See how to use AI in construction to support cost estimating without losing traceability, official sources, calculation memoranda, and technical responsibility.
Check it out!
AI in construction can support the preparation and review of construction cost estimates, but it does not replace official price sources, market research, calculation memoranda, professional judgment, or technical responsibility. The TCU’s Guide to Cost Engineering in Public Works, published in 2026, treats Artificial Intelligence as an instrumental support tool and identifies specific precautions to prevent production speed from being confused with estimate reliability.
For the base cost estimate, the safest application is to use AI to organize information, identify inconsistencies, suggest verification paths, and accelerate preparatory tasks. Values, composition codes, productivity rates, standards, quantities, prices, and assumptions must remain anchored in verifiable sources. A model-generated response is not primary market evidence and cannot replace SINAPI, SICRO, quotations, comparable contracts, or project technical documents.
The real gain comes from the combination: automation to expand analytical capacity and engineering to validate, justify, and assume responsibility for the result.
What the TCU allows for AI in construction estimating
The use of AI in the base estimate is a method-and-evidence control within the procurement planning phase. To place these precautions within the TCU’s full control framework, see What the TCU Reviews in Construction and Engineering Services Procurement.
The 2026 Cost Engineering Guide introduces Artificial Intelligence into the estimating process pragmatically. The technology can support activities such as organizing the Work Breakdown Structure, reviewing consistency between design and worksheet, identifying omissions, suggesting candidate compositions, and organizing price-research information.
The central word is support.
The detailed estimate still needs to comply with applicable legal and technical requirements: quantities linked to the design, justifiable unit costs, BDI, social burdens, calculation memoranda, and documents supporting the estimates. Law No. 14,133/2021 does not create a methodological exception because a particular activity was automated.
An AI may indicate that a service appears to be missing. The professional who knows the design confirms the omission. It may suggest a SINAPI code. The estimator must verify whether the code exists, is current, represents the service, and uses the correct location and base date.
Generative AI is not a price database
Language models generate responses from learned patterns and provided context. They should not be treated as transaction databases that are inherently up to date.
This distinction is decisive in Cost Engineering.
If someone asks a model “what is the price of 30 MPa ready-mix concrete in Paraná?”, the answer may be plausible and still be unsuitable because it does not state, in an auditable manner:
- collection date;
- exact location;
- commercial condition;
- volume considered;
- transportation;
- taxes;
- payment terms;
- price source;
- quotation validity;
- statistical representativeness.
A market estimate needs to be reproducible. A textual response does not replace this requirement.
The article on Public Works Estimating details the evidence chain that must remain documented even when AI tools are used.
The eight precautions when using AI in estimating
The TCU Guide organizes points of attention that can be converted into a control policy for a public body, engineering firm, or project team.
1. Traceability: the AI response is not a primary source
The first rule addresses the greatest weakness of indiscriminate use of the technology.
If an AI states that a certain productivity rate is 0.25 h/m², that statement is not the technical origin of the coefficient. The professional needs to identify where the number came from or replace it with a verifiable reference.
The same applies to:
- SINAPI and SICRO codes;
- price-adjustment indexes;
- BDI parameters;
- waste percentages;
- labor coefficients;
- material consumption;
- hourly equipment costs;
- standards requirements;
- dates and numbers of TCU decisions.
The correct chain is:
AI identifies or suggests → professional verifies → primary source supports → memorandum records.
When the tool works on controlled internal documents, traceability improves, but the need to cite the source document remains.
2. Validation by a qualified professional
Automation does not shift professional responsibility.
An engineering estimate involves decisions about execution method, productivity, design compatibility, quantities, mobilization, equipment, risks, and local conditions. These decisions require technical education and judgment.
AI may reduce repetitive effort, but the final product remains a human technical statement. Where applicable, the estimating worksheet must remain linked to the corresponding technical responsibility.
The professional must be able to explain why a particular result was accepted. “The AI calculated it” is not a calculation memorandum.
