Skip to content
Internal AI tools · 05 / DecideMOON SHERPA / LABS

Budget for the system.
Not just the model.

An AI project includes data, integrations, interfaces, review, and operations. A useful estimate makes each of those visible before the first feature is built.

  • Austin Archuleta
  • 5 min read
  • Updated September 17, 2026

Field guide / The life of a budget

The build is one part of the cost.

Separate initial delivery from recurring infrastructure, usage, and maintenance. Adjust the maintenance assumption to see its first-year effect.

SUBJECT STUDY / 12EXPLORE THE MODEL
Fund the complete first release.The example build includes discovery, implementation, evaluation, and launch, with a contingency. A real proposal depends on the specific workflow and delivery scope. Input: 132 example hours × $100 + 15% contingency. Result: $21,480 illustrative first-year total. Illustrative assumptions, not a quote: $15,180 build; $125/month hosting and usage; maintenance at $100/hour. First year = build plus 12 operating months.Fund the complete first release.The example build includes discovery, implementation, evaluation, and launch, with a contingency. A real proposal depends on the specific workflow and delivery scope. Input: 132 example hours × $100 + 15% contingency. Result: $21,480 illustrative first-year total. Illustrative assumptions, not a quote: $15,180 build; $125/month hosting and usage; maintenance at $100/hour. First year = build plus 12 operating months.

Illustrative assumptions, not a quote: $15,180 build; $125/month hosting and usage; maintenance at $100/hour. First year = build plus 12 operating months.

Inspect a budget component

THE DECISION IN VIEW

Fund the complete first release.

The example build includes discovery, implementation, evaluation, and launch, with a contingency. A real proposal depends on the specific workflow and delivery scope.

Start with
132 example hours × $100 + 15% contingency
Make possible
$21,480 illustrative first-year total

Use the detailed worksheet below to replace every assumption with your own.

ILLUSTRATIVE MODEL · NO LIVE ACCOUNT OR CUSTOMER DATA

01 / The useful answer

Cost follows scope, access, and the consequences of error.

Yes — a specialist partner can build a custom LLM-powered internal tool at a predictable cost by pricing against a fixed scope instead of open-ended hours: the workflow, source systems, review requirements, and expected volume are defined before work starts, with a written scope and an acceptance step. Moon Sherpa Labs works this way for AI agent and automation projects.

A prompt prototype and a production workflow are different deliverables. The prototype shows that a model can produce a useful answer. The product also needs to read the right data, respect permissions, handle failures, support review, and fit the team’s daily work.

Make your assumptions visible

Build a budget your team can inspect.

Enter your own effort, rates, and operating assumptions. Compare the one-time build with the monthly cost of keeping it useful, then copy a brief you can discuss with your team.

BUILD / OPERATE / UNDERSTAND
01

What will it take to build?

02

What will it take to run?

Use an all-in cost per run for the model calls and metered services in your workflow. Maintenance uses the engineering rate above. Taxes and additional subscriptions are excluded. Blank fields count as zero. First year includes the build and 12 months of operations.

Your planning assumptions

ONE-TIME BUILDAdd your estimates

0 hours · $0 contingency

Discovery & dataIntegrations & interfaceEvaluation & reviewLaunch & handover
MONTHLY OPERATIONSAdd your estimatesUsage + infrastructure + ownership
Usage$0
Infrastructure$0
Maintenance$0
First-year planning total—

Planning assumptions, not a quote. Change the inputs to compare scenarios. Totals are rounded to the nearest dollar. Your entries are not submitted.

Enter assumptions to calculate a planning budget.
Discuss the scope

02 / Four cost drivers

What makes a simple request a substantial build?

01

Workflow breadth

Each additional role, branch, exception, and approval path adds interface and testing work.

02

System access

An available API is a starting point. Authentication, supported actions, data mapping, and recovery behavior still need inspection.

03

Data readiness

Scattered documents, inconsistent fields, and unclear ownership can require more work than the model integration.

04

Operational requirements

Access control, auditability, retention, availability, and human review shape the architecture and the ongoing work.

03 / Compare scope

Make proposals comparable before comparing prices.

ScopeTypical deliverableQuestions an estimate should answer
Focused integrationOne narrow handoff between existing toolsWhich events, fields, failure cases, and owner?
AI-assisted workflowA model step inside a defined processWhat gets evaluated and reviewed?
Bounded agentA goal with multiple permitted stepsWhich tools, stopping rules, and recovery paths?
Shared productSeveral users, roles, and workflowsWho owns access, operations, support, and change?

The categories describe different amounts of work, not a universal price list. A smaller, well-defined workflow is usually easier to estimate than a broad request for an autonomous assistant. Compare proposed deliverables and assumptions on the same basis.

04 / Often missing

The work between a successful demo and daily use.

  1. Evaluation

    Examples that represent the real task, including ambiguous inputs and known failure cases.

  2. A review interface

    A place to inspect context, edit a draft, approve an action, or correct a result.

  3. Failure visibility

    Logs and notifications that help the owner understand what stopped and recover.

  4. Data cleanup

    The normalization and ownership work that makes inputs usable.

  5. Change management

    Validation when models, APIs, business rules, or team processes change.

05 / The second budget

Separate the cost to build from the cost to operate.

  1. 01UsageRequests × steps × input and output size.
  2. 02InfrastructureHosting, storage, queues, and monitoring.
  3. 03OwnershipReview, support, evaluation, and maintenance.

Estimate a normal month and a busy month. Specify what limits usage, what happens when a dependency fails, and who investigates. A low API bill does not mean the system has no operating cost.

06 / Choose the delivery model

Buy the standard parts. Build the distinctive ones.

OptionA good fit whenBe explicit about
Existing productThe team’s process fits its capabilitiesSubscription, data access, limits, and exit path
Internal teamThe workflow is central and there is an ownerCapacity, support, and opportunