COMPASS is a loop, not a report
The framework follows the Theory of Constraints' process of ongoing improvement: find the constraint, elevate it, and when it moves — start again. Run these five steps as a quarterly rhythm.
- Locate your stage. Take the 3-minute diagnostic, or place yourself against the stage descriptions below. Answer on evidence — what is deployed, who can build, what your data supports today — not on ambition. Most organizations overplace themselves by one stage.
- Name the binding constraint. One constraint governs your progress at any given time. Validate the diagnostic result with three conversations: an executive, a delivery lead, and an end-user. Ask each, "what is actually stopping us?" If their answers diverge, that divergence is itself data — and usually points at leadership alignment.
- Elevate the constraint. Concentrate resources on the binding constraint before optimizing anything else. The stage playbooks below tell you how, response by response. Resist the instinct to fix everything at once — that instinct is why broad transformation programs stall.
- Track the stage indicators. Each stage has four success indicators. Review them monthly, in the same forum where you review financials. What gets reviewed at that table gets done.
- Re-diagnose and advance. When your indicators have held for a quarter, retake the diagnostic. The constraint will have moved — that is success, not instability. A new binding constraint means you've earned the next stage's problems.
Explore — earn the right to invest
You are here if AI is aspiration, pilots, or proofs-of-concept — and internal capability is near zero. The goal of this stage is not transformation; it is one proven use case, one champion, and a leadership team that understands what it is buying.
1 · Run an AI discovery & diagnostic phase
Inventory your core processes and score candidate AI use cases on two axes: business value and implementation complexity. Involve process owners, not just IT — the highest-value use cases are usually invisible from the technology side. Rank them and pick from the high-value, low-complexity corner.
Done when: you hold a ranked shortlist of use cases, each with a named business owner and an estimated value metric.
2 · Pilot one high-value, low-complexity use case
One use case, not a program. Define the ROI metric before you start — hours saved, error rate, cycle time, revenue per lead — and instrument its measurement from day one. Timebox the pilot to 60–90 days. Refuse scope additions ("while we're at it…") until the pilot ships; broad initiatives without a proof point stalled in every case studied.
Done when: the pilot is live with real users and its metric is being captured against a pre-pilot baseline.
3 · Designate an internal AI champion
Select for organizational authority and peer credibility, not technical enthusiasm. Protect their time formally — an agreed share of their week, visible to their manager. Endorse them publicly from the top, and give them co-leadership of the pilot alongside any external partner. This single design decision correlated with success more than any other in the research.
Done when: the champion's role, time allocation, and decision authority are written down and announced.
4 · Structure consulting for capability transfer
If you use external help, put transfer in the contract: named internal counterparts who co-deliver, documentation obligations, and internalization milestones — "by month X, our team operates this without the partner." Delivery-only engagements create dependency that compounds; capability-transfer engagements make the next project cheaper than the last.
Done when: your agreement names who learns what, by when — and payment maps to those milestones.
5 · Launch executive AI literacy before major spend
Short, recurring executive sessions built on your own use cases — not generic demos. The goal is scoping judgment: what AI can and cannot do in your business, what good data looks like, what realistic timelines are. The research is unambiguous: leadership literacy, not budget, is the primary determinant of success.
Done when: your executive team can articulate the pilot's purpose, metric, and limits without the champion in the room.
Implement — close the adoption gap
You are here if AI runs in specific functions but integration is partial, engagement uneven, and ROI still argued rather than shown. The defining risk of this stage is the implementation–adoption gap: tools that are technically live and behaviorally dead.
1 · Set a minimum viable data standard
Do not launch a company-wide data overhaul. List the specific data your next use case needs; assess only that slice for quality, access, and ownership; close only those gaps in a time-boxed readiness sprint with a named owner per source. Governance grows use case by use case, in parallel with delivery — never as a prerequisite that delays it.
Done when: the data for your next use case has a quality bar, an owner, and a date.
2 · Deploy in phases, module by module
Roll out function by function, with an adoption target and a review gate per phase. A phase that misses its gate gets fixed or killed before the next begins. Modular deployment turns one big bet into a sequence of small, reversible ones.
Done when: each live module has passed a gate with its adoption numbers on the table.
3 · Co-design adoption with end-users
Before go-live, run working sessions with the people whose day changes: what disappears, what replaces it, what's in it for them. Set behavioral milestones — e.g., share of cases processed through the tool — and track them after launch with the same discipline as uptime. Passive non-engagement doesn't announce itself; you have to measure for it.
