90-day roadmap for implementing AI

A 90-day AI roadmap should turn an operational need into a measurable use case, validate data, establish controls, deploy a pilot, and demonstrate its financial impact. The goal is not to transform the entire company in three months, but to generate sufficient evidence to decide whether to scale up, correct, or stop.

Taxonomic definitions

1. AI Value Discovery. A structured process to identify and prioritize artificial intelligence use cases based on their economic impact, technical feasibility, data availability, level of risk, and adoption capability. Its result is not a list of ideas, but an ordered portfolio of opportunities with financial assumptions.

2. Minimum viable product of AI. Controlled version of an artificial intelligence solution that uses real data, users, and processes to verify whether the use case produces repeatable results. It should include metrics of accuracy, adoption, productivity, cost, and risk; a single isolated demonstration does not constitute a viable minimum viable product.

3. AI government. Set of policies, responsibilities, controls, records, and oversight mechanisms that determine how artificial intelligence systems are selected, developed, used, monitored, and retired. Includes data protection, human validation, traceability, security, intellectual property, and vendor management.

Comparison table: roadmap for implementing AI in 90 days

Stage	Period	Business objective	Mandatory deliverables	Indicator of progress	Executive decision
Lineup	Days 1-15	Defining the financial and operational problem	Sponsor, baseline, scope, risks and responsibilities	Case associated with cost, revenue, risk, or labor capital	Continue or discard
Prioritization	Days 16-30	Selecting a viable case	Use case matrix, data evaluation and business case	Expected value higher than the total cost and usable data	Authorize pilot
Construction	Days 31-60	Check the technical feasibility	Integrated pilot, tests, controls and dashboard of metrics	Quality and performance within agreed thresholds	Switching to controlled operation
Deployment	Days 61-75	Validate use under real conditions	Pilot group, training, support and monitoring	Adoption, time savings, exceptions and satisfaction	Expand, correct or stop
Evaluation	Days 76-90	Demonstrate economic value and scalability	ROI, payback, residual risks and expansion plan	Verifiable benefit and stable process	Scaling, redesigning or closing

Why implement AI in 90 days

The adoption of AI is no longer the main differentiator. The challenge is to turn its use into financial impact.

Stanford’s AI Index 2026 indicates that organizational adoption of AI reached 88%. However, McKinsey reported that nearly two-thirds of organizations had not yet initiated a business scaling initiative and only 39% attributed some impact on EBIT to the use of AI. Most projects remain caught between experimentation and operational transformation.

A 90-day roadmap addresses that gap. It establishes a period long enough to test the solution in operation, but short enough to avoid open-ended programs, indefinite budgets, and pilots that survive without demonstrating a return.

At the end of the 90 days, the management must be able to answer five questions:

• What economic problem has AI solved?
• How much value has it produced or can it produce?
• What data do you need to operate?
• What risks does it introduce?
• Should it be scaled up, corrected, or stopped?

The inability to answer them means that technology has been built, but not business capacity.

Principles to avoid an AI pilot without return

Starting with a result account, not with a tool

The first document of the project should not be a technological architecture. It should be a value hypothesis.

Each use case must be associated with at least one of these results:

• Reduction in operating hours.
• Reduced errors or rework.
• Reduction of cycle time.
• Prevention of losses, fines or interruptions.
• Increase in conversion, retention or margin.
• Release of capacity for higher-value activities.
• Reduction of inventory or working capital.
• Acceleration of attention and resolution.

“Using generative AI” is not an outcome. “Reducing 25% the time spent classifying and responding to requests” can be measured, audited, and translated into money.

Choosing a process, not a flashy task

The most difficult projects to scale usually automate a single activity without changing the entire flow.

For example, generating an AI response is of little use if it is then:

• A collaborator must copy it manually to another system.
• There are no reliable data available to customize it.
• Approval takes longer than drafting.
• There is no record of the decision.
• The exception still requires multiple emails.
• The solution does not integrate with CRM, ERP, ITSM or internal systems.

The participants with the highest financial impact in the McKinsey survey were also those who reported greater workflow redesign, executive engagement, KPI tracking, and the integration of AI into business processes.

