+55 11 4193 5660 PT Free assessment

AI Transformation

Six layers between where you are and a company that runs on AI.

Not a chatbot on the website, and not a pilot that dies in committee. It is foundation, knowledge, agents, products, operation and culture, built in the right order, with the model running wherever your data allows.

I want the AIMI assessment Explore the layers

Why almost everyone stalls

The model was never the problem.
Everything around it was.

85%

of AI pilots never reach production scale, for lack of governance, integration or platform discipline.

McKinsey Global Institute · The State of AI

What sets the leaders apart

Companies capturing outsized value from AI share five measurable traits. The distance is not an accident: it is an architectural and organizational choice.

  • Platform approach +20 p.p.
    66%
    46%

    A single platform for data, models, governance and integration instead of fragmented point solutions.

  • Governance and security +21 p.p.
    63%
    42%

    Formal policy, risk control and security protocols built before scaling.

  • Innovation mindset +15 p.p.
    53%
    38%

    Experimentation and learning as operational discipline, not as cultural aspiration.

  • Team composition +21 p.p.
    50%
    29%

    AI engineers, domain specialists and change facilitators, not only data scientists.

  • Agentic AI adoption +17 p.p.
    36%
    19%

    Autonomous agents running real business flows, beyond assistants and chat.

AI leaders Everyone else

  • 20–40% productivity gain among knowledge workers using corporate AI assistants. McKinsey · market benchmark
  • 60–70% of routine knowledge-work tasks can be automated by agentic AI systems. McKinsey & Company · Agentic AI Outlook
  • 25–30% reduction in service desk operating cost with AI-first triage and classification. McKinsey Digital · AI-Enabled Operations

The market in 2026

Adopting is now consensus.
Operating is not.

The figures in this section come from the AI Index Report 2026 by Stanford HAI. It is an independent academic effort, now in its ninth edition, not a survey sponsored by anyone selling the technology.

  1. Organisations already using AI

    88%

  2. Use generative AI in at least one business function

    70%

  3. Agent deployment, across nearly every business function

    single digits

Adoption is nearly universal. Operating with agents has barely begun. That distance is exactly what the six layers cover.

  • 26%

    Measured gain in software development

    Customer support landed between 14% and 15%, and marketing output reached 50%. The gain shows up where the work is structured and the result is easy to measure, and shrinks where the task demands judgement.

  • 362

    Documented AI incidents in 2025

    There were 233 in 2024, a 55% rise in a single year. Meanwhile, the safety benchmarks of the models themselves advance slowly and lose ground under deliberate attack.

  • 22% to 94%

    Hallucination range across 26 top models

    The spread on a new accuracy benchmark. This is why evaluation, verified retrieval and guardrails come before scaling, and not after the first error in production.

  • 12% → 66%

    Agents’ leap on real computer tasks

    One year of progress on OSWorld. Even so, they fail roughly one attempt in three. Autonomy is earned through evidence, never assumed.

  • 11%

    Of companies still have no AI policy at all

    Down from 24% in a year, but the stated obstacles are unchanged: lack of knowledge (59%), budget constraints (48%) and regulatory uncertainty (41%).

  • 36%

    Already cite ISO/IEC 42001 as a reference

    The AI management system standard entered the list of most cited frameworks, with the NIST AI RMF just behind at 33%. AI governance is becoming an auditable certification.

Source: Stanford HAI · AI Index Report 2026

Our paper on ISO/IEC 42001

The stack

The level only rises if the layer
below it is full.

Layer 1 of 6

AI Foundation

Models, data and infrastructure: in the cloud, in the enterprise, or in-house.

The layer almost nobody builds properly, and the one that decides everything else. A single gateway for any model, routing by cost and capability, data classified and vectorised, and an honest choice between cloud, enterprise and on-premise based on what the data allows , not on what the trend demands.

  • Multi-model gateway: swap the LLM without rewriting the application
  • Routing by cost, latency and how critical the task is
  • Deploy in cloud, enterprise or local (sensitive data never leaves the perimeter)
  • Data pipeline, vectorisation and sensitivity classification
  • Claude
  • GPT
  • Gemini
  • Llama
  • Mistral
  • Bedrock
Layer 2 of 6

Your company intelligence, operated

What your company knows, available to whoever needs it, and to the agents too.

