AI agent strategy · Mid-size and large companies
AI doesn't need more ideas. It needs decisions.
I help companies decide where to apply AI, which agents are worth building, and what capabilities to put in place before any of it reaches production.
What happens when you click: you describe your initiative and a council of agents challenges it live. Free, in a couple of minutes, no sign-up.
- ForProduct, operations, strategy and innovation leaders
- You leave withA prioritised portfolio of initiatives and a plan you can execute
- No biasI tell you when not to use AI: I don't make money implementing
- ForProduct, operations, strategy and innovation leaders
- You leave withA prioritised portfolio of initiatives and a plan you can execute
- No biasI tell you when not to use AI: I don't make money implementing
The problem
Getting access to AI is no longer the problem. Knowing what to do with it is.
- 01
Too many ideas
Every department showed up with its own initiative. There are pilots, vendors and demos everywhere, and nobody knows which ones contradict each other.
- 02
No real prioritisation
Things get prioritised by enthusiasm, or by whoever raised their hand first. Not by dependencies, risk, or which initiative unlocks the others.
- 03
No operating model
Nobody defined what a person still does, what an agent decides, who approves, and what happens when the agent gets it wrong.
The stance
The only AI consultancy that will tell you when not to use AI.
A consultancy whose business is implementation cannot recommend that you don't implement. That's a structural limitation, not a matter of judgement.
A good share of the initiatives that arrive as “we need an agent” do not need one. Some are solved with deterministic automation. Others are a process problem dressed up as an AI problem. And some simply do not have the data. See for yourself: bring an idea to the council on this page and look at the verdict.
Those discards are the most valuable deliverable of the work: they're the budget you won't burn.
One example in depth · Stage 03
The question is not where to put an agent. It is how the work gets divided.
A solved case: a collections process, six activities. For each one, who does it — a person, a rule or an agent — and why. The verdict is at the bottom.
| Mode | Process activity | Why |
|---|---|---|
| Human only | Negotiate a write-off outside policy | It commits the company and requires judgement about the commercial relationship. |
| Human + AI | Prepare the case before calling the customer | The analyst decides; AI pulls together history, payments and context in seconds. |
| AI assisted | Draft the restructuring proposal | AI drafts, a person approves before it goes out. |
| Agent executed | Log the agreement in the CRM and schedule follow-up | Bounded, reversible action with a trace. No approval needed. |
| Agent autonomous | Rank the day’s caseload by likelihood of recovery | Decides within defined limits; the supervisor sees the criteria applied. |
| Automation | Send the reminder 3 days past due | Fixed rule. Putting a model here would be expensive, slow and worse. |
Out of six activities, only two justify an agent. One needs no AI at all. That detail is the difference between a pilot that scales and one that dies in production.
Free: in a couple of minutes the council tells you whether yours is an agent, a rule, a data or a process case.
How we work together
Start for free. Go deeper when you need to.
- 01
Quick verdict
Free · A couple of minutes
You describe your idea and a council of agents challenges it live: process, business, opportunity and economics. You leave with a verdict and the questions you would need to answer.
- 02
Full analysis
CLP 149,000 · The same day
Your real case with the full method: interview, sourced cases, readiness, autonomy and economics in ranges. A report to decide and present.
You request it from the free verdict, with your case already loaded.
- 03
Programmes
By scope · 2 to 10 weeks
When there are several initiatives or a transformation under way: Scan, Strategy or Advisory, with me. It starts with a 45-minute conversation at no cost.
See what each programme includes
- Entry point
Transformation Scan
- Maturity and context diagnosis
- Prioritised opportunity map
- Preliminary readiness for the top 5
- Executive closing session
For when the question is still “where do we start?”.
- Core engagement
Agentic Strategy
- All five stages of the method
- Agent Portfolio and Blueprints
- Capability & Dependency Map
- Governance and autonomy levels
- Roadmap in waves with business case
For when there are already scattered initiatives and one strategy is missing.
- Ongoing
Advisory
- Review of new initiatives
- Assessment of vendor proposals
- Review of architectures and pilots
- Living roadmap and board reporting
For when the transformation is already under way and the judgement has to hold.