3. AI does not replace SINAPI, SICRO, or other official references
Law No. 14,133/2021 establishes an order of parameters for the estimated value of construction and engineering services. For transportation infrastructure, SICRO occupies a central position; for other construction, SINAPI is the primary reference in cases covered by the federal rule.
An AI tool can help search, classify, or compare compositions. It should not create a fictitious third table between the design and the official system.
When the service is not adequately represented in the public reference, the solution remains to develop a custom composition, adapt a composition with justification, or use another permissible parameter, documenting the reason.
The analysis of the five effects that may distort SINAPI and SICRO prices shows that even official references should not be applied mechanically; with AI, the need for justification increases rather than decreases.
4. Hallucination: plausible codes, standards, and productivity rates may be false
Generative models may produce syntactically convincing information with no factual correspondence.
In Cost Engineering, this may generate errors such as:
- nonexistent SINAPI code;
- composition from a different state or month;
- reference to a superseded standard;
- statutory provision with incorrect content;
- productivity incompatible with the equipment;
- physically impossible consumption;
- wrong unit of measure;
- price without a base date;
- duplication between inputs and services;
- a formula that looks correct but is logically wrong.
The risk is high because the result often looks organized. A visually consistent worksheet may hide dozens of fabricated assumptions.
Validation must be systematic, not intuitive.
5. AI use needs to be documented in the administrative record when it influences the estimate
If the tool was used only to correct spelling in a text, its effect on the pricing decision may be irrelevant. If it was used to identify compositions, structure a WBS, classify quotations, or suggest coefficients, its use becomes part of the methodology.
Good practice is to record:
- purpose of use;
- tool or environment used;
- date;
- set of documents provided;
- type of output produced;
- checks performed;
- human decisions made from the output;
- sources that confirmed accepted data.
The objective is not to attach thousands of lines of conversation with no practical value. It is to allow an auditor, reviewer, or successor to understand how automation influenced the estimate.
6. Data, secrecy, and confidentiality need to be controlled
Projects may contain sensitive information: drawings of critical facilities, still-confidential values, procurement strategies, personal data, supplier documentation, and information protected by confidentiality.
Entering this content into an AI service without governance may constitute improper transfer of information.
Before use, the organization needs to know:
- where data are processed;
- whether they are retained;
- whether they may be used for training;
- who has access;
- which authentication controls exist;
- whether an isolated corporate environment exists;
- which document categories are prohibited.
Brazil’s General Data Protection Law adds obligations where personal data are present, but governance needs to go beyond LGPD. Trade secrets and critical-infrastructure information also require protection.
7. Version control is mandatory
Tools, models, and databases evolve rapidly. The same request may produce a different output after the model or supporting documentation is updated.
For this reason, relevant results need to be frozen in the process record.
Recording only “ChatGPT was used” or “AI was used” is insufficient. What matters is enabling reconstruction of the decision stage.
The estimate also needs its own version:
- design revision used;
- base date of references;
- quantity file;
- version of compositions;
- current assumptions;
- date of automated analysis;
- subsequent human changes.
This discipline reduces the risk of validating an AI output against a different design from the one used to produce it.
8. Custom compositions still require real productivity and a technical memorandum
AI can help break a service down into labor, materials, and equipment. This is useful for starting a composition that does not exist in official databases.
But the final coefficient needs a technical basis.
A custom composition should answer:
- which execution method was considered;
- which crew performs the service;
- which hourly production is expected;
- which losses were adopted;
- which equipment is required;
- which field condition limits productivity;
- which historical data exist;
- which tests, prototypes, or comparable projects support the value.
AI can structure the question. It cannot invent the productivity rate that makes the desired price work.