Limit the scope to increase the evidence

During the first 90 days it is recommended:

• Evaluate between 10 and 20 opportunities.
• Preselect from two to four cases.
• Building a lead pilot.
• Working with a limited group of users.
• Using a defined set of data.
• Establishing exit criteria before development begins.

A small pilot with reliable metrics has more value than five demonstrations without adoption or financial responsibility. The old discipline of measuring before expanding still works; even AI hasn't managed to retire it.

Days 1 to 15: Aligning business, risks, and responsibilities

1. Name the sponsor and the owner of the process

Every project must have two distinct figures:

• Executive sponsor: eliminates obstacles, authorizes resources, and is responsible for organizational priorities.
• Owner of the process: It knows the operation, validates results, and accepts or rejects changes in the workflow.

TI can build or integrate the solution, but it should not define by itself whether the result generates value. The area responsible for the affected results state should share ownership of the project.

2. Documenting the current problem

The baseline should record the performance before AI.

At least:

• Monthly volume of transactions.
• Average time per activity.
• Number of people involved.
• Error rate.
• Percentage of rework.
• Total cycle time.
• Average cost per transaction.
• Incidents, returns or exceptions.
• Lost revenue or avoidable costs.
• Current level of service.
• User or customer satisfaction.

Without a baseline, any subsequent improvement will be an opinion.

3. Define the financial hypothesis

The hypothesis must take a verifiable form:

If we apply AI to the process X, We expect to reduce or increase the indicator Y from their current baseline to the goal Z, generating an estimated impact of $N in a certain period of time.

Example:

If we automate the classification and preparation of responses for service requests, we expect to reduce the average time for attention from 18 to 12 minutes, freeing up 1,200 hours per year and reducing the operational cost of the process.

4. Create the initial risk register

Before selecting technology, it is necessary to document:

• Personal or confidential data involved.
• Decisions that may affect customers or employees.
• Need for explanation or evidence.
• Risk of incorrect answers.
• Supplier department.
• Intellectual property exhibition.
• Impact of an unavailability.
• Systems with which the integration will be carried out.
• Interventions that require human approval.

The NIST AI Risk Management Framework recommends incorporating reliability considerations during the design, development, use, and evaluation of AI systems. Its specific profile for generative AI adds actions to manage the particular risks of these models.

Available by the 15th

• Executive summary of the case.
• Sponsor and owner of the process.
• Operational and financial baseline.
• Value hypotheses.
• Initial risk inventory.
• Scope and exclusions.
• Preliminary criteria for success.

Decision gate

The case should not proceed if:

• It has no business owner.
• There is no information available to build the baseline.
• The expected benefit cannot be quantified.
• Its use requires data without authorization.
• The operation has no capacity to participate.
• The potential risk exceeds the expected benefit.

Days 16 to 30: prioritize the use case and prepare the data

1. Construct a prioritization matrix

Each opportunity must be rated on a common scale.

Criterion	Evaluation question	Suggested peso
Financial impact	How much does cost, income, risk or capital affect?	25%
Data feasibility	Do the data exist, are they accessible, and are they of sufficient quality?	20%
Technical feasibility	Can it be integrated with the current architecture?	15%
Time to value	Can it produce evidence within 90 days?	15%
Risk	What consequences would an error have?	15%
Adoption	Do users have incentives and the ability to use it?	10%

Prioritization should favor cases with high impact, available data, and controllable risk. It does not necessarily have to choose the project with the highest potential value, but rather the one that can generate the best evidence adjusted for risk.

2. Evaluate the data preparation

The review must cover:

• Owner of each source.
• Age and frequency of updates.
• Empty or duplicate fields.
• Consistency between systems.
• Historical coverage.
• Existing labels or classifications.
• Use permits.
• Personal or restricted information.
• Representation biases.
• Extraction and update mechanisms.

The data need not be perfect, but they must be sufficiently reliable for the selected case.

If the project requires 95% accuracy and the available data only allows reaching 70%, the problem is not corrected by buying a more expensive model.

3. Defining the validation method

Before construction begins, it is necessary to agree on the following:

• Test set.
• Minimum number of cases.
• Answer considered correct.
• Error tolerance.
• Situations that require human intervention.
• Method for registering exceptions.
• Responsible for approving results.
• Frequency of reevaluation.

