When do we need an AI agent instead of a simpler automation?

An agent can be useful when a task involves interpreting varied inputs, deciding which information to retrieve, or adapting the next step to what it finds. A predictable task with clear rules may be better handled by a conventional automation. We assess the workflow before choosing the approach.

Is this a finished product or a custom build?

This is a custom implementation service. Build Lean can configure existing components, integrate your systems, and develop the tools, workflow logic, and interfaces needed for the agreed scope. The proposal identifies those components and their dependencies.

Can an agent take actions without approval?

It can be designed to carry out specified actions automatically within defined permissions and rules. Other actions can require a person to approve the exact proposed change. We agree those boundaries for each workflow and implement the checks in the application and tools.

Which systems and documents can it use?

That depends on the APIs, exports, permissions, licensing, and document formats available. We assess both retrieval and update capabilities. An integration may use an existing connector or require custom development. Document search may also need extraction, indexing, and revision information.

Are answers always accurate and supported by sources?

No. Retrieval gives the model relevant information, but it can still miss material or interpret it incorrectly. We define how source references, missing information, and unresolved questions are presented, then test the system against representative tasks. Human review remains part of workflows where the consequences require it.

Can we change the AI model later?

The architecture can be designed to reduce dependence on a single provider. Changing models can still require adjustments to prompts, tool use, supported features, and data handling. A replacement should be evaluated against the workflow’s quality, response-time, and cost requirements before rollout.

Where does our data go?

That depends on the selected hosting, model provider, retrieval services, and logging configuration. Information retrieved for a task may be sent to the model provider for processing. The design should document which data each component receives, how access is controlled, and the applicable retention arrangements.

What happens when the agent gets stuck or an update fails?

The workflow defines stopping conditions, retry limits, and escalation behavior. Missing information can trigger a request for clarification. Failed updates should return a visible status. If an action may have succeeded before a connection failed, the system should check the result before retrying to avoid duplicate changes.

Who owns the build, and what costs continue?

Ownership, access, and handover are documented in the agreement, alongside the terms of third-party and open-source components. Ongoing costs can include model usage, hosting, search or storage services, software licenses, monitoring, and maintenance. The proposal separates implementation from recurring services and support.

AI Agents

Custom AI agents for your business workflows

Build Lean develops AI agents that retrieve business records, prepare drafts, and use defined tools to carry out tasks. We combine AI models, integrations, and application rules around a specific workflow, with clear limits on access, actions, and human approval.

Book a Call See how it works

What It Is

What is an AI agent?

An AI agent uses a large language model, or LLM, to interpret a request, choose a next step, and use tools to work toward an objective. It can retrieve information, assess what is missing, and prepare an answer or proposed action.

The model chooses steps within the tools and limits the application provides. Application code checks permissions, validates inputs, and enforces approval requirements. A run can finish, ask for more information, or stop for human review.

Build Lean designs and builds this system around your workflow. That can include configuring existing components, connecting source systems, developing custom tools, and building the interface your team uses to review results.

Result returns to the agent

Request or trigger

A question, message, or scheduled event

Agent

Interprets, selects tools, drafts

Checked tool call

Permissions, validation, approval rules, run limits

Records or action result

Returned from the connected system

Possible outcomes and handoffs

Answer or draftAwaiting human reviewConfirmed updateStopped or handed off

Request or trigger
A question, message, or scheduled event
Agent
Interprets, selects tools, drafts
Checked tool call
Permissions, validation, approval rules, run limits
Records or action result
Returned from the connected system

The result returns to the agent, which can make another checked tool call or finish.

Possible outcomes and handoffs

Answer or draftAwaiting human reviewConfirmed updateStopped or handed off

Illustrative agent flow. Available tools, access rules, approval steps, and stopping conditions are defined for each build.

How this fits with your other systems

An agent can use a Business MCP Layer to access defined tools and a Data Lakehouse to retrieve prepared business data. A broader Agentic Operations Platform can provide shared interfaces, approvals, and workflows around multiple capabilities. The architecture depends on the task and the systems already in place.

Explore Business MCP Layers Explore Data Lakehouse Explore Agentic Operations Platform

Technical details: Retrieval, tools, and orchestration

Retrieval: Supplies information for the task. Document search may use retrieval-augmented generation, or RAG, to provide relevant passages to the model. Search can combine exact terms, filters, and semantic matching. Structured records may be retrieved through APIs or database queries.
Tools: Expose specific operations, such as finding a project, reading an invoice, or creating a task. Each tool has defined inputs, outputs, permissions, and error handling. MCP can provide a standard interface where it suits the selected application.
Orchestration: Controls how the task progresses. It manages tool results, review steps, retries, and limits on time, cost, or number of steps. Some stages follow fixed rules; others allow the model to choose what to do next.

Benefits

What a custom AI agent can help your team do

Reduce the searching, preparation, and repeated handoffs involved in a defined business workflow.

Bring the relevant information together

Retrieve the records and documents needed for a task from selected sources. Give your team a starting point with record IDs, document references, and available timestamps they can check.

Prepare work for review

Turn source information into a draft response, task, or summary. Show the supporting material and unresolved questions so a reviewer can approve, edit, or reject the proposed next step.

Carry approved work into your systems

Connect defined actions to the tools your team already uses. Where updates are included, validate the request, apply the required approval rules, and return the resulting record ID or failure status.

