· By Ricardo Torres Oliva
A leadership team that decides by the same criteria
Why every meeting about artificial intelligence starts again from zero in the committee, and the minimum vocabulary to decide in a single session.
Abstract
Six competent functions can receive the same instruction from the board and execute six different initiatives, incomparable with one another, without any of them being wrong: the sentence admitted all six readings. What fails is the specification, not the willingness to change, and the remedy is to agree, not to persuade. Written for the senior management team that receives an artificial intelligence decision already taken and has to turn it into operations, the text identifies what gets lost between board and management, fixes a vocabulary of six terms with operating definitions, and delivers a framework of eight criteria — five binary gates and three for priority — with which a committee evaluates initiatives in a single session. It closes with what the framework does not fix on its own and with seven actions for the coming week.
An artificial intelligence decision that comes down from the board without its criteria does not come down whole: it comes down as a budget and as a deadline. Each function completes it with whatever criterion it has to hand, and the result is a portfolio of initiatives that nobody designed. This document delivers the vocabulary and the framework that prevent that dispersion.
1. The problem
The decision arrives as a sentence and a number. "We are going to use artificial intelligence in customer service, with this budget and this deadline." What does not arrive is the reasoning: why that process and not another, what risk was accepted in choosing it, and what was considered a sufficient result.
Six competent functions receive that sentence and translate it honestly. Operations reads automate. Sales reads more conversion. Technology reads platform. Finance reads fewer heads. Legal reads exposure. Talent reads conflict. None of them is wrong, because the sentence admits all six readings. Six months later, the committee reviews six initiatives that share neither the definition of success nor the unit of measurement, and which therefore cannot be compared with one another or prioritized.
The usual diagnosis of that situation is resistance to change. It is incorrect and it is also costly, because it points the remedy toward persuasion when the problem is one of specification. Nobody resisted. Each one filled in what was missing.
The magnitude of the matter is not marginal. The most replicated heuristic of adoption distributes success at roughly 10% algorithm, 20% technology and data, and 70% people and processes — the source itself asks that it be read as an order of magnitude, not as a measurement [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo]. And the hard evidence points the same way: high-performing companies are three times more likely than their peers to have redesigned their workflows around artificial intelligence [source: AI Literacy A Multi-Dimensional Analysis of Governance, Revenue Systems, and Epistemic Rigor in the Agentic Era]. Redesigning a process is management's job, not the board's and not the vendor's.
2. Why the usual approaches fail
The alignment announcement. The all-hands meeting where the decision is explained and commitment is requested. It communicates the what; the criterion does not travel inside a presentation. A criterion is not communicated: it is practiced on cases until two managers, facing the same file, reach the same conclusion.
The AI committee without a framework. It meets, listens to presentations and approves by conviction and seniority. The doctrine is explicit about this figure: committees that exist to distribute responsibility without taking it [source: The Phoenix Doctrine v1.1]. A committee without a cut-off framework turns evaluation into a contest of presentation quality.
The single mass training event. One event, with no reinforcement, no evidence and no path afterwards. The Research_Base calls it a skills placebo and treats it as an anti-pattern, not as an insufficient step [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo].
Adoption by decree. Licenses for everyone, redesign for no one. It produces shadow AI and the illusion of progress: scattered use, with no owner, no measurement and outside the security perimeter [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo].
The isolated laboratory. The innovation team that produces brilliant demonstrations that never touch operations — pilots with no business owner and no baseline [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo].
All five share one cause: they attempt to resolve with communication, structure or enthusiasm a problem that is one of absent shared context.
3. What exactly gets lost on the way down
Three things travel badly between the board and management, and they are always the same three.
The reason for the priority. Why that process. Without it, each function assumes that its own process also qualifies, and the budget fragments into small initiatives that never reach measurable scale.
The risk tolerance that was accepted. How much error is admissible and in exchange for what. Without it, one function deploys with high autonomy on a sensitive process while another blocks a trivial process behind manual approval. Both believe they are following the same instruction.
The definition of done. What figure, on what cases, defines that the initiative is finished. Without it, "done" means deployed, and learning stops exactly where it should have begun.
There is an exact analogy, and it is worth using because it is operational, not decorative. An agentic system decides badly when it receives a context that is incomplete or saturated with noise; the discipline that corrects this — curating what information enters each decision — has a declared owner, because ownerless context is a recognized anti-pattern: nobody curates what the system sees [source: Context Engineering La Disciplina del Contexto en Sistemas Agénticos]. A management layer works the same way. The difference is that in most companies nobody curates management's context, and so each manager decides with whatever context they managed to obtain on their own.
The golden rule of that discipline also carries over: more is gained by removing noise than by adding information [source: Context Engineering La Disciplina del Contexto en Sistemas Agénticos]. The committee's reflex response is to add — more reports, more steps, more approvals. The first correct move is to remove.
4. The minimum vocabulary, agreed in writing
Six terms. The third column is the one that does the work: a definition is useful when it closes off sentences that could previously be said without consequence.
| Term | Agreed operating definition | What can no longer be said |
|---|---|---|
| Automation | A deterministic rule executed by a machine: same input, same output, and the failure is visible. | Calling a rules-based flow "AI" in order to access the AI budget. |
| Agent | A system that decides intermediate steps toward a goal, chooses tools and can fail in ways nobody wrote down. It does not produce a verdict: it produces a rate [source: Evals Ingeniería de Confiabilidad y Evaluación de Sistemas Agénticos]. | "It works" as a statement with no number behind it. |
| Baseline | The measurement of the process before anything is installed: cost, cycle time, volume and error, with method and date. | "It improved quite a bit" with no recorded starting point. |
| Eval | A curated set of real cases with expected outcomes, which produces a success rate. Nothing is done until it passes its evals, and whoever judges is not whoever built it [source: Evals Ingeniería de Confiabilidad y Evaluación de Sistemas Agénticos]. | Presenting a demonstration as proof of reliability. |
| Pilot | A system without evals [source: Evals Ingeniería de Confiabilidad y Evaluación de Sistemas Agénticos]. | Using "pilot" as a stage that justifies not measuring. |
| Harness owner | The named person who approves increases in autonomy, reads the evaluation results and answers for the system before the business [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo]. | A system in operation whose owner is a department rather than a person. |
One clarification on the last term: in a mid-sized company these are hats before they are positions. The same person can be responsible for two systems. What cannot happen is that the hat does not exist, because an agent without an owner is shadow AI with a budget [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo].