Documentation scattered everywhere, runbooks living in three people’s heads, decisions buried in email threads. We turn that into a living base: continuous ingestion, curation, per-area permissions and reliable retrieval. It is what makes an agent answer with your company context instead of the internet average.

  • Continuous ingestion of docs, wikis, tickets, code and meetings
  • RAG with evaluated retrieval, not merely plugged in
  • Permission by area: each person retrieves only what they may see
  • Curation and detection of stale or conflicting knowledge
  • RAG
  • Embeddings
  • pgvector
  • Reranking
  • Knowledge graph
  • Recall evaluation
Layer 3 of 6

Agents across the whole company

Support, IT, finance, sales, logistics, HR: each area with its own.

Agents that do not stop at chat: they execute. Connected to the ERP, the CRM, the service desk, IoT devices and internal APIs, with granular permissions, human approval where the risk demands it, and an audit trail on every step. One agent per process, not a generic chatbot for everything.

  • Native integration with ERP, CRM, ticketing and IoT
  • Tools with scoped, revocable permissions
  • Mandatory human-in-the-loop for high-impact actions
  • Multi-agent orchestration for processes that cross areas
  • Tool use
  • MCP
  • Orchestration
  • Human-in-the-loop
  • ERP
  • CRM
Layer 4 of 6

Micro SaaS and internal solutions

Software that never paid off before, and now does.

That internal system always stuck in the queue because the return never justified six months of a squad. With AI in the development cycle, it now fits in weeks. We build micro SaaS for your market and internal tools that solve the specific problem, without forcing the company to bend around an off-the-shelf product.

  • Micro SaaS aimed at your market, with your business rules
  • Internal tools that replace the critical spreadsheet and the fragile macro
  • Portals and copilots embedded in the systems the team already uses
  • AI-accelerated development cycle, with human review
  • Micro SaaS
  • Embedded copilots
  • APIs
  • Integrations
  • Process automation
Layer 5 of 6

Control tower

The entire business chain in a single view, for the board and the executive team.

The tower is not the technical team’s dashboard. It is where the board and the chiefs see the company link by link, from demand to cash, with each indicator operated by AI and flagging deviation before it becomes a problem. It works for any business because the chain is the same: retail, manufacturing, services and software change the names of the links, not the shape. The Command Center and the SOC are still there, now as what feeds this view.

  • An indicator per link of the chain, from acquisition cost to cash cycle
  • Consolidation for the board with the origin of every number visible
  • Deviation flagged by the agent, rather than a report to dig through
  • L1/L2 Command Center and 24x7 SOC, both model-assisted
  • Board view
  • Indicator per link
  • Command Center
  • 24x7 SOC
  • AI observability
  • SLA
Layer 6 of 6

Engineering and culture with AI

Your team running it alone, because a transformation that depends on a vendor is not a transformation.

The last layer decides whether the previous one survives. We move the practices inside your team: how to develop with AI without losing review, how to write the evaluation before the prompt, when to trust the agent and when to doubt it. The stated goal is for you to stop needing us for this.

  • AI-assisted development with a defined review standard
  • Evaluation practices: tests before prompts, metrics before deploys
  • AI usage policy and risk criteria by task type
  • Enablement by profile: engineering, operations, business and leadership
  • Enablement
  • AI code review
  • Evals
  • Usage policy
  • Training
  • Playbooks

Keep scrolling to fill the tank

The assessment

AIMI · AI Maturity Index

Before proposing anything, we measure. AIMI shows how ready the company is to use AI with scale, governance, security and real value creation, and exactly where it is stuck today.

  1. L1 Fragmented AI

    Maturity profile at level L1, across the six AIMI pillars

    AI exists, but nobody knows what it returned

    most companies are here
  2. L2 Controlled AI

    Maturity profile at level L2, across the six AIMI pillars

    Rules are in place. Usage has not spread

    most companies are here
  3. L3 Integrated AI

    Maturity profile at level L3, across the six AIMI pillars

    Wired into the process, with results in the metrics

  4. L4 Optimised AI

    Maturity profile at level L4, across the six AIMI pillars

    It became a capability, not a project

Strategy · Data · Platform · Governance · Operation · Value

The polygons are a reading of the AIMI scale definitions, not a market survey: they show the typical SHAPE of each level, where growth is usually lopsided before it evens out. The band where most companies sit is our own finding, measured in assessments. The number for your environment comes out of the calculator below.

