Proof of concept, pilot, or productive implementation of AI: how much to invest in each stage

A proof of concept should fund evidence; a pilot, operational validation; and production, reliability, safety, and scale. The mistake is to fund all three as if they were the same project. The budget should grow only when data, performance, adoption, risk, and unit economics show that there is a business case capable of producing a ROI.

Taxonomic definitions

1. Proof of concept of artificial intelligence — PoC

A AI concept test It is a limited-scope experiment designed to demonstrate whether a use case is technically and economically viable.

It does not intend to operate on a scale or serve the entire organization. Its role is to answer questions such as:

• Can AI solve the problem with the available data?
• Is the precision obtained sufficient for the process?
• Is there a measurable improvement compared to the current baseline?
• What technical dependencies would appear when scaling up?
• Does the potential benefit justify continuing to invest?

A PoC that only produces an attractive demonstration, but not metrics, costs, and exit criteria, has still not fulfilled its purpose.

2. Pilot of artificial intelligence

A AI pilot bring the use case to a more realistic environment: real users, real data, selected integrations, and sufficient security controls to measure behavior, adoption, and results.

The difference is important: a PoC demonstrates that an idea is feasible. It can work; a pilot demonstrates whether It works within the business.

3. Productive implementation of AI

A Productive implementation It is an artificial intelligence system integrated into the operation and subject to business requirements regarding availability, identity, security, auditing, monitoring, support, continuity, performance, costs, and governance.

NIST recommends that AI systems be evaluated before deployment and regularly monitored during operation, including performance, security, privacy, reliability, and emerging risks.

PoC vs pilot vs production: investment, goal and exit criteria

Variable	Proof of concept	Pilot	Productive implementation
Question that must be answered	Can it work?	Does it work with real users, data, and processes?	Can it operate safely, profitably, and sustainably?
Range	A limited use case	A controlled population or process	Business operation
Data	Representative sample	Real operational data	Governed and recurring pipelines
Integrations	Minimals	Selected critical systems	End-to-end integration
Users	Reduced team	Controlled group	Users expected to be operational
Security	Basic controls	Identity, access and evidence	Security, IAM, auditing, incident response
Observability	Experimental metrics	Quality, latency, cost and behavior	Continuous monitoring, alerts, traceability, and SLA
Financial metric	Economic potential	Observed value	ROI, TCO, margin and operating cost
Final decision	Discard, redesign or pilot	Stop, adjust or produce	Operate, optimize and scale
Public reference for investment	From very limited prototypes of US$10,000 to publicly released enterprise PoCs of US$75,000–US$150,000	Benchmark published of US$150,000–US$450,000 for enterprise pilot	Benchmark published of US$450,000–US$2 million for enterprise rollout

Important: these figures They are not Scanda rates nor a neutral average of the market. They serve to dimension why the term “PoC” alone does not allow a project to be quoted.

Microsoft Marketplace, for example, publishes a four-to-six-week fixed-scope offer for a Copilot prototype for $10,000. In the AWS Marketplace, there are PoC programs where certain projects can receive up to $25,000 funding from the hyperscaler. At the enterprise end, Alice Labs publishes bands of its own for 2026 ranging from $75,000–$150,000 for PoC, $150,000–$450,000 for pilot, and $450,000–$2 million for rollout; this should be interpreted as a reference published by a vendor, not as a universal fee.

The conclusion for Finance is more useful than any single figure: the price depends less on the stage label and more on data, integrations, autonomy, criticality, regulation, competition, and the level of operation required.

Proof of concept, pilot and production: what really changes

The AI budget should not be approved as a one-time check from the beginning.

Each stage must make a different decision.

That approach is especially relevant because the current bottleneck is no longer getting access to an artificial intelligence model. It is about turning the model into a capability that works within business processes, systems, and controls.

S&P Global Market Intelligence reported that the proportion of organizations that abandon most of their AI initiatives before production increased from 17% to 42%, and that an average organization eliminates approximately 46% of its proof-of-concept projects before bringing them into production.

This changes the financial logic.

The question should not be: “How much will the entire project cost?”

First it must be: “How much should we invest to obtain sufficient evidence to justify the next investment?”

¿How much does each stage of an AI project cost??

Stage 1. Proof of concept: Invest in evidence

A PoC should focus budget on resolving uncertainties.