Done when: adoption rate is a tracked number with a target, reviewed monthly.
4 · Build an internal AI learning program
Convert consulting dependency into internal capability: role-based training paths, a second-tier champion in each function, and lessons from every project written into shared playbooks rather than left in people's heads. This is how Explore-stage knowledge transfer becomes Implement-stage muscle.
Done when: a second-tier champion exists outside the original pilot team, and playbooks outlive personnel changes.
5 · Stand up an ROI tracking framework
Baseline before deployment, then keep a simple ledger per use case: cost, usage, outcome metric. Publish the results internally — visible ROI is simultaneously your funding argument and your strongest culture countermeasure against detachment and evaluation paralysis.
Done when: any executive can see cost and return per AI use case on one page.
Scale — turn wins into infrastructure
You are here if AI is integrated across functions and the question has shifted from "does it work?" to "how do we do this faster, safer, and everywhere?" The constraint is no longer proof — it is platformization, governance, and sustaining the learning culture under growth.
1 · Platformize your AI capability
Extract what your projects share into reusable assets — deployment templates, assessment protocols, internal SDKs — and treat each as a product with an owner and a roadmap. Advanced organizations in the research distinguished themselves precisely here: lessons encoded into assets, not anecdotes.
Done when: a new use case starts from your assets, and time-to-deploy is measurably falling.
2 · Pre-engage governance stakeholders
Bring compliance, legal, security, and worker representatives in before builds, not at launch. Agree guardrails once, then build inside them. Pre-engagement converts your slowest stakeholders from a late-stage veto into a design input.
Done when: a standing governance forum reviews upcoming AI work on a fixed cadence.
3 · Run iterative implementation cycles
Short cycles with explicit learning reviews, whose lessons feed back into the platform assets. Bureaucracy is the dominant constraint at this stage; iteration cadence is your counterweight to it.
Done when: cycle retrospectives produce changes to templates and protocols, not just notes.
4 · Make explainability a design requirement
For regulated or human-facing applications, specify up front how outputs are produced, reviewed, and overridden — and name an accountable owner per system. Explainability retrofitted after deployment is both expensive and unconvincing; specified at design time, it is your license to operate in high-stakes domains.
Done when: every high-stakes AI system has documented logic, a review path, and a named owner.
5 · Measure innovation, not activity
Track reuse rate, time-to-deploy per use case, and adoption durability — not project counts. These metrics reward the platformization and learning behaviors that keep a Scale-stage organization from silently sliding backward.
Done when: innovation metrics sit beside financials in leadership reviews.
Constraint countermeasures — quick reference
| If this binds you | First moves |
|---|---|
| Leadership | Executive literacy sessions before further spend; personal sponsorship with named accountability; a champion with real authority. Nothing else compensates for this one. |
| Talent | Stage-matched capability building — broad literacy early, champions and learning programs mid-journey, specialized roles at scale; consulting contracts that transfer knowledge inward. |
| Data | Use-case-scoped minimum standard; time-boxed readiness sprint; run it parallel to the pilot, never as a precondition. |
| Culture | Behavioral co-design with end-users; decision gates with deadlines against evaluation paralysis; published ROI against quiet detachment. |
Common questions
Does COMPASS apply outside Mexico?
Yes. The framework was developed and validated in the Mexican mid-market — one of the most demanding environments a framework can be built in: scarce specialized talent, uneven data infrastructure, constrained budgets. The four constraints and three stages it maps are structural, and the same patterns are documented in mid-market firms worldwide. What varies by country is the intensity of each constraint — which is exactly what the diagnostic measures, because it asks about your organization, not your geography.
Is it only for mid-sized companies?
It was built for organizations of roughly 50–250 people — but the evidence base includes a department of a global enterprise, and departments of large organizations behave like mid-sized firms in constraint terms. If your unit can't simply outspend its constraints, COMPASS applies.
Do we need consultants to use it?
No. The framework is designed for self-application — and when you do bring in external help, it prescribes how: structured for capability transfer, with co-leadership and internalization milestones, never delivery-only dependency.
How long does a stage take?
There is no fixed timeline, and pretending otherwise is how transformation theater starts. You advance when your stage indicators hold for a quarter and the re-diagnosis shows the constraint has moved. Some organizations cross Explore in six months; others need two years — the sequence matters, the calendar doesn't.