  1. L1 0.0 – 1.5

    Fragmented AI

    Experimental

    Isolated proofs of concept, controlled tests, limited business impact and no real scale. AI exists, but nobody can say what it returned.

  2. L2 1.6 – 2.4

    Controlled AI

    Still without scale

    Minimum controls, approvals, security criteria and the first standards in place. The house starts to have rules, but usage has not spread yet.

  3. L3 2.5 – 3.2

    Integrated AI

    Connected to the business

    AI wired into the corporate foundation, with active governance and integration into business processes. The result already shows up in the department’s own metrics.

  4. L4 3.3 – 4.0

    Optimised AI

    Ready to expand

    A scalable operation with real-time metrics, FinOps discipline, clear ROI, automation and expansion across departments. AI has become a capability, not a project.

Six pillars measured · scored 1 to 4

  • Strategy and executive sponsorship
  • Data and knowledge
  • Platform and architecture
  • Governance risk and security
  • Operation and delivery
  • Value and adoption

What AIMI answers

  • Is AI still in pilot, or does it already produce business results?
  • Are governance, security and control mechanisms in place?
  • Are the data and knowledge assets ready and connected?
  • Does the operation measure ROI, risk, adoption and performance?
  • Can the company scale AI across several business areas?

Evidence we ask for before scoring

  • Approved roadmap with executive sponsorship
  • A single source of truth and integration of core data
  • Standardised architecture and observability
  • Policies, auditability, logs and compliance controls
  • Defined deployment, support and ownership processes
  • Savings, productivity gains and adoption measured

“AIMI is the index that shows the distance between an isolated AI experiment and a genuine AI operating model inside the company.”

I want my company AIMI

AIMI calculator

Where you are.
Where you want to get.

Drag the controls and watch the index move. It takes a minute and gives back the read that normally comes after a two-week assessment.

This is a self-assessment, not an audit: the scores come from your own reading of the company. It is meant to rank priorities and show the distance to where you want to be, not to certify maturity. In the AIMI assessment we run with your team, every pillar is backed by evidence.

  • 1

    How to read it Each area tries on its own. There is no owner and no budget line for AI.There is someone accountable and a budget, but still no business target written down.An approved roadmap, sponsored at board level, with a target tied to a business metric.The AI plan is reviewed on the same cycle as the company plan, and budget moves with measured results.

  • 1

    How to read it Data scattered across systems, spreadsheets and inboxes. Nobody knows which version counts.Core data has owners and a register, but integration is still manual or by export.A single source for core data, with a catalogue, lineage and controlled access.Data and documents served through an interface of their own, with measured quality and history ready for a model to query.

  • 1

    How to read it Each initiative picks its own tooling, and nothing carries over to the next one.There is a standard environment to build in, but getting to production is still handcrafted.A common platform, with a delivery pipeline, separated environments and observability.Models, agents and tools come and go by configuration, and switching vendor is not a project.

  • 1

    How to read it Nobody knows who uses AI in the company, on which key, or what has already been sent outside.A written policy exists, but compliance is left to each team.Controls applied on the path of the call, with a record per request and periodic review.Evidence appears on its own, ready for audit, and AI risk sits in the same committee as corporate risk.

  • 1

    How to read it What ships depends on whoever built it. There is no on-call and no rollback plan.A delivery process exists, but support for what is already in production is reactive.A named owner, a contracted SLA, a tested rollback plan and a rehearsed incident response.Model performance, cost and accuracy monitored like any other critical service.

  • 1

    How to read it Return is not measured. Success is the demo having worked.There are numbers from the pilot, but nobody followed them once it became routine.Savings, time gained and adoption measured per use case, against the prior baseline.The return lands in the budget of the area that uses it, and a use case that does not pay is switched off.

US$

Maturity radar across the six pillars

  • Where we are
  • Target

AIMI index today

1,0

L1 Fragmented AI

  • L1Fragmented AI
  • L2Controlled AI
  • L3Integrated AI
  • L4Optimised AI

Reading

Validate this index with us

The company brain

Data in a silo answers a team.
Related data answers the business.

Every question a director actually asks cuts across three or four systems that never spoke to each other. That is why the answer took a week and arrived stale. Once the whole company’s knowledge becomes a single mesh, with per-area permissions preserved, that same question is answerable in minutes, with the evidence attached. That is where AI touches revenue: not by writing better, but by shortening the distance between noticing and deciding.