The difference between assistive use and automated decision-making
Not every use of AI carries the same risk.
| Use | AI role | Risk | Recommended control |
| Organize service list | Assistive | Low | Estimator review |
| Compare terminology | Assistive | Low to medium | Verify source and unit |
| Detect possible omission | Assistive | Medium | Check design and scope |
| Suggest candidate composition | Analytical | Medium | Confirm code and applicability |
| Extract quantities from text | Analytical | Medium to high | Validate against drawing/model |
| Propose productivity | Decision-making | High | Require history or method engineering |
| Define unit price | Decision-making | High | Official source or verifiable research |
| Approve estimate | Decision-making | Unacceptable without human responsibility | Formal technical approval |
The use policy should be proportional to the risk of the decision.
Where AI creates the most value without making the decision
The most mature application is not asking a model to “make the estimate.” It is using the model to expand review capacity.
WBS structuring
AI can turn the technical memorandum, scope, and design documents into an initial Work Breakdown Structure proposal. The result helps verify whether relevant disciplines and work packages were considered.
This WBS needs to be reconciled with the design, delivery model, and estimating worksheet.
Consistency review between design and worksheet
A model can compare lists extracted from memoranda, specifications, and the estimate and flag terms present in one document but absent from another.
Example: the technical memorandum mentions waterproofing a reservoir, but there is no corresponding service in the worksheet. This does not prove an omission; it creates a verification point.
Identification of candidate compositions
AI can help map a design description to keywords in a composition database. The output should be treated as a shortlist, not as the selected composition.
Quotation analysis
It can organize received proposals, normalize descriptions, highlight scope differences, and identify missing fields.
The final statistical analysis should use data actually received and a stated methodology.
ABC Curve review
After the estimate is formed, AI can support analysis of the cost ABC Curve to identify which items deserve in-depth validation.
The value is not in recalculating the cumulative percentage, which is a simple deterministic task. It is in cross-referencing the most relevant items with specification risks, market availability, economies of scale, and price sensitivity.
What should remain in deterministic tools
Generative models are not the best tool for every task.
Repeatable and critical calculations should remain in a spreadsheet, code, estimating system, or validated deterministic mechanism.
Examples:
- sums;
- application of BDI formulas;
- social burdens;
- monetary updating;
- ABC Curve;
- unit conversion;
- percentage distribution;
- numerical scenario analysis;
- quantity consolidation;
- automated comparison of revisions.
AI can explain the result, help find an inconsistency, or generate a review script. The final calculation should be reproducible independently of the natural language used in the conversation.
How to control hallucination in SINAPI and SICRO codes
A simple procedure reduces much of the risk.
- Ask AI only for candidate compositions; never state that the output will be adopted.
- Search each code in the corresponding official database.
- Check description, unit, base date, and location.
- Compare inputs and coefficients.
- Verify whether the design service uses the same construction technology.
- Record the reason for selecting or discarding it.
- If no composition is suitable, develop a technically justified solution.
This flow creates traceability between suggestion and decision.
AI ceases to be the “source” and becomes a search and triage mechanism.
Example: a nonexistent composition suggested convincingly
Imagine a model returns:
> “SINAPI 104532 — Installation of 200 mm perforated metal cable tray, including supports, labor, and accessories.”
The description looks perfectly plausible. The number may not exist, may belong to another service, or may come from a different database.
If the estimator enters the code directly into the worksheet, a false reference with an official appearance is created.
The appropriate control is to search the source for the code. If it does not exist, the generated text serves only as a research lead. The composition needs to be replaced by a real reference or built as a custom composition.
This is a small difference in workflow and a major difference in auditability.
AI can detect omissions, but it can also create unnecessary services
When given a technical memorandum, a model may suggest items “normally present” in similar projects. This helps uncover gaps, but may also introduce scope that the design does not require.
For this reason, every suggestion needs to be classified:
- explicitly specified in the design;
- technically necessary to execute a specified item;
- applicable standards requirement;
- risk provision;
- optional improvement;
- item unsupported by the scope.
Only technically justified categories enter the base estimate.
Mature designs make AI more useful
Automation quality is limited by the quality of the documents provided.
If the design has inconsistent drawings, generic memoranda, quantities without a supporting memorandum, and contradictory specifications, AI can only reorganize the uncertainty.
