Skip to content
be-human.ai
← Back to the journal

AI by industry

AI for Small Engineering Firms: Project Files and Submittals

A practical plan for small engineering firms to track project files and submittal status with an AI agent while engineers keep stamps, specs and sign-off.

Be Human typographic cover with the title: AI for Small Engineering Firms: Project Files and Submittals
Conceptual illustration created with AI · Be Human

An AI agent can help a small engineering firm track project files and submittal status by keeping one current register of drawings, revisions, specifications and submittal packages, then reporting what is late, missing or waiting on a reviewer. It should not stamp drawings, make design decisions, or approve a submittal. The engineer of record keeps those decisions, and the agent makes sure they are made on time with the right documents in front of the right person.

Key takeaways

  • Start with a register of project files and submittal status, not with drafting or design work.
  • Keep stamps, specification interpretation, deviations and final acceptance with a named engineer.
  • Give the agent read access first, and write access only for logging, reminders and drafts for review.
  • Define one observable success criterion, such as every open submittal having a named reviewer and a next review step.
  • Review the agent's permissions and project memory on a fixed schedule, because project files change every week.

What should an agent track in project files and submittals?

Small firms usually keep project documents in several places at once: a shared drive, a project folder that one engineer maintains by hand, email threads with contractors, and a few local copies on laptops. The first problem is rarely a lack of documents. It is that nobody can answer, quickly and with confidence, which revision is current, which submittal is waiting on whom, and what was promised in the last coordination call.

A useful agent starts with two records. The first is a file register: drawing number, title, discipline, current revision, issue reason, issuing engineer, and the location of the controlled copy. The second is a submittal log: the item, the specification section it answers, the contractor or vendor who sent it, the date it was received, the assigned reviewer, the current status, and the next required action. The agent can keep both records current by reading the document store, email notices and the firm's project tool, then flagging mismatches such as a drawing that was superseded but still referenced in a transmittal.

The same logic applies to requests for information. If your firm also manages RFIs, keep them in the same register so that a late answer to a question can be traced to the submittal it blocks. For a deeper look at this pairing, see Superintelligent Agents for Civil Engineers: RFI and Submittal Logs.

Which decisions must stay with engineers?

An agent can organize a submittal review, but it cannot carry professional responsibility. The line should be written down before any tool is connected, so that everyone in the firm knows what the agent may touch and what it may not. A practical boundary for most small firms includes these decisions:

  • Applying a professional seal or stamp to any drawing, calculation or report.
  • Judging whether a product, material or method meets the specification or the applicable code.
  • Accepting, rejecting or accepting with comments a submittal, including any substitution or deviation.
  • Interpreting an ambiguous specification clause or resolving a conflict between drawings.
  • Changing scope, schedule commitments or fees, or releasing documents to a client or authority.

The agent can prepare material for each of these decisions. It can pull the relevant specification section, list the cut sheets that were received and the ones that are missing, and draft a comment summary for the engineer to edit. The engineer reads the draft, changes what is wrong, and signs the response. This is the kind of human oversight that the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) treats as part of governing an AI system: someone accountable, a defined scope, and a record of what was decided and why.

How do you set up the first phase, step by step?

Most firms should begin with one active project or one submittal type, not with the whole portfolio. The goal of the first phase is to prove that the register is accurate and that reminders reach the right person. A workable sequence looks like this:

  • Inventory where project files live today, and name one owner for each active project.
  • Build the submittal register with the fields listed earlier, and fill it from the last three months of correspondence.
  • Connect read access to the document store and to the register. Leave write access off.
  • Define the statuses the agent may set, such as received, logged, under review, returned and closed. Only an engineer may move an item to accepted or rejected.
  • Set reminders for items with no assigned reviewer, items that have passed the review period in the contract, and revisions that have not been acknowledged by the contractor.
  • Send one weekly summary to the project lead, listing open items by owner, and keep a log of every action the agent takes.
  • Review every error in the first few weeks, correct the rules, and only then consider adding another project or tool.