The workspace is included in all three programmes. The quick verdict and the full analysis are done by agents, and every report goes through a critic before delivery. The programmes add what an agent does not replace: a person who challenges every decision with you and answers for it.
The fee for each way of working is set by scope, in the first conversation.
What stays in your company with a programme
- 01
AI Opportunity Portfolio
Every opportunity raised, scored and with its status — including the discarded ones and why.
- 02
Human + Agent Operating Model
How the work is divided per process: what stays human, what gets automated and what justifies an agent.
- 03
Agent Blueprints
A full spec per priority agent: mission, knowledge, tools, actions, autonomy, escalation, guardrails and KPIs.
- 04
Capability & Dependency Map
Which shared capabilities to build first and which initiatives each one unblocks.
- 05
Governance Model
Autonomy levels, controls, traceability and owners per agent.
- 06
Transformation Roadmap
A real sequence in waves, with dependencies, investment and the criteria for deciding whether an initiative moves forward or stops.
The platform
Consulting operated by agents.
The method does not end in a PowerPoint. It lives in a workspace where every opportunity is challenged by specialised agents — business, process, architecture, governance and economics — before it enters the roadmap, and the roadmap recalculates when a dependency changes. The council in this mockup works right here on this page.
Agents analyse information, organise decisions and speed up the work. Strategic direction always stays in human hands.
OPP-014 · Requested by Service Management
“We want an agent that answers customer queries.”
- Process Analyst61% of queries are about the status of a shipment. Today the customer cannot see that anywhere.
- Business StrategistThe stated goal is to bring call volume down, not to improve answer quality. Those are different things.
- AI OpportunityThis is not an AI problem: it is a visibility problem. Exposing shipment status removes 61% of the queries with no agent at all.
- EconomicsConversational agent: 14-month payback. Shipment status in the app: 3 months.
Recommendation · No agentReframed as OPP-014b. The remaining 39% gets reassessed later.
Free and live: four specialists who deliberate without seeing each other and a chair who issues the verdict, in a couple of minutes.
The method
Before building agents, let’s design the organisation that needs them.
Five stages. The model is the last decision, not the first: we start from the business outcome and technology only shows up at the end. And the analysis is not by use case, it is of the work that actually happens — which is where the opportunities nobody proposed turn up.
- 01
Understand
Objectives, capabilities, processes, decisions, roles, systems and knowledge. The unit of analysis is not the use case: it is the work that actually happens.
Deliverable Organization Model
Evidence map - 02
Redesign
How that work should be done once people and agents can collaborate, decide and execute. Current state against target state, activity by activity.
Deliverable Human + Agent
Operating Model - 03
Agentify
What stays human, what is deterministic automation, and what genuinely justifies an agent. This is where the method says no.
Deliverable Agent Portfolio
Agent Blueprints - 04
Enable
Which shared capabilities — identity, knowledge, tool layer — have to be built first because they unblock several initiatives at once.
Deliverable Capability Map
Dependency Graph - 05
Govern
Autonomy, guardrails, evaluation and traceability proportional to each agent’s risk. Oversight goes up as autonomy goes up.
Deliverable Governance Model
Transformation Roadmap
Three decisions, up close
We don’t prioritise initiatives. We prioritise what unblocks initiatives.
The classic roadmap mistake is ranking agents by individual ROI. Almost always, three shared capabilities explain why half the portfolio is blocked.
None of the three base capabilities is, in itself, an AI project. They are integration, permissions and access to knowledge. And they are the reason 70% of the portfolio cannot start yet.
Autonomy is not decided once. It is raised in steps.
Every agent has a current level and a target level. Human oversight and controls are sized against that level, not against how excited the project is.
- L0 No AIEntirely human work.
- L1 · CurrentAssistGathers and summarises information. Makes no proposals.
- L2 RecommendProposes an answer; the person decides.
- L3 · TargetExecute with approvalPrepares the action and waits for sign-off.
- L4 Bounded autonomyDecides and executes within defined limits.
- L5 AutonomousRuns on its own; escalates exceptions only.
Current levelTarget level at 12 monthsJumping from L1 to L4 is the most expensive mistake in the field
A maturity percentage is useless. Readiness is per initiative.