This principle connects the topic to Design Review. Before automating the estimate, it is worth verifying whether the design contains sufficient information to be estimated.
A model does not turn an incomplete preliminary design into a detailed design. At most, it can help make what is missing explicit.
AI and parametric estimating: a powerful and dangerous combination
Parametric estimating uses relationships between technical variables and historical costs. AI and statistical techniques can help find correlations, classify comparable projects, and test hypotheses.
The risk arises when correlation is confused with causation or when the model uses non-comparable projects.
A parametric estimate needs to state:
- data universe;
- explanatory variables;
- outlier treatment;
- base date;
- location;
- project type;
- error measure;
- application range;
- out-of-sample validation, where applicable.
AI can support modeling, but it does not eliminate the need for statistics and Cost Engineering.
AI applied to market research needs to preserve the original evidence
A tool can read multiple quotations and create a consolidated table. This saves time, especially when suppliers use different formats.
However, the consolidated worksheet does not replace the documents received.
The process should retain:
- original proposal;
- supplier identification;
- date;
- validity period;
- scope;
- commercial conditions;
- freight;
- taxes;
- quantities;
- unit of measure;
- person responsible for collection.
If AI incorrectly normalizes “unit,” “kit,” “set,” and “meter,” the comparison may be mathematically correct and economically invalid.
Data protection and sensitive information need to be part of the policy
Law No. 13,709/2018 regulates the processing of personal data, but engineering estimates may involve confidentiality risks even without personal data.
Critical-infrastructure projects may contain:
- locations of security systems;
- telecommunications routes;
- electrical diagrams;
- vulnerability information;
- confidential prices;
- names of strategic suppliers;
- contracts and commercial conditions.
An internal policy may classify documents by level, defining which may be processed in an external service and which require a controlled corporate environment.
RAG and controlled databases improve quality, but do not create automatic truth
A retrieval-augmented generation architecture allows AI to consult selected documents before responding. In engineering, this is much more useful than relying only on the model’s general knowledge.
A controlled knowledge base may contain:
- SINAPI and SICRO manuals;
- exported compositions with base date;
- technical booklets;
- organization specifications;
- project history;
- internal productivity data;
- selected case law;
- internal standards;
- project designs and technical memoranda.
Even so, the output needs to indicate the origin of the data. If the database contains an outdated or incorrect document, the system will merely retrieve the error more efficiently.
Document governance and AI governance need to advance together.
The usage log does not need to become a complete transcript of the conversation
Documenting AI in the administrative record does not mean attaching every interaction without curation.
A compact technical memorandum may record:
| Field | Example record |
| Purpose | Review consistency between technical memorandum and worksheet |
| Documents | PB-R02, MD-R03, estimate R04 |
| Tool | Corporate AI environment |
| Date | 17/09/2026 |
| Result | 14 candidate discrepancies identified |
| Validation | 9 discarded, 3 corrected, 2 justified |
| Final sources | Design, SINAPI, quotations, and quantity memorandum |
| Responsible party | Professional who reviewed and approved |
This structure is more auditable than saving an extensive history without identifying which outputs influenced the decision.
How to review an estimate produced with AI support
Independent review should start from the premise that the technology may have made a coherent-looking error.
An efficient sequence is:
- freeze the estimate revision;
- validate totals and formulas through a deterministic mechanism;
- review the ABC Curve of services;
- review the ABC Curve of inputs;
- verify codes in official databases;
- test items without an official code;
- compare critical quantities with the memorandum and design;
- test units of measure;
- verify relevant market prices;
- audit BDI and burdens;
- review standards references;
- document discrepancies and corrections.
Priority should be economic. It makes no sense to spend the same effort validating an item representing 0.01% and one representing 15% of the estimate.
The ABC Curve is the best entry point for AI validation
The ABC Curve makes it possible to rank financial exposure.
In AI-assisted estimating, Class A items deserve enhanced control because any hallucination, duplication, or productivity error has a material impact.
