LLMs Can’t Fix Your Org Chart: It’s About Aligning People, Process and AI
By GAPx
At GAPx, we’ve observed this across portfolios and sectors: large language models don’t save broken operating models, they magnify them. If incentives are misaligned, decision rights are fuzzy, and processes run on goodwill and spreadsheets, you won’t get leverage from AI, you’ll get louder noise. Culture sets the rules of the game and communications teach the rules to everyone else. If you want compounding value, fix the way work flows and the way people align around that work, not just the code that runs.
The real blocker isn’t the model, it’s the operating model
Many organisations run energetic AI pilots that stall on contact with the org chart. Conway’s Law predicted this decades ago: systems mirror the communication structures of the organisations that build them. Fragmented ownership creates fragmented systems, and your AI will faithfully inherit those seams. If marketing, product, data, and risk speak different dialects, your models will too. The cure begins with structure and decision clarity, not another proof of concept.
Pilots also fail because leadership can’t see how model outputs change frontline decisions. The work is still routed through manual checkpoints, unowned datasets, and committees that meet monthly. In our experience, that gap isn’t technical at root. It’s governance and accountability: who owns the decision, who owns the data that feeds it, and who is accountable for outcomes when the model is wrong. Clear decision roles raise organisational throughput; ambiguity slows everything down. (USC Center for Effective Organizations)
There is a second trap. Teams optimise for demo metrics, not business metrics. A model that looks good in offline evaluation can die when it meets the messy realities of forecast cycles, pricing windows, and service levels. Companies that do capture value treat AI as a rewiring exercise, redesigning workflows so model outputs enter the decision at the right moment with the right human oversight. The evidence keeps pointing in the same direction: impact tracks with workflow change, not with model novelty. (McKinsey & Company)
The opportunity: an operating model that turns LLMs into leverage
A modern operating model for AI is surprisingly practical. It has clear product ownership, value-stream teams that span business and tech, human-in-the-loop controls where risk matters, and a governance layer that makes models auditable and improvable. When you put those basics in place, LLMs and predictive systems start acting like force multipliers rather than side projects. That looks like redesigned work, not just new tools. (BCG, MIT Sloan Management Review)
Risk management is part of the operating model, not an afterthought. You need explicit controls for data quality, evaluation, guardrails, and escalation paths. Mature frameworks exist: the NIST AI Risk Management Framework lays out practical processes for mapping context, measuring risks, and managing them with governance and human oversight. Treat it like a playbook for aligning model behaviour with organisational intent. (NIST Technical Series, NIST)
Decision rights remain the quiet superpower. You can reorganise around value streams, but unless someone “has the D” for the decision the model influences, you’ll default to consensus that arrives too late. Formalising decision ownership, escalation, and cadence clears the path for AI to change outcomes rather than add analysis. (Harvard Business Review)
Culture and communications are the multiplier
Great operating models fail in quiet ways when culture resists new behaviours or when people do not hear, remember, or believe the narrative for change. Treat culture as the operating system and communications as the user interface. Start with sponsor clarity: one senior owner who says the same simple why, what, when, and how every time. Turn managers into multipliers with talking points, micro-rituals, and team huddles that translate strategy into local action. Replace change theatre with behaviour change by scripting the first ten minutes of the new way of working: who does what, with which tool, at which moment in the decision. Keep a predictable cadence: monthly all-hands for direction, weekly value-stream updates for progress, and lightweight changelogs that explain what changed, why it changed, and how to use it. The culture you design in public is the execution you will get in private.
A composite example: rewiring pricing with people in the loop
Consider a portfolio retailer with regional pricing set in quarterly bursts. The company implements demand forecasting and an LLM assistant that proposes price changes with narrative rationale. The transformation doesn’t come from the assistant, it comes from the design choices around it:
Ownership: a pricing product owner owns the end-to-end decision.
Value-stream team: commercial, data science, finance, and ops sit in one pod with a weekly decision cadence.
Controls: the LLM drafts proposals; analysts validate high-impact moves and sign off where thresholds are exceeded.
Governance: changes are logged with the data used, model version, and the human approver; exceptions trigger review.
Metrics: the team tracks gross margin, conversion rate, price realisation, and the hit rate of model-recommended moves.