Readiness is not a stage of the method: it is the gate every initiative has to pass before it becomes a pilot. And it is assessed one by one, not for the company as a whole.
The business case is solid and the process is clear. It still does not start. Four conditions have to be resolved first:
- Collections history lives in three systems that do not talk to each other.
- There is no write API on the collections core.
- Nobody has been assigned the agent owner role.
- There is no definition of what it means for the agent to have worked.
With that resolved, the same agent moves to 84 and goes to pilot.
Why me
I can hold my own in all three rooms.
In the executive room I talk about value, customers, processes and priorities. In the architecture room, about RAG, MCP, tool calling, memory, evaluation and cost. In the product room, about journey, adoption and friction.
More than twenty years building web and mobile platforms, the last few focused on systems built around AI and agents. I do not do the implementation — your team or your vendor does that — but I know enough that nobody can sell you something that cannot be sustained.
Built from experimentation, not from PowerPoints.
This method did not come out of an industry report. It came out of building and running agent systems in production: platforms where specialised agents run research, simulate users, generate campaigns, evaluate products and coordinate with each other.
Everything I propose — orchestration, context engineering, human-in-the-loop, evaluation, model routing, multimodal generation — I had to solve in production first. That is why I can tell you when something is not going to work.
- Multi-agent orchestration
- Context engineering
- Model routing
- Human-in-the-loop
- Evaluation
- Observability
- Multimodal generation
- User simulation
VISITOR QUESTION
|
TRIAGE
|
+--------+-----+-----+--------+
v v v v
PROCESS BUSINESS OPPORTUNITY ECONOMICS
| | | |
+--------+-----+-----+--------+
v
CHAIR
|
VERDICT · DISSENT · QUESTIONS
|
INTAKE
|
BOOKING · BRIEF · INVITE
|
HUMAN: THE CONVERSATIONThe shortest proof is on this very page: the council you can convene is this architecture at work. Independent agents in parallel, a structured verdict, cost limits, and a person exactly where commitments are made: in the conversation.
Frequent questions
What almost everyone asks before getting started.
These answers are the synthesis. To see how they are reached — four agents arguing, one of them dissenting — open the deliberation. Or bring the council your own question.
Where do I start if I have ten loose AI ideas?
Do not start by choosing. Start by laying out the full portfolio with its status and dependencies: the sequence falls out on its own once you can see what unblocks what. That is a two-week Transformation Scan.
How do I know whether a use case justifies an agent?
An agent is justified by decisions, not by volume. If you can write the rule on one page, automate. If the answer already exists and nobody can see it, expose the data. The agent is for what is left.
What do I need in place before running a pilot?
Readiness is not a company-wide maturity percentage: it is assessed per initiative. The same agent can score 92 on business case and 22 on evaluation, and that 22 is the one that decides.
What do I get for free?
A quick verdict on your initiative in a couple of minutes: four agents analyse it from process, business, opportunity and economics, and the chair decides whether it is a case for an agent, for automation, for visibility, for process, or whether it should stay human. No sign-up.
What is the difference between the quick verdict and the full analysis?
The quick verdict starts from one sentence and uses general judgement. The full analysis, at CLP 149,000, works on your real case: it interviews you, researches sourced cases, applies the five stages of the method and delivers a report with readiness, autonomy and economics in ranges, reviewed by a critic before it reaches you.
Who does the analysis: people or agents?
Specialised agents, each on its own part: interview, research, analysis and critique. An automatic critic reviews every report, followed by a window of up to 3 hours in which Roberto can read and correct it before it goes out. If no changes are needed, it is delivered automatically at that time.
Let's talk
Bring your AI chaos. Let’s leave with a portfolio.
If there are already several initiatives in play, let us start with a 45-minute session, at no cost. You tell me what you have running today and I tell you plainly what I see: what I would scale, what I would stop, and what is blocked by something nobody is looking at.
- You book 45 minutes Pick a time. No long forms, no commitment.
- We review what you have Initiatives in play, priorities and what is blocking them.
- You leave with a clear next step What I would scale, what I would stop and where to start.