An illustration of a company knowledge mesh: 10 data domains, each with its own cloud of records, and 12 lit bridges connecting different domains. Each bridge represents a business question that can only be answered once two systems that never spoke to each other start to.

ERP and finance Sales and CRM E-commerce Logistics Service desk Infrastructure Security People Marketing Documentation
  • What the company already has records, documents and logs, each in its own system
  • What only exists once related the business question that dies inside a silo

The mesh is neutral and only the bridges are lit, and that is the argument of the figure: the value is not in having the data, it is in the data touching. This illustrates the domain structure almost every company has; it is not a screenshot of a client. The placement of each node is arrangement, not measurement.

One mesh, six readings

Each seat asks a different thing, and each answer crosses different domains. It is the same base: what changes is the slice and the permission.

  • Board

    Board

    Is the plan holding, and what changed since the last meeting?

    • ERP and finance
    • Sales and CRM
    • Marketing
    before
    a pack closed the month before
    with the mesh
    the same reading, on yesterday’s data

  • CEO

    CEO

    Why did margin drop, and exactly where?

    • ERP and finance
    • Logistics
    • Sales and CRM
    before
    three teams extracting, a week to reconcile
    with the mesh
    the breakdown on screen, with the number behind it

  • COO

    Operations director

    Where does the order stall before it ships?

    • E-commerce
    • Logistics
    • Service desk
    before
    a spreadsheet per carrier, refreshed on Monday
    with the mesh
    the bottleneck flagged while it is happening

  • CFO

    Finance director

    What does it really cost to serve each customer?

    • ERP and finance
    • Logistics
    • Service desk
    before
    averaged allocation, which hides the customer who loses money
    with the mesh
    cost per customer, with freight, returns and tickets inside

  • CTO

    Technology director

    Which change caused this incident?

    • Service desk
    • Infrastructure
    • Documentation
    before
    correlation held in the memory of whoever was on call
    with the mesh
    the deploy window and the ticket linked automatically

  • CISO

    Security director

    Should this access still exist?

    • Security
    • People
    • Infrastructure
    before
    a quarterly review, done from an export
    with the mesh
    the gap between headcount and permissions, flagged the same day

The loop, and where the person comes in

The machine reads, relates and proposes. A person still decides, and what was decided returns to the mesh. Full autonomy exists only where it was explicitly delegated, and never for sensitive, financial or access actions.

  1. 01

    Relates

    The mesh links system data, documents and conversations, with per-area permissions preserved.

  2. 02

    Observes

    Specialist agents read the mesh continuously and compare it against what was expected.

  3. 03

    Proposes

    Whatever departs from the expected becomes a hypothesis with evidence attached, not a bare alert.

  4. 04

    Decides

    A person approves, adjusts or refuses. Sensitive, financial or access actions are never automatic.

    human
  5. 05

    Executes

    What was approved runs in the systems of record, with every step logged.

  6. 06

    Learns

    The outcome returns to the mesh, and the next proposal is born with that feedback inside.

10 domains, 12 bridges drawn, and per-area permissions that survive the trip: whoever could not see the data in the system of record still cannot see it in the answer.

Control tower

The board does not need a report.
It needs the chain, live.

The tower is not the technical team’s dashboard. It is where the board and the executive team see the whole company, link by link, with each indicator operated by AI and flagging deviation before it becomes a problem. The Command Center and the SOC still exist, and they are now what feeds this view rather than what it is.

What reaches the board

The chain, link by link

  1. 01

    Demand

    attract and qualify whoever can buy

    • Acquisition cost CMO AI-operated
    • Lead quality CMO AI-operated

    also called marketing, lead generation, traffic

  2. 02

    Sales

    turn interest into an order, within margin

    • Stage conversion CRO AI-operated
    • Quote margin CFO Hybrid

    also called commercial, funnel, e-commerce, tenders

  3. 03

    Supply

    have something to sell, without money parked in stock

    • Stockout and overstock COO AI-operated
    • Replenishment cost CFO Hybrid

    also called purchasing, contracts, capacity, licences

  4. 04

    Operations

    produce or execute what was sold

    • Availability CTO AI-operated
    • Rework COO Hybrid

    also called factory, project, platform, service

  5. 05

    Delivery

    get it to the customer on the agreed date

    • On-time delivery COO AI-operated
    • Cost per delivery CFO AI-operated

    also called logistics, rollout, go-live, provisioning

  6. 06

    After-sales

    keep the customer, and find out early when they are leaving

    • Churn risk CRO AI-operated
    • Cost to serve CFO Hybrid

    also called support, customer success, warranty, renewal

  7. 07

    Finance

    close the books and say whether the month worked

    • Margin by line CFO Hybrid
    • Cash cycle CFO Human

    also called controlling, billing, collections

Who operates each number

  • AI-operated the agent computes it, compares it against expectation and flags the deviation. A person looks at what was flagged, not at the whole spreadsheet.
  • Hybrid AI prepares and proposes, a person confirms. This is where the numbers that depend on judgement or on a contract live.
  • Human still decided by people, with AI only organising the material. Nobody automates what answers to the auditor.

The chain above is deliberately generic, and that is what lets the same tower serve different businesses: retail, manufacturing, services and software change the names of the links, not the shape. Of the 14 indicators drawn, 8 are operated end to end by AI, and the rest still go through people.

In numbers

What changes when the operation
stops fighting fires.

What left human hands

the agent does it a person does it

  • Time to resolve an incident MTTR
    −84%

    Triage and runbook executed by the agent before anyone is paged

  • Alerts that reach a human operational noise
    −82%

    Automatic deduplication and correlation with deploys and changes

  • Human effort on repetitive procedures manual effort
    −72%

    Runbook executed by the agent, with a person stepping in only on exceptions

  • Audit preparation effort certification
    −60%

    Evidence collected continuously, rather than assembled in audit week

  • Level 1 tickets service desk
    −54%

    Agents handling the repetitive work before it becomes a ticket

And what improved another way

  • Support NPS satisfaction
    +80%

    Availability and accuracy in the fix: solved first time, and solved fast

  • Monthly cloud spend AWS · Azure · GCP
    −41%

    Continuous rightsizing, reservation review and shutdown of idle capacity

And where the change is an order of magnitude

  • Root cause analysis

    25 to 30 min 60 seconds

    25 to 30× faster

    The agent correlates alerts, changes and logs across the layers involved before anyone opens a terminal

  • Wait time in support

    over 30 min under 1 min

    over 30× less waiting

    The problem is solved before it becomes a queue, and that is what lifts NPS

Ranges observed across our projects. The number for your environment comes out of the assessment.

See the cases

Autonomy with control

No agent starts out unsupervised.
It earns its autonomy.

Every agent starts out only recommending. It moves up a stage when the accuracy rate on its specific process justifies it, and it never moves up on sensitive, financial, legal or access-related actions.

An agent flow: it receives the alert, investigates, handles the low-risk part on its own, and stops to ask for human approval on anything high-impact. Approved, it runs in production. Declined, it cancels and goes back to recommending.

Alert: disk full production, 03:12 Agent investigates reads logs, correlates, plans Low-risk action clears rotated logs on its own High impact? expand a production volume ? Production executes and records Declined cancels and records yes no back to recommending

The agent stopped here and is waiting. What do you do?

  1. Assistive

    The AI recommends the action. A human executes it.

    The starting point for every new agent: the model suggests, the person decides. It exists to calibrate accuracy before any automation.

  2. Semi-autonomous

    The AI executes approved steps, with human validation at control points.

    The agent drives the flow and stops at defined checkpoints. This is the stage where most processes settle.

  3. Controlled autonomy

    The AI runs bounded processes end to end, inside policy, risk and audit.

    Reserved for well-understood processes, with fallback, rollback and complete logging. Sensitive, financial or access-related actions still require human approval.

What every agent carries, no exceptions

Before any agent reaches production, these six items are declared and tested. This is not a good-intentions checklist: it is what separates auditable automation from a loose bot on the corporate network.

  • The tools and systems it is allowed to reach
  • The actions it is allowed to perform
  • Confidence thresholds and exception triggers
  • An audit log and observability for every decision
  • Fallback and rollback mechanisms
  • Mandatory human approval on sensitive, financial, legal or access actions

Agents by area

What they actually do.

A top view of an operation running on AI. Each room is a sector, each label is an agent, and the number of people orchestrating is stated in every room: AI does not empty the room, it changes who does what inside it.