4. Prepare the business case

The total cost should consider:

• Consulting and design.
• Licenses.
• Consumption of models.
• Infrastructure.
• Integrations.
• Data preparation.
• Security.
• Training.
• Support.
• Monitoring.
• Managing change.
• Annual maintenance.

AI can reduce manual labor and, at the same time, create new costs of integration, supervision, and consumption. Ignoring them inflates the ROI in PowerPoint and deflates it in production.

Available by the 30th

• Prioritized matrix of use cases.
• Selected case.
• Data evaluation.
• Preliminary architecture.
• Business case.
• Validation plan.
• Barriers to success and arrest.

Days 31 to 60: build and validate the pilot

1. Design the solution around the workflow

The architecture must specify:

• User who starts the process.
• Data source.
• Model or automation engine.
• Consulted systems.
• Business rules.
• Points of approval.
• Activity registration.
• Handling exceptions.
• Scaling up to a person.
• The result is returned to the operating system.

2. Incorporate controls from the development stage

The minimum controls include:

• Access based on roles.
• Separation between environments.
• Encryption of information.
• Registration of instructions and responses.
• Filters for confidential data.
• Validation of sources.
• Limits of autonomy.
• Human review for sensitive decisions.
• Version management.
• Shutdown mechanism or rollback.

In 2025, McKinsey reported that 51% of the participants belonging to organizations using AI had observed at least one negative consequence; nearly a third noted consequences related to inaccuracy. Stanford recorded 362 documented AI incidents, compared to 233 in 2024. These data justify integrating controls before deployment, not after the first incident.

3. Perform technical and business tests

The tests must measure:

• Accuracy.
• Relevance.
• Response time.
• Availability.
• Cost per transaction.
• Special rate.
• Hallucinations or responses without foundation.
• Proper use of fonts.
• Compliance with rules.
• Resistance to malicious instructions.
• Perceived quality by the user.
• Stability of integrations.

4. Compare AI with the current process

The solution must be evaluated against:

• Manual operation.
• Traditional automation.
• Fixed rule.
• Conventional search.
• AI-assisted process.
• Process with greater autonomy.

AI will not always be the right alternative. In deterministic and stable tasks, a traditional rule or automation can be cheaper, more explainable, and more sustainable.

Available by the 60th day

• Minimum viable product.
• Basic integrations.
• Test results.
• Updated risk register.
• Technical documentation.
• Flow of exceptions.
• Initial KPI dashboard.
• Recommended controlled deployment.

Days 61 to 75: deploy under controlled conditions

1. Select the pilot group

The group should represent the actual operation, but limit exposure.

It is advisable to include:

• Experienced users.
• Users with less knowledge of the process.
• A supervisor.
• The owner of the process.
• Technical support.
• Risk or safety manager, when applicable.

2. Training on use and limits

The training should explain:

• What the tool solves.
• What should not be solved.
• What information can be used.
• How to verify a response.
• When to climb a person.
• How to report an error.
• Who is responsible for the final decision?.
• What activity is recorded.

Adoption should not be measured by the number of licenses issued, but by the frequency of correct usage and the percentage of processes completed with acceptable results.

3. Operate with reinforced monitoring

During this stage it is recommended to check daily:

• Transactions processed.
• Responses rejected.
• Exceptions.
• Security incidents.
• Cost of consumption.
• Time saved.
• Manual corrections.
• User comments.
• Integration failures.
• Cases where human validation was omitted.

4. Avoid automating the error

A poor process does not improve by being executed faster.

When the pilot reveals contradictory rules, incomplete data, or unnecessary approvals, the team must redesign the workflow before expanding automation.

Available for delivery on the 75th day

• Active users.
• Evidence of adoption.
• Comparison against baseline.
• Error and exception log.
• Real operating costs.
• Adjustments to the workflow.
• Stabilization plan.

Days 76 to 90: demonstrating ROI and deciding on scaling

1. Calculate the observed benefit

The results should be separated into four categories:

Direct savings

• Hours deducted.
• Lower cost per transaction.
• Reduction of external services.
• Consolidation of tools.
• Less rework.