It must include, at a minimum:

• definition of the business problem;
• baseline of the current process;
• inventory and data analysis;
• selection of architecture and models;
• information preparation;
• prototype construction;
• evaluation dataset;
• quality tests;
• latency measurement;
• consumption measurement;
• initial identification of risks;
• Preliminary calculation of the economic value;
• an estimate of the cost of reaching production.

It should not include oversized infrastructure, high business availability, or integrations that are not necessary to test the hypothesis.

The main deliverable is not the chatbot, agent, model, or dashboard.

It is a documented decision to go, redesign or stop.

NIST precisely incorporates the business value, benefits, costs, risk tolerance, and scope of application within the AI risk mapping stage, before making deployment decisions.

What should justify the investment of a PoC

Before approving the pilot, there must be quantifiable answers to:

• quality of the result;
• case coverage;
• error rate;
• response time;
• availability and quality of data;
• cost per execution;
• potential savings;
• potential revenues;
• reduction of rework;
• critical risks;
• required integrations;
• Probable cost of the next level.

A PoC that ends by simply saying “the technology works” leaves Finance practically where it started.

Stage 2. Pilot: investing in operational reality

Here are costs that many proof-of-concept tests avoid.

The pilot must gradually incorporate:

• real users;
• real processes;
• permits;
• identity profiles;
• integrations;
• interfaces;
• exceptions;
• human supervision;
• security;
• logging;
• observability;
• continuous evaluation;
• training;
• support;
• feedback mechanisms.

The question now is different:

Does the solution produce value when it interacts with the real complexity of the company?

This difference explains part of the so-called The gap between pilot and production.

McKinsey found in its 2025 global study that 88% of the respondents reported regular use of AI in at least one function, but only about a third claimed that their organization was scaling up their AI programs.

Implementing does not equate to scaling.

What a pilot should measure

A pilot must generate evidence on five levels.

1. Performance

• quality;
• precision;
• resolution rate;
• errors;
• exceptions;
• latency.

2. Economy

• cost per execution;
• cost per user;
• cost per transaction;
• infrastructure;
• consumption of models;
• residual human operation.

3. Adoption

• active users;
• frequency;
• abandonment rate;
• completed processes;
• human intervention required.

4. Risk

• material errors;
• information leakage;
• unauthorized access;
• decisions outside the authorized scope;
• traceability;
• possibility of reversal.

5. Business impact

• hours recovered;
• cost avoided;
• reduction of rework;
• reduction of cycle times;
• capacity increase;
• incremental margin;
• Expected losses avoided.

It’s not enough to ask if the users were “satisfied”.

The pilot must produce evidence that can be used to support a business case.

Stage 3. Productive implementation: investing in reliability and scale

Production doesn't just mean changing a URL or enabling more users.

The budget should include capabilities that can be almost invisible during PoC:

• identity management;
• separation of permits;
• management of secrets;
• integration with business systems;
• government of data;
• information protection;
• automatic evaluation;
• observability;
• Versioning;
• CI/CD;
• prompt management;
• model management;
• fallback;
• human intervention;
• error handling;
• continuity;
• response to incidents;
• auditing;
• support;
• training;
• change management;
• optimization of consumption;
• supplier monitoring.

NIST notes that the behavior and functionality of AI systems must be monitored when they are in production, in addition to maintaining incident management, recovery, change, and eventual deactivation mechanisms.

In IA agentic, this point becomes even more relevant. The global study Pulse of Agentic AI 2026 Of Dynatrace’s findings, 52% of the participants found that the main barriers to bringing agents into production were security, privacy, or compliance, and 51% had technical difficulties managing and monitoring agents on a scale.

That is to say: the costs that appear when scaling are not just tokens.

They are the costs of converting probabilistic capacity into a manageable operation.

What should the budget buy at each stage?

A PoC buys knowledge

It should reduce uncertainty about:

• viability;
• data;
• architecture;
• performance;
• economic value;
• risks.

A pilot purchases operational evidence

It should reduce uncertainty about:

• adoption;
• integrations;
• real behavior;
• costs;
• exceptions;
• supervision;
• Controls.

Production buys reliability

It should reduce uncertainty about:

• continuity;
• scalability;
• security;
• compliance;
• support;
• traceability;
• recurring costs;
• capacity for evolution.

This is a healthier way to budget for AI than starting with a global figure and then hoping the project will later find a justification.

How to build the budget for an AI initiative

Step 1. Define the economic baseline before the model

Before selecting a technology, the current cost of the problem must be known.