Keep the system understandable and maintainable

Document the model, integrations, access rules, and operating responsibilities. Select models against the task’s quality, response-time, cost, and data requirements. Evaluate changes before introducing them into the workflow.

By Industry

How an AI agent could support your industry

Illustrative workflows showing how an agent could retrieve information, prepare work, and use permitted actions. Sources, integrations, and review requirements are assessed for each project.

01 · Construction

Turn a project question into a draft RFI

A project manager asks about a discrepancy between a drawing and a specification. An agent could retrieve the relevant documents, identify the revisions recorded in the source system, and prepare a draft request for information with references and unresolved questions.

The project manager reviews the interpretation and wording. If included in the build, an authorized tool can create the approved draft in the project system and return its RFI ID.

DrawingsSpecificationsRFIsProject records

See General Contractors

02 · Real Estate

Prepare maintenance requests for action

An agent could read a maintenance message, retrieve the relevant property and unit records, and check for related open work orders. It can summarize the issue, identify missing information, and propose a category or follow-up task.

Application rules determine routing and required review. Where included, a permitted tool can create or update a work order after the required checks and approval.

Maintenance messagesPropertiesUnitsWork orders

See Property Management

03 · Legal

Prepare a matter summary and follow-up tasks

An agent could retrieve permitted matter documents and correspondence, prepare a summary with source references, and identify questions that need an attorney’s attention.

An attorney reviews the interpretation and proposed tasks. Creating a task or saving an approved note can be included as a separate permitted action. Document access, external communications, and filing actions require their own scope and controls.

Matter recordsCorrespondenceContractsTasks

See Legal

These examples describe possible implementations. Available records, document quality, permissions, and supported system actions determine the scope.

How It Works

Illustrative Workflow

From a project question to a draft RFI

1Request
Start With the Question and Project
A project manager asks: “The drawing and specification appear to disagree on this item. Gather the references and prepare an RFI draft.”
The application checks the requester’s project access. The agent identifies the documents and details needed to investigate the question.
2Retrieve
Retrieve and Compare the Relevant Material
The agent selects permitted search and lookup tools to retrieve drawing references, specification passages, and recorded revision information.
It can make another lookup when a reference points to additional material, within the run’s defined limits. Missing documents or conflicting revision information are surfaced for review.
3AI DraftHuman Review
Prepare the Draft for a Person to Assess
The agent drafts the question, includes the supporting references, and identifies any assumptions or unresolved points.
The project manager reviews the source material and edits the draft. The proposed RFI requests clarification; it does not authorize a design change or direct work in the field.
4Human ReviewSystem Update
Create the Reviewed Record and Confirm the Result
If RFI creation is included, the authorized reviewer approves the exact details to be submitted. The application validates required fields, checks for duplicate requests, and calls the project system.
A successful result includes the new RFI ID or link. A failed or uncertain result remains unresolved until its status is checked. The system does not report completion simply because the agent produced a draft.

Who does what

The AI selects retrieval steps and generates draft wording. Application rules enforce access, required fields, approval, and execution limits. The reviewer approves the content. The connected system confirms whether the record was created.

Illustrative workflow. The source documents, tool operations, approval process, and exception handling would be defined for the selected project system.

How we build your agent

Start with one workflow, define what a successful result looks like, and test the full path from request to confirmed outcome.

01 / 06
Define the Task and Boundaries
Identify the users, trigger, source records, expected output, and actions the agent may take. Agree where human review is required and when the system should stop or ask for help.
02 / 06
Connect the Relevant Information
Assess APIs, document access, data quality, and available connectors. Configure retrieval around the sources, including record matching, document references, and refresh requirements.
03 / 06
Build the Tools and Workflow Logic
Configure suitable existing components and develop the missing integrations or tools. Define their inputs, outputs, and limits. Use fixed rules for predictable steps and model-driven choices where the task benefits from them.
04 / 06
Implement Access and Approval Controls
Define who can retrieve each kind of information and request each action. Enforce restrictions in the application and connected tools. Bind approvals to the proposed update and define how changed records or repeated requests are handled.
05 / 06
Evaluate Answers, Actions, and Exceptions
Test representative tasks against agreed expected results. Check retrieval quality, source references, tool selection, unauthorized requests, missing information, and failed updates. Measure task completion, review corrections, response time, and operating cost.
06 / 06
Deploy and Define Ongoing Support
Document the configuration, hosting, dependencies, and handover. Agree who monitors failures, maintains integrations, reviews model changes, and supports the team after launch.

Start with one recurring task

Choose a task that repeatedly sends someone searching for information, preparing a draft, or copying details into another system. We assess the available sources and actions, then scope a first release your team can evaluate.

Your proposal identifies what will be configured, integrated, and custom-built. It also sets out testing, deployment, documentation, and operating responsibilities.

A defined implementation can be scoped as a Custom Software project. An AI Transformation partnership can support ongoing priorities, rollout, and adoption.

Book a Call Explore Custom Software Explore AI Transformation

FAQs

AI agent questions

What to consider before introducing an agent into your business workflows.

Start With One Workflow

Which recurring task should your team spend less time on?

Tell us what starts the task, where the information lives, and what a completed result should look like. We’ll discuss the systems involved, the review steps, and a practical first scope.

Book a Call

Free · 30 minutes · No preparation required