Capacity released

• Hours reassigned to activities of greater value.
• Higher volume of cases handled without increasing staff.
• Decrease in backlog.
• Lower dependence on specialists.

Incremental income

• Higher conversion rate.
• Reduction of drop-out rates.
• Less launch time.
• Increase in availability.
• Higher retention.

Risk avoided

• Fewer mistakes.
• Less penalties.
• Reduction of incidents.
• Greater compliance.
• Evidence for audit.

The freed capacity should not be automatically counted as savings. It becomes a financial benefit when it reduces future hiring, eliminates overtime, increases volume, or is reallocated to an activity that produces verifiable value.

2. Compare results against thresholds

The committee should review:

• Improvement compared to baseline.
• Precision.
• Adoption.
• Unit cost.
• Special rate.
• Residual risks.
• Time to recover from failures.
• User satisfaction.
• Annualized benefit.
• Revenge.
• Capacity for integration.
• Effort to climb.

3. Making an explicit decision

By day 90 there should only be four possible decisions:

• Scaling: There is proven value and controlled risk.
• Optimize: The case has potential, but it requires adjustments.
• Redesigning: the technology works, but the process or the data do not.
• Stop: the benefit does not justify the investment or the risk.

Closing a case without a return to court does not represent a failure. Keeping it alive to protect internal enthusiasm does destroy budget and credibility.

Available for delivery on day 90

• Executive report of results.
• Observed and projected ROI.
• Estimated payback.
• Residual risks.
• Lessons from the pilot.
• Objective architecture.
• Operational model.
• Prioritized backlog.
• Scaling budget.
• Formal decision of the committee.

How to calculate the financial impact of AI

Annual gross profit

Annual gross profit = verifiable labor savings + incremental income with margin + avoided losses + costs of substituted technologies

Net profit

Net profit = gross profit − implementation costs − operational costs − adoption costs − cost of controls

ROI

ROI = [(accumulated net profit − investment) / investment] × 100

When the investment has already been included in the net profit, it should be avoided to subtract it a second time.

Recovery period

Payback in months = initial investment / monthly net profit

Illustrative example

An operation with 12 analysts automates two hours per person per day:

• Integrated labor cost: $450 MXN per hour.
• Effective adoption: 70%.
• Working days: 220 per year.
• Estimated annual gross profit: $1,663,200 MXN.
• Annual investment and operation: $900,000 MXN.
• Estimated net profit: $763,200 MXN.
• Estimated annual ROI: 84.8%.

This calculation is illustrative, not a benchmark. In a real case, the time for human review, exceptions, cost of consumption, integrations, and the adoption curve must be taken into account.

Indicators that must reach the executive committee

Financial indicators

• Gross and net profit.
• ROI.
• Revenge.
• Cost per transaction.
• Cost per active user.
• Expected impact on EBITDA.
• Incremental income.
• Avoided losses or fines.

Operational indicators

• Cycle time.
• Productivity per person.
• Processed volume.
• Backlog.
• Rate of rework.
• Availability.
• Compliance with SLA.
• Time to resolve.

Indicators of the model

• Precision.
• Responses rejected.
• Rate of hallucination.
• Exceptions.
• Derivations to humans.
• Quality of the sources.
• Latency.
• Cost per inference.

Indicators of adoption

• Active users.
• Frequency of use.
• Percentage of the process executed with AI.
• Use outside of politics.
• Satisfaction.
• Manual corrections.
• Abandoning the tool.

Risk indicators

• Incidents.
• Unauthorized access.
• Data display.
• Errors with economic impact.
• Decisions reversed.
• Audit findings.
• Supplier's department.
• Changes in the model's performance.

Government and minimum controls

Central inventory of solutions

The company must register:

• System name.
• Owner.
• Supplier.
• Models used.
• Data sources.
• Authorized users.
• Purpose.
• Risk level.
• Checks.
• Cost.
• Date of revision.
• Status: Pilot, production, restricted or withdrawn.

Classification by risk level

A practical classification could consider:

• Low risk: Internal support without sensitive decisions.
• Moderate risk: operational recommendation revised by a person.
• High risk: decisions that affect money, employment, credit, health, safety or compliance.
• Unacceptable risk: use without authorization, supervision, traceability, or legal basis.