Office floor plan · top view

41 agents 16 people 10 sectors

An office floor plan seen from above, with one sector per room. Every workstation has a lit monitor and the figure of whoever operates it: people, with round heads, and AI agents, with square heads and antennae. There are meeting rooms, a kitchen, stairs and a central corridor.

circulation Meeting room Kitchen Meeting room Stairs Open area A-01 Support and service desk A-02 Infrastructure and NOC A-03 Security and SOC A-04 Data and knowledge A-05 Finance A-06 Sales and pre-sales A-07 Logistics and store operations A-08 People A-09 Engineering T-00 Control tower N

Of the total occupancy, 72 per cent are AI agents and 28 per cent are people.

57 stations
  • 72% AI agents 41 stations
  • 28% people 16 stations
  • control tower runs the other rooms

Each room is a sector, and the occupancy of each one is the composition declared in the data: one station per agent, one per person orchestrating. These are types of agent we build, not a client inventory.

AI-first service desk

The service desk stops being a queue
and becomes a flow.

AI does not show up here as a chatbot. It becomes an operational layer inside the service cycle, helping the organization receive, understand, prioritize, route, support and resolve faster and more consistently.

4 AI executes 2 AI proposes and stops 3 A person decides

010203040506070809
AI executes human audits afterwards
01

Monitoring and event

Deduplicates, correlates with deploys and changes, and drops what is known noise.

02

Service desk

Records, classifies, sets priority and asks for what is missing before opening.

03

Service request

Fulfils anything in the catalogue end to end, within the authorized scope.

07

Knowledge management

Writes the article from the resolved ticket and flags what went stale.

AI proposes and stops waiting for approval
04

Incident management

Investigates, drafts the plan and runs runbooks. Stops for approval on anything destructive.

08

Service level

Measures, flags what fell out of band and writes the report. The SLA conversation is human.

A person decides AI prepares the material
05

Change enablement

Prepares the record, the rollback plan and the risk assessment. The board approves.

06

Problem management

Groups recurring incidents and gathers the evidence. Root cause is a human conclusion.

09

Continual improvement

Surfaces the pattern nobody had seen. What enters the backlog is the team decision.

The order follows a ticket from intake to closure. What moves up and down between the lanes is the degree of autonomy: it is decided practice by practice, not once for the whole operation. Change and problem stay with people on purpose, because both end in judgement rather than execution.

  1. Intake and understanding

    The AI receives the request through a portal, email, chat, form or collaboration tool. It interprets the intent, extracts the relevant details, detects signs of urgency and builds a structured understanding of the case.

  2. Classification and triage

    Based on content, history, knowledge base, business rules and SLA definitions, it classifies the ticket, determines category and severity, identifies sensitive cases and decides the next step.

  3. Using the knowledge

    It searches the approved internal knowledge (procedures, policies, previous resolutions) to propose the best action. It drafts the reply to the user, recommends diagnostic steps and summarises what is complex.

  4. Routing and orchestration

    It hands the case to the right team, queue or resolver group. Integrated with corporate systems, it triggers approved actions: opens derived requests, enriches the case with context and asks for the missing information.

  5. Accelerating resolution

    On repetitive or low-risk requests, it resolves or automates part of the path. On complex cases, it accelerates the analyst with a summary, a suggested next step and a structured handover note.

  6. Continuous improvement

    The model learns from ticket outcomes, quality ratings, resolution times, escalation patterns, knowledge gaps and SLA breaches, refining classification, coverage and routing.

Where the model runs

Any model. Wherever your data allows.

We do not resell any AI vendor, so the model choice is technical: whatever solves the task at the lowest cost, in the place the data is allowed to be. Switching models later is configuration, not a rewrite.

Cost, capability and objective

22 models in operation, 11 of them run inside your walls

125× between the expensive end and the cheap one

A map of twenty-two language models, with cost per million tokens on the horizontal axis, on a logarithmic scale, and four capability tiers on the vertical axis: volume and latency, work at scale, the workhorse, and frontier. Each tier states what it is for. Colour separates three licences: proprietary API-only, open-weight with a restrictive licence, and open source with a permissive one.

