is simple: fewer open decisions, more pilots scaled or stopped, and dissent recorded before decisions rather than blame after them.

Limits of this argument

I stress-tested this framework the way I would review a colleague's work. A few honest caveats:

  • Scope. The framework does not cover honest error, where genuine debate still picks the wrong option. Nor does it cover drift, where a good decision goes stale because nobody revisits it. A fourth mode, failure to learn after acting, may be worth adding.
  • Evidence quality. The BCG and McKinsey figures come from large but self-reported surveys, and the link between workflow redesign and performance is correlation, not proof of cause. The MIT 95% figure is widely quoted but rests on a small sample, so I treat it as directional. The decision-making and psychological-safety data predate GenAI.
  • Regional fit. All three theories come from US research. In the hierarchical cultures common across APAC, I suspect Silent is the dominant failure mode, and "the leader speaks last" on its own will not unlock voice.
  • AI can make Silent worse. AI assistants tend to agree with whoever is asking. A red-team agent only helps if it is explicitly instructed to disagree and its objections are logged.

Smith and Berg would also warn against using this framework as an either/or tool. The aim is not to maximize dissent or speed. It is to hold speed and voice, autonomy and control, in useful tension.

HTH...

A Tech Artist 🎨

Sources

 

Posted by at No comments:
Labels: , ,
Subscribe to: Posts (Atom)