The proof that the vocabulary is agreed is not that everyone has read it. It is that two managers from different areas, looking at the same initiative separately, classify it the same way.
5. Initiative evaluation framework
Eight criteria, applicable by a committee in a single session, at roughly five minutes per initiative. The first five are binary gates: either they are met or the initiative goes back. The last three order priority among those that passed.
| # | Criterion | The question in the session | Evidence presented | Cut-off rule |
|---|---|---|---|---|
| 1 | Owner | Who answers for this before the committee? | A name and the declared hat | Without a proper name, it is not evaluated |
| 2 | Baseline | What is the measurement of the process today? | Figure, method and date | Estimate without method: goes back |
| 3 | Definition of done | What rate, on which real cases, defines success? | A set of cases with expected outcomes | "Improve efficiency": goes back |
| 4 | Destruction | What stops being done if this works? | A named process, report or step, with a date | "Nothing, it adds to what we do now": goes back |
| 5 | Oversight | How many human interventions per day does it consume, and who performs them? | Estimated rate × minutes per intervention = hours/day | Supervisor without budgeted hours: goes back |
| 6 | Full cost | What does a successful task cost, including retraining, data and integration? | Breakdown, not license price | Orders priority |
| 7 | Reversibility | If it has to be switched off in thirty days, what is kept and what is lost? | Shutdown and portability plan | Orders priority |
| 8 | Redistribution | Where does the freed-up time go, and has that been communicated? | Written policy | Orders priority |
Three notes on use, which are where the framework is won or lost.
Gate 4 is the distinctive one, and the one that causes the most discomfort. Deploying on processes that have not been redesigned produces likeable pilots and zero advantage [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo]. An initiative that does not name what ceases to exist is a layer on top of the previous operation, and its cost is the sum of both.
Gate 5 is calculated, not estimated. A system that asks for ten five-minute approvals a day consumes one hour a day of its supervisor; with eight such systems, the supervisor is the process [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo]. The approval queue is a job that gets designed, with a maximum daily volume and rotation, because an exhausted approver degrades the system's most important control into a formality.
The committee's most valuable product is not its approvals. It is the list of what went back and why. That list is the criterion making itself explicit, and after three or four sessions it stops growing: the functions begin to present already filtered. That is the moment the framework stopped being a procedure and became vocabulary.
6. What the framework does not fix on its own
Structure. The pattern the evidence favors is a small core — standards, evaluation, governance: the owner of the how — with federated capability in the business areas, where the owners of each system live: the owners of the what. The anti-pattern is the AI department that centralizes all execution: it becomes a bottleneck and produces systems with no business owner [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo].
The policy for redistributing the savings. Freed-up time is reassigned under an explicit policy — higher-value work, training, oversight — and communicated. Invisible savings are perceived as a threat; redistributed savings, as a tool [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo]. This is a cultural declaration by senior management, and it arrives late if it is issued after the first deployment.
Fear, which is rational. A good part of the exposed roles have no easy transition, and denying it destroys the trust that adoption needs [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo]. The corresponding position in the doctrine is uncomfortable and worth holding: the worst way to minimize the pain of the transition is to deny it; telling the team that everything will be fine without investing in real training is a lie that comes due [source: The Phoenix Doctrine v1.1]. Announcing replacements before there is evidence to support them is the other extreme, and it has a name of its own: headcount theater.
The voice of those who operate. The failure-reporting channel only works if what is reported turns into an improvement and the improvement is announced with credit to whoever originated it. Those who see their corrections become improvements stop sabotaging and start training [source: Organización Humano-Agente La Empresa Híbrida y el Rediseño del Trabajo].
7. What to do on Monday
- Write the six definitions from section 4 on one sheet and have the committee agree them. If two managers define "agent" differently, that discussion is the whole session, and it is worth it.
- Run the AI initiatives that are live today through the five gates. Publish how many go back and on which criterion. The first round usually returns most of them; that is the expected result, not a failure of the framework.
- Put a first and last name on every AI system in operation, including those contracted without going through the committee.
- Calculate the daily oversight hours each deployed system consumes and check whether they are budgeted anywhere. Usually they are not.
- Ask each function for a single sentence per initiative: what stops being done if it works. Those that cannot write it are the ones to review first.
- Agree and communicate the policy for redistributing freed-up time before the next approval, not after the next deployment.
- Set a monthly review of the same metrics on the same cases. Comparability over time is worth more than the precision of any isolated measurement.
What we have not covered here
This document assumes the decision already taken and the money already approved. It does not cover the prior decision: whether to invest at all, what to require in a proposal, what to ask a vendor and what not to buy. Those questions are resolved at the board, with another audience and another legal responsibility, and they have their own document.
Nor does it cover the technical architecture of what is operated — agent anatomy, agentic security, context engineering — or the detailed definition of autonomy levels, which is a framework of its own that this text mentions without developing. It does not cover training the full operating workforce, which is a different system from the management agreement described here. And it does not cover how that agreement is built with the real team, on its own initiatives and on its own calendar, which is the work of Phoenix TEAx, from VoltAi Academy.