Volume and latency cost per call is what decides Work at scale has to think, but not much The workhorse where almost every production agent lives Frontier long reasoning and hard decisions US$ 0,1 US$ 0,3 US$ 1 US$ 3 US$ 10 cost per million tokens · logarithmic scale more expensive → capability Phi-4 Gemma 3 Gemini Flash-Lite GPT-5 nano Mistral Small Claude Haiku Llama 4 Scout Gemini 2.5 Flash Llama 3.3 70B DeepSeek V3 Qwen 2.5 72B GPT-5 mini Llama 4 Maverick DeepSeek R1 Command R+ Mistral Large Claude Sonnet Grok Gemini 2.5 Pro GPT-5 o3 Claude Opus

Colour is the licence, and it decides whether a model can enter your datacenter: proprietary exists only through an API; open-weight downloads and runs, but the licence restricts commercial use; open source is Apache 2.0 or MIT and runs anywhere. The horizontal axis is order of magnitude from public list prices, not a price table: price changes often, varies by region, contract and volume, and drops with every generation. The vertical axis is an ordered tier, not a benchmark score: between neighbouring models the ranking shifts with the benchmark, but the tier holds, and it is the tier that decides a project. Height within a tier is only arrangement so the names do not stack. A new model ships every week; the shape of the map is what does not change.

total cost

Total cost as usage grows, in three zones. In the first, public cloud is cheapest; in the second, enterprise starts to pay off; in the third, on-premise becomes cheapest. The vertical axis has no scale.

usage, over time Start Public cloud Predictable Enterprise High and steady Local / on-premise Public cloud Enterprise Local / on-premise

Hover the chart, or use the arrow keys, to see the cost ranking at each moment

The vertical axis has no scale, and that is deliberate: price depends on model, region, contract and volume, and an invented number here would be worth less than nothing. The figure claims the shape of the cost, where it flips, and what the floor is when the choice is right at each moment. The number for your case comes out of the assessment.

What cost says, and what strategy says

  • Getting started

    pilot, irregular volume, first use case

    cost says Public cloud is cheapest here
    strategy says If the task touches sensitive data, local is right from day one. Migrating later costs more than starting right.

    strategy wins

  • Predictable volume

    steady usage, legal asking for a contract

    cost says Enterprise starts to pay off
    strategy says It is also when contract and isolation start being required. Cost and strategy point the same way, and the decision is easy.

    both agree

  • High, steady volume

    mature operation, usage that does not drop

    cost says Local becomes cheapest
    strategy says It only works if someone runs it. Hardware without an on-call team is savings on paper and downtime in practice.

    strategy limits it

  • Public cloud

    Bedrock, Azure OpenAI, Vertex, direct API

    the data
    leaves the perimeter, in the region you choose
    the bill
    drops to nothing when idle, grows with usage
    typical models
    Claude · GPT · Gemini

    decides when speed matters more than control

  • Enterprise

    corporate contract with isolation

    the data
    leaves, but stays out of vendor training
    the bill
    a commitment at the base, usage on top
    typical models
    Claude · GPT · Gemini

    decides when legal needs a signed contract

  • Local / on-premise

    open model in your datacenter

    the data
    never leaves, and that is why it exists
    the bill
    expensive to build, cheap to use heavily
    typical models
    Llama · Mistral · Qwen

    decides when the data cannot leave, full stop

Hybrid In practice it is what almost everyone ends up running, and it is not indecision: it is the correct reading that the three zones coexist inside the same company. Public work runs cheaply in the cloud, sensitive work stays inside, and the routing is a classification rule rather than a case-by-case choice.

Claude, GPT, Gemini, Llama, Mistral, Qwen and self-hosted open models, through Bedrock, Azure OpenAI, Vertex, direct API, vLLM or Ollama.

AI governance The gateway is what makes that true. One door for every AI call in the company: a virtual key per application, a spend ceiling per team, data policy applied before sending, and an audit trail. We are implementation specialists, with LiteLLM, inside your own infrastructure. See the LLM gateway

Non-negotiable principles

These are not guidelines.
They are operating requirements.

Every practice, every initiative and every deploy in the program has to line up with these four principles. When one of them cannot be met, the initiative does not ship, and we say so upfront, not afterwards.

Security and privacy first

The NIST AI RMF is the baseline. Security is a starting requirement, not a last-minute consideration: masking, anonymisation, personal data protection and defence against adversarial attack go in on day one.

NIST AI RMF

Accountability and FinOps

Token, model and API costs managed with rigour and transparency. Every initiative has to demonstrate clear business value and be tracked with FinOps practice so it can scale sustainably.