Human-in-the-loop

Human intervention should be defined by risk, not by custom.

It must be established:

• Which results are reviewed.
• Who approves them.
• What evidence do you consult?.
• How much time does he have to decide?.
• How to document a correction.
• When the solution can be suspended.

Continuous monitoring

The approval of day 90 is not permanent. Models, data, costs, and behaviors can change.

Periodic reviews are needed to detect:

• Model drift.
• Data drift.
• Cost increases.
• Changes in the provider.
• New vulnerabilities.
• Prejudices.
• Adoption failure.
• Non-compliance with policies.
• New regulatory requirements.

Team responsible for the roadmap

Rol	Primary responsibility
Executive Sponsor	Priority, budget and elimination of obstacles
Owner of the process	Operational results and value validation
AI leader or product	Comprehensive coordination of the use case
Finance	Business case, ROI, payback and validation of the benefit
Data	Quality, access, lineage and preparation of information
Architecture and IT	Integration, performance, operation, and continuity
Cybersecurity	Access, protection, monitoring and response
Legal and compliance	Privacy, contracts, regulation, and intellectual property
Users of the process	Validation, adoption and detection of exceptions
Managing change	Communication, training and follow-up of use

The committee does not need to meet to approve every technical detail. It must intervene at four moments:

• Selection of the case.
• Pilot approval.
• Authorization for deployment.
• Decision to escalate.

Scanda's role in a 90-day roadmap

MaIA by Scanda organizes the adoption of AI from value identification to measurement and expansion. Its approach starts with the analysis of critical processes, data validation, and case prioritization based on business impact before making a technological investment.

The proposal includes:

• AI Value Discovery.
• Diagnosis of maturity.
• Process automation.
• Specialized agents and assistants.
• Integration with business systems.
• Government and traceability.
• Measurement of results.
• Scaling based on evidence.

This approach seeks to replace scattered initiatives with a path to adoption that includes quick wins, metrics, accountable parties, and defined controls.

Predictive conclusion

Over the next 12 to 18 months, access to AI models will no longer be a relevant competitive advantage. The difference will lie in each organization’s ability to integrate AI into real workflows, govern its decisions, and demonstrate impact on EBITDA, revenue, productivity, or risk.

Recent evidence shows a contradiction: adoption has already reached high levels, but scaling up and the financial impact remain concentrated in a minority. Incidents, errors, and exposure associated with implementations without sufficient controls are also increasing.

Therefore, the most valuable result of a 90-day roadmap is not a chatbot or a demonstration. It is a repeatable organizational capability to decide:

• Where to use AI.
• What data does it need?.
• What autonomy can be granted to them?.
• How much value it produces.
• When a person should intervene.
• Which cases deserve a larger investment.

Companies that establish this system will be able to scale more quickly and with less exposure. Those that accumulate isolated pilots will increase their technological spending without building a defensible operational advantage.

Frequently asked questions

1. How much should a company invest in its first AI project? The investment must be defined after establishing the baseline, the potential benefit, and the total cost of operation. The first project should be small enough to limit exposure, but comprehensive enough to measure data, integration, adoption, security, and ROI. It is not recommended to purchase an enterprise platform before validating at least one use case with real information and users.

2. Can AI be implemented without having perfect data? Yes. The data must be adequate for the purpose and level of risk of the case, not perfect in absolute terms. An internal assistant may tolerate levels of uncertainty that would be unacceptable in financial, regulatory, or security decisions. The company must measure quality, coverage, timeliness, permissions, and traceability before piloting.

3. What should the company have achieved by the end of the 90 days? It must have a proven use case, a baseline, operational metrics, adoption and risk, preliminary financial results, documented controls, a reference architecture, and a decision on scaling. If there is only a technological demonstration, the roadmap is not complete.

Turn 90 days of experimentation into a business decision

Scanda helps you identify where there is real value to apply artificial intelligence, validate the preparation of your data, design a controlled pilot, and build the business case to scale it.

We don't start with the tool. We start with the process, the financial impact, and the evidence your committee needs to decide.

Schedule an AI Value Discovery session with MaIA by Scanda and define which use case can generate measurable results within the next 90 days.