For example:

• annual volume of transactions;
• hours invested;
• loaded labor cost;
• errors;
• rework;
• waiting times;
• outsourcing;
• penalties;
• loss of productivity;
• uncaptured sales opportunities;
• Affected margin.

Without a baseline, there is no ROI; there is only a technological comparison.

Step 2. Calculate the available economic value

One way to structure it is:

Potential economic value = recovered capacity + avoided cost + incremental margin + reduction of expected loss

To avoid overestimating the business case, the theoretical savings must be adjusted by:

• actual level of adoption;
• percentage of the process that can actually be automated;
• the need for human supervision;
• residual error rate;
• time required to reach the altitude.

The CFO should analyze Margin and cost avoided, Not just “hours saved”.

An hour of freed up capacity does not automatically increase EBITDA if that capacity is not reused, reduces costs, or allows for increased production.

Step 3. Separate construction cost and operating cost

The budget must clearly distinguish:

Build

• discovery;
• data;
• development;
• integration;
• tests;
• security;
• deployment;
• training.

Run

• inference;
• storage;
• vector bases;
• APIs;
• licenses;
• observability;
• support;
• evaluation;
• maintenance;
• model updates;
• change management;
• human intervention.

This separation prevents a project from appearing profitable because only the construction cost was presented.

Step 4. Calculate TCO, not just the initial cost

The executive calculation should use: TCO = implementation investment + recurring costs + operation + support + controls + evolution

Afterwards: ROI = (adjusted economic benefit − TCO) / TCO

And, for risk reduction projects: Expected loss = probability of the event × financial impact

The expected reduction between the current scenario and the scenario with AI can be incorporated into the business case when there is sufficient evidence to support it.

Step 5. Establish financial gates

Each stage requires explicit criteria of continuity.

Gate PoC → Pilot

Advance only if:

• the problem still has economic value;
• the data proved to be usable;
• the performance exceeds the agreed baseline;
• there is no material risk blocking;
• The expected cost of scaling remains justifiable.

Gate Piloto → Production

Advance only if:

• there are results with real users;
• adoption is sufficient;
• the integrations are viable;
• the critical controls were validated;
• recurring costs are understood;
• there is operational ownership;
• There is an updated business case.

Gate Production → Scale

Expand only if:

• the result remains stable;
• the economic value materializes;
• unit costs remain under control;
• the system can be monitored;
• the risks remain within tolerance;
• The process can be replicated.

¿What a PoC must demonstrate before moving to pilot mode?

An organization should be able to answer clearly:

• Which indicator changed?
• What was the baseline?
• What was the result?
• With what data was it tested?
• What percentage of cases were excluded?
• How much does an execution cost?
• What are the main sources of error?
• What would I need to change to operate with real data?
• What risks remain open?
• What would be the approximate cost of the next step?

The need for discipline is significant.

S&P Global found that 46% of the organizations surveyed did not identify a single business goal with a “strongly positive” impact derived from their investments in generative AI, while only 19% reported that level of impact across most of their objectives.

The goal of a PoC, therefore, should not be to demonstrate enthusiasm.

It should avoid making larger investments in cases where capital is not yet justified.

What must a pilot demonstrate before it goes into production?

The pilot must try three things simultaneously.

1. That AI works

Not just in ideal cases.

Also with:

• incomplete data;
• exceptions;
• different users;
• variations in the process;
• ambiguous situations.

2. That the business uses it

A technically correct but underused solution will hardly generate the expected business case.

3. That the company can operate it

It should be clear:

• who is responsible for the system;
• who approves changes;
• who monitors results;
• who handles incidents;
• when a person intervenes;
• when the system stops.

McKinsey also found that their organizations classified as AI high performers They were 2.8 times more likely than the rest to have fundamentally redesigned their workflows when deploying AI.

The implication is important: the ROI is not usually seen by inserting AI into an intact process.

It often requires redesigning the process around the new capacity.

How to evaluate ROI, TCO and impact on EBITDA

A CIO can demonstrate that the model works.

The CFO needs to demonstrate that the investment creates value.

That is why it is advisable to connect each technological KPI with its economic equivalent.