FinOps · Value-driven

Governance and compliance

The platform aligns with IT service management practice, AI governance and responsible data use. Who decides, who executes and who audits are all defined. Data protection compliance is not negotiable.

ITIL · ISO/IEC 42001 · LGPD

Agility with alignment

Agile method turns experiment into production capability, supported by market benchmarks and proven references. Innovation with speed, but always anchored in evidence and prioritisation.

Agile · Evidence · Prioritisation

Translated into technical controls

  • Every tool the agent can reach is declared, scoped and revocable
  • Destructive or high-impact actions stop and ask for human approval
  • Every run is recorded: input, reasoning, tool, result and who approved it
  • Classification of what may leave the perimeter, with automatic routing
  • Test sets built from real cases run on every prompt or model change
  • Quota and cost ceiling per team and use case, with alerting and automatic cutoff

How we run it

From the assessment to your team running it alone.

Timelines from programs we have already run. Each phase has a defined deliverable and an exit point: you can stop at any of them without leaving half a solution running. A first six-month cycle is usually the right horizon: long enough to prove value on a real case, not just to produce a plan.

  1. Assessment and map

    2 to 3 weeks

    We run AIMI: maturity measured across the six pillars, processes, data and systems mapped, and the six-layer stack applied to your company: what already exists, what is missing, and the order to build it in. It includes the use cases we recommend not doing.

    Deliverable AIMI index, prioritized backlog and business case

  2. Foundation

    4 to 8 weeks

    Model gateway, data pipeline, usage policy, cost control and observability. It is the layer nobody wants to pay for and without which everything after it becomes rework. So we deliver it alongside the first use case, and value shows up right away.

    Deliverable A working AI platform plus the first use case in production

  3. Scale area by area

    6-week cycles

    Knowledge and agents go live area by area, in short cycles. Every cycle has a success metric agreed before it starts and an exit point: you can stop whenever you want without leaving half a solution running.

    Deliverable Agents in production per area, with audited metrics

  4. Control tower and autonomy

    continuous

    Everything moves to the control tower, under SLA. In parallel we transfer the practices to your team, with a deadline and an agreed exit criterion. A transformation that depends on the vendor forever is not a transformation.

    Deliverable Operation under SLA and a team able to run it without outside help

We operate as one delivery team: shared accountability with the business and technology areas, and knowledge transfer so the organization can scale on its own, not as a vendor that hands over a deck and leaves.

How we measure

The program is not measured by
tools shipped. It is measured by impact.

These are the executive indicators we agree on before starting. The values reflect the average result observed over twelve months in organizations running a structured AI operating model with governance discipline.

  • +10%

    Revenue increase

    From AI-assisted selling, faster delivery, a better customer journey and new capabilities the technology enables.

  • 20%

    Savings realised

    From automation, reduced manual effort, tool rationalisation and greater efficiency in service operations.

  • 41%

    Efficiency gain

    In speed, throughput, consistency and execution capacity, driven by assistants and autonomous operations.

The AI OKRs we track

  • Adoption and active use of the approved AI capabilities
  • Number of AI initiatives in production with measurable value
  • Efficiency gain in the targeted workflows and service operations
  • Reduction of manual effort in the selected processes
  • Financial impact delivered, in revenue or in cost
  • Adherence to security, trust and compliance in what is delivered

What we say upfront, not afterwards

  • Business and technology teams will need structured support, training and a clear path to capability in order to move from curiosity to practical adoption.
  • There will be a need to balance speed and guardrails: allowing experimentation without losing control.
  • Some use cases prove value quickly; others require more groundwork in data, systems or process clarity.
  • The organisation will need visible executive sponsorship and recurring communication to keep the pace.
  • Long-term advantage comes from building repeatable capability, not from an isolated pilot.

Reference lenses we apply

  • NIST AI RMF
  • ISO/IEC 42001
  • ITIL 4
  • LGPD
  • McKinsey State of AI
  • PwC Responsible AI
  • Deloitte AI Governance
  • BCG Workforce Transformation

In production every day

Shall we find out which layer your company is on?

A technical session at no cost: we apply AIMI to your environment and hand back the six-layer read for your case, with the order of priority and what is not yet justified.

Talk to a specialist +55 11 4193 5660