For each relevant item, the review may verify:
- origin of the quantity;
- applicability of the composition;
- price of the dominant input;
- productivity;
- economies of scale;
- local condition;
- duplication with another service;
- presence of risk already covered in another component.
This process improves quality even when no AI has been used. With AI, it becomes an additional safety barrier.
How to use AI to compare design revisions and estimate revisions
A high-value application is change analysis.
When the design moves from R02 to R03, the system can help identify:
- new elements;
- removed elements;
- material changes;
- capacity changes;
- specification revisions;
- dimensional changes;
- potential impact on existing services.
The output should generate a list of candidate impacts. The estimator verifies each one and updates the revision memorandum.
This improves change governance and reduces the risk that the worksheet remains linked to an earlier design revision.
An internal AI policy for Cost Engineering
The TCU Guide creates room to convert these precautions into a corporate procedure.
A policy may contain five layers.
Permitted uses
- document organization;
- item classification;
- text review;
- identification of inconsistencies;
- checklist generation;
- research support;
- comparison between revisions;
- preliminary WBS structuring.
Conditional uses
- suggesting compositions;
- extracting quantities;
- analyzing quotations;
- parametric estimating;
- productivity modeling;
- risk analysis.
In these cases, the policy requires specific validation and an external source.
Uses prohibited without formal validation
- creating a final price without a source;
- inventing a composition code;
- approving an estimate;
- defining BDI on its own;
- replacing the responsible professional;
- entering a confidential document into an unauthorized environment.
Recordkeeping
Define when use must be documented and what level of detail is necessary.
Audit
Provide for periodic review of the policy, authorized tools, and samples of use.
How to require AI governance in the Terms of Reference
If the Public Administration hires a company to prepare an estimate, design, or cost analysis, it may establish traceability requirements for relevant automation without mandating a specific technology.
The Terms of Reference may require that:
- all cost sources be identified;
- custom compositions have a productivity memorandum;
- automated results be validated by a qualified professional;
- AI tools not be treated as primary sources;
- confidential documents be handled in an authorized environment;
- material changes generated by automation have a review record;
- the contractor deliver editable files and calculation memoranda;
- technical responsibility remain clearly identified.
This approach regulates the result and evidence, not the software brand used.
A public body does not need to prohibit AI to preserve control
A blanket prohibition may be impractical and difficult to enforce. Productivity tools already incorporate AI capabilities in editors, spreadsheets, search systems, and design platforms.
It is more effective to establish control requirements:
- no material decision without a verifiable source;
- no confidential information in an unauthorized environment;
- no custom composition without a supporting memorandum;
- no approval without a responsible human professional;
- all relevant automation subject to review.
The focus shifts from “was AI used?” to “is the product reproducible, substantiated, and technically accountable?”
The responsible professional needs to understand the output before signing
Signing a worksheet produced by an automated tool without understanding its logic transfers the risk to the professional, not to the software.
Technical responsibility requires the ability to defend:
- quantities;
- estimating method;
- compositions;
- productivity;
- prices;
- BDI;
- burdens;
- risk assumptions;
- base date.
The article on ART, RRT, and TRT in Public Works details the role of technical responsibility in procurement documents.
Cost Engineering with AI requires more documentation, not less
It is tempting to imagine that automation reduces the need for calculation memoranda. The correct effect is the opposite.
The more stages are automated, the more important it becomes to document:
- inputs;
- rules;
- versions;
- exceptions;
- validations;
- final source of each material decision.
Cost Engineering for Construction and Engineering Services can incorporate automation without losing the traceability required in public estimating.
Technology is useful when it reduces mechanical effort and frees the professional to investigate relevant items, challenge assumptions, and review the design.
Checklist of eight controls before accepting an AI output
Before incorporating AI-generated information into the estimate, confirm:
- there is a verifiable primary source for the material data;
- a qualified professional reviewed the output;
- no official reference was replaced by an invented value;
- codes, standards, units, and dates were confirmed;
- material use was recorded in the process memorandum;
- no restricted data were sent to an unauthorized environment;
- the design, estimate, database, and tool have identifiable versions;
- custom compositions have supported productivity and execution methods.