Technical metric	Operational translation	Financial translation
Greater precision	Fewer errors	Less rework and cost
Lower latency	Lower cycle time	Greater capacity
Greater automation	Less manual intervention	Lower cost per transaction
Greater availability	Less interruptions	Lower operational loss
Less false positives	Less unnecessary revisions	Lower labor cost
Better autonomous resolution	Less escalations	Lower cost of care
Higher adoption rate	Greater utilization	Higher capture of the expected value
Better traceability	Greater control	Lower operational and regulatory exposure

The MIT CISR found in its study of AI business maturity that organizations in the pilot-focused stage were 6.8 percentage points below the average growth of their industry and 1.4 points below in profit; organizations ranked in their most mature stage were 13.9 points above in growth and 9.9 points above in profit. The study precisely identifies the transition from pilots to scaled AI-enabled work forms as a relevant turning point.

This does not mean that AI scaling guarantees those results.

It reinforces an idea: value creation occurs when AI stops being an experiment and becomes an operational capability.

¿When to stop an artificial intelligence project?

Canceling an initiative can also be a good investment decision.

It is advisable to stop or redesign when:

• the problem does not have sufficient economic value;
• the necessary data are not available;
• The improvement against the baseline is marginal;
• the operating cost eliminates the benefit;
• Too much human intervention is required;
• the residual risk exceeds the tolerance;
• integration requires disproportionate investment;
• adoptions are low;
• there is a more economical conventional alternative;
• the organization cannot yet assume responsibility for the operation.

The PoC abandonment rate reported by S&P Global should not be interpreted solely as a technological failure. It also demonstrates the need to use the early stages to discard early and cheaply anything that should not receive production budget.

The financial problem arises when a company invests as if it were building production before checking whether it actually has a business case.

The Scanda approach: scaling evidence, not experiments

The logic of Scanda is based on the premise that automation and applied AI must advance in an organized manner.

The current Scanda offering proposes identifying value first, reviewing the feasibility of the data, prioritizing use cases, defining an adoption path, implementing, and then measuring and expanding based on the results obtained.

Applied to investing, this means:

• do not start with the tool;
• Do not confuse a prototype with a product;
• do not climb without metrics;
• not approving production without TCO;
• not automating a process that should first be redesigned;
• not turning every successful experiment into an enterprise implementation;
• Do not measure success by the number of agents or models deployed.

The asset that is being built is not just an agent.

It is the organizational ability to identify, validate, implement, and govern AI repeatedly.

Predictive conclusion

During the next budget cycles, the business conversation about AI will tend to shift from “how many projects do we have” to “how many produce sustainable financial results.” Current data already show the tension: adoption continues to grow, but the transition from experimentation at scale remains limited; at the same time, security, monitoring, integration, and governance appear as relevant barriers to production. Therefore, it is predictable that CFOs, CIOs, and COO will raise the bar for approval.

A PoC must demonstrate evidence. A pilot must demonstrate operational economy. Production must demonstrate control. And the investment will tend to focus on fewer use cases, but with a clearer path towards impact on productivity, margin, costs, and risk. The advantage will not be in spending less on the PoC. It will be in knowing exactly what evidence should be purchased for each dollar before authorizing the next.

FAQ

1. How much does an artificial intelligence concept test cost? There is no universal pricing because the scope can vary drastically. As a verifiable reference, Microsoft Marketplace publishes limited prototypes starting at $10,000, while an enterprise implementation provider that publishes its bands for 2026 places a PoC between $75,000 and $150,000.

The budget should be defined by the number of data sources, integrations, criticality, security, models, evaluation, and functional scope, not just by the word “PoC”.

2. When is an AI pilot ready to go into production? When it has demonstrated results with real users and data, the critical integrations work, the risks are within tolerance, the recurring cost is understandable, and there is an operational responsible person. There must also be sufficient evidence to update the ROI and TCO before committing to the production investment.

3. What costs arise when an AI solution is brought to production? In addition to the model, data governance, integrations, identity, security, observability, evaluation, storage, support, continuity, audit, training, change management, and operation are also included. NIST recommends that systems be evaluated before deployment and continuously monitored during operation.

Do not finance production when you still need a proof of concept. The most dangerous cost of an artificial intelligence project is not always investing too much. It can also be scaling too early into a case that has not yet demonstrated value. With MaIA by Scanda We help you structure the adoption of artificial intelligence in an organized way: identify the use case, validate the data, define the baseline, establish KPIs, build the value proof, and determine what the project needs to demonstrate before receiving the next investment.

If you have an AI case under evaluation, let’s talk to determine if you need a PoC, a pilot, or a production implementation today, and build the business case before committing more budget.