Nothing exotic here. It’s structure. The result is a shift from quarterly batches to weekly adjustments, with a measurable lift in margin and a shorter time from signal to decision. Studies continue to show that organisations see value when they reshape workflows and keep humans in the loop at the right points. (McKinsey & Company)
The practical roadmap
1) Assess: map decisions before you map data
Start by mapping critical decisions where AI could change outcomes: what is the decision, who makes it, what inputs feed it, when is it made, and what metric defines “better”? Follow with a light capability matrix (data quality, model ops, experimentation, governance) and a RACI that names the Product Owner, Model Owner, Data Steward, and Risk/Compliance counterpart. Clear roles beat vague ambition. (USC Center for Effective Organizations)
2) Redesign: organise around value streams
Move from project teams to cross-functional pods aligned to value streams such as pricing, retention, underwriting, or claims. Embed data and ML inside the pod, give it a delivery cadence, and set decision SLAs so models land inside the window where choices are made. This is the shift from adjacency to integration. Leading operating-model guidance for AI stresses cross-functional ownership and cadence as prerequisites for scale. (BCG)
3) Govern: make models auditable and safe to use
Adopt a risk framework that is practical for practitioners. Define model cards, data lineage, evaluation suites, bias checks where relevant, and thresholds for human review. Agree escalation paths. Document context and intended use. Compliance becomes faster when it is designed into the lifecycle rather than stapled on after deployment. The NIST AI RMF is a solid reference to structure this without freezing innovation. (NIST Technical Series)
4) Adopt: invest in skills, incentives, and change
Adopt means capabilities, incentives, and story. Build a role-based enablement plan that maps skills to decisions, not tools to job titles. Give managers a communication kit: the one-page narrative, the three metrics that matter, the next small action for their team. Align incentives so teams win when frontline metrics move, not when projects ship. Make leadership behaviours explicit: visible sponsor time in value-stream reviews, quick decisions on blockers, praise for teams that retire work as well as add it. Keep the organisation informed with a steady rhythm of short, useful updates that show the link from model output to business outcome. Adoption follows confidence, and confidence follows clear instructions and fast feedback.
5) Scale: from pilot to workflow to portfolio
Codify what worked into playbooks: data interfaces, evaluation standards, decision cadences, and rollout checklists. Then move horizontally to similar value streams in adjacent brands or geographies. Scaling is a management problem, not an algorithmic one; companies that industrialise these patterns see sustained returns rather than a spike of early wins. (BCG)
How to track value without getting lost in dashboards
Start with frontline financials, then trace to operations. If you measure only model accuracy, you will optimise the scoreboard, not the game.
Frontline financial metrics
Revenue: conversion lift, average order value, price realisation.
Margin: gross margin improvement, markdown reduction, loss-ratio gains.
Cost-to-serve: handling time, automation rate, first-contact resolution.
Operational flow metrics
Cycle time: time from signal to decision to action.
Forecast accuracy: MAPE or bias on demand or cash forecasts.
Case resolution: backlog burn down, re-open rate.
CSAT/effort: customer satisfaction and effort scores.
Throughput: tickets per agent, quotes per underwriter, claims per adjuster.
Tie these metrics to test designs that create credible counterfactuals. Use A/B tests where you can randomise, pre/post with matched controls where you cannot, and difference-in-differences for staged rollouts. This is well-trodden ground: disciplined business experimentation reduces opinion and raises signal, and the online experimentation canon explains how to run these tests at scale with integrity. Build a small “control tower” that reports adoption, model performance, and these outcome metrics per value stream, with owners accountable for variance. (Harvard Business Review, PMC)
Track the human side with the same discipline. Measure adoption and depth of use by role, not just log-ins. Track enablement completion and time-to-competence for the roles that touch the decision. Use lightweight pulse checks to gauge understanding and trust in the new decision loop. Monitor message reach and recall across sponsors, managers, and frontline teams. Tie these leading indicators to the lagging financials so you can see whether the story is landing before the numbers move.
A final point on governance and value: keep humans in the loop where risk concentration is high, and make that oversight measurable. Logs of human approvals, overrides, and exceptions become inputs to model improvement, audit readiness, and trust. Risk frameworks encourage documented human oversight that is proportionate to impact and context. (NIST Technical Series)
Common anti-patterns to retire
Tool-centrism: buying platforms before mapping decisions.
Shadow ownership: data owned by everyone and no one.
Siloed squads: model teams far from the decisions they influence.
Pilot theatre: demos that never make contact with the frontline.
Metrics fog: accuracy charts with no link to revenue, margin, or cost.
Companies that avoid these traps converge on a simple truth: durable ROI comes from aligning structure, incentives, and cadence so AI changes how work gets done. Surveys of AI leaders keep reinforcing that process redesign and human-in-the-loop practices correlate with EBIT impact. (McKinsey & Company)
Closing in the GAPx voice
We do not see LLMs as silver bullets. We see them as a catalyst that rewards well-run organisations. If you want AI to compound, align people and process first, then wire in governance and clear communications that make the change simple to follow. Give decisions owners, organise around value streams, show your work in public, and measure the effects where they count. That is how portfolios turn capability into cash flow, and how operators turn pilots into a performance system that lasts. If you are ready to align culture, communications, and operating model with AI, we are ready to help you build the engine that makes it pay.