If any answer is negative, the output is not yet ready to become part of the base estimate.
Final considerations
AI in construction can substantially increase the capacity to analyze designs and estimates, but the gain depends on a control architecture. The technology is especially useful for organizing large volumes of information, comparing documents, identifying discrepancies, and directing human review toward the most material items.
In the base estimate, the red line is clear: AI does not create economic evidence by itself. Price needs a source; quantity needs a design or supporting memorandum; productivity needs a method or history; a decision needs a responsible professional.
The eight precautions presented in the TCU’s 2026 Cost Engineering Guide provide an appropriate governance structure: traceability, professional validation, preservation of official references, hallucination control, documentation of use, data protection, versioning, and technical support for custom compositions.
An organization applying these controls can use AI to gain speed without trading auditability for convenience. This is the balance point: automate repetitive work and preserve with the professional what cannot be delegated — judgment, decision-making, and technical responsibility.
Technical references
[1] TRIBUNAL DE CONTAS DA UNIÃO. Engenharia de Custos em Obras Públicas: um guia de perguntas e respostas. Item 5.1.37 and other estimating guidance. Brasília: TCU, 2026. Available at: [TCU — Infrastructure and publications](https://portal.tcu.gov.br/infraestrutura).
[2] TRIBUNAL DE CONTAS DA UNIÃO. Licitações e Contratos: Orientações e Jurisprudência do TCU. Item 4.4.3.6 — Detailed estimate of the total cost of the works. Available at: [TCU — Detailed estimate](https://licitacoesecontratos.tcu.gov.br/4-4-3-6-orcamento-detalhado-do-custo-global-da-obra/).
[3] BRASIL. Lei nº 14.133, de 1º de abril de 2021. Lei de Licitações e Contratos Administrativos, especially Articles 18, 23, and 59. Available at: [Planalto — Lei nº 14.133/2021](https://www.planalto.gov.br/ccivil_03/_ato2019-2022/2021/lei/l14133.htm).
[4] BRASIL. Lei nº 13.709, de 14 de agosto de 2018. Lei Geral de Proteção de Dados Pessoais — LGPD. Available at: [Planalto — Lei nº 13.709/2018](https://www.planalto.gov.br/ccivil_03/_ato2015-2018/2018/lei/l13709.htm).
Frequently asked questions
It can be used as a support tool, but the estimate still needs verifiable sources, calculation memoranda, official references, and validation by a responsible professional. AI should not be treated as a primary price source.
No. SINAPI and SICRO remain official references in cases provided for by Law 14,133 and applicable regulations. AI may help search or compare compositions, but each code and data point needs to be verified in the official database.
It is the generation of plausible but false or unsupported information, such as nonexistent composition codes, incorrect standards, unrealistic productivity rates, switched units, or prices without a source.
When AI materially influences the estimate, it is advisable to record the purpose, documents used, output type, checks, and human decisions. The objective is to preserve traceability, not necessarily to attach the complete conversation.
It can support decomposition and structuring, but labor, material, equipment, and productivity coefficients need to be supported by the execution method, historical data, tests, references, or other verifiable technical evidence.
Responsibility remains with the officials and professionals who prepare, review, and approve the document according to their duties. The tool does not replace technical responsibility or professional judgment.
Document organization, comparison between revisions, identification of inconsistencies, shortlisting candidate compositions, classification of quotations, and directing human review are high-value uses when followed by validation.
The organization should classify information, use authorized environments, understand data retention and access, and prevent confidential or sensitive content from being sent to tools without adequate controls.
Complementary technical materials
Related solutions
Related services
- Cost Engineering for Construction and Engineering Services: quantities, compositions, BDI, and reference basis
- Design Review in Engineering Projects: technical review, interfaces, and project maturity
Main content on the topic
- Public Works Estimating: how to structure quantities, costs, BDI, and risks
- SINAPI and SICRO Prices: 5 effects that may distort market value