[Image placeholder]

A small firm can also use the Be Human approach to project management offices as a model for reporting. The same discipline applies: the report must say where information came from, and a person must be able to see what changed since the last update.

How should permissions and project memory be set?

Permissions are where small firms most often take a shortcut that later causes trouble. An agent connected to a full shared drive with edit rights can move, rename or overwrite files by mistake, and it becomes hard to tell which person or process caused a change. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) recommends granting only the access a task needs and checking that access on each request. For this use case, that means read access to the project folders, write access to the register and the reminder queue, and no access to signing tools, billing or client release folders.

Vendor documents need particular care. A submittal package may contain text that tries to instruct a reader or an agent, whether by accident or on purpose. The [OWASP guidance on prompt injection](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes the principle to apply: treat document content as data to be summarized and checked, never as instructions that change the agent's behavior.

Project memory should hold the register, the naming conventions, the review periods from each contract, and the names of the people responsible for each discipline. It should not hold unrelated client files or personal data beyond what the project needs. A deeper discussion of how a firm can keep one shared memory across projects appears in Superintelligent Agents for Engineering Firms: Project Memory.

Illustrative scenario: a packaged unit submittal waiting on a manufacturer

Illustrative scenario: a contractor sends a submittal for a packaged mechanical unit for a building under construction. The specification requires a performance table and a set of cut sheets. The agent logs the package when it arrives, notes that the cut sheets are missing and that no reviewer has been assigned, and sends a reminder to the project lead with both points listed. It also drafts a request to the contractor asking for the missing sheets, and the project lead approves that message before it goes out.

When the complete package arrives, the agent places the performance table next to the specification clauses it answers and lists two values that fall outside the stated range. The mechanical engineer reviews the comparison, decides that one value is acceptable under the stated tolerance, and writes a short response that accepts the unit with one condition. The engineer signs the response in the firm's normal process. The agent then updates the status to returned, attaches the signed comments to the register, and schedules a check to confirm that the contractor acknowledged the condition. At no point did the agent judge the equipment or approve the submittal.

How Be Human's managed approach handles this

Be Human designs this kind of setup around one useful role, defined with the memory and tools it needs, explicit approval boundaries, and an observable success criterion. For a firm like this, the first engagement usually covers the register, the reminder rules, the permission model and the weekly summary, with the engineers deciding which decisions remain manual. Managed operations can then include monitoring of register accuracy, recurring evaluations against a set of known test cases, permission reviews as the team changes, and incident response when a file is misfiled or a status is set incorrectly. The approach follows a principle that [Anthropic's guidance on building effective agents](https://www.anthropic.com/engineering/building-effective-agents) also stresses: begin with simple, well-bounded workflows with human checkpoints, and add complexity only where measured results justify it.

The free Be Human AI diagnostic can help you find the first task worth handing to an agent in your firm.

Frequently asked questions

Can an AI agent approve a submittal? No, and it should not be set up to try. The agent can log the submittal, check it against the specification section, draft comments and track the response. The engineer decides whether the item is accepted, rejected or accepted with comments, and signs the response.

What is the best first use of an agent in a small engineering firm? The best first use is usually a submittal and file register that is accurate and reminds the right person about open items. It is narrow, easy to check, and it reduces the chance that a late review holds up a construction schedule. Expanding into drafting or design support should wait until the register has proven reliable.

Which systems does the agent need to connect to? At minimum, it needs read access to the project document store and write access to the register and reminder queue. It should not have access to signing tools, billing, or folders used for client release. Each connection should be granted only for the project and task it serves.

How do we know the agent is working correctly? Define success in observable terms, such as every open submittal having a named reviewer and a next step, and check the register against the source documents each week. Keep a log of every action and review any errors with the project lead. Recurring evaluations with known test cases make drift visible before it causes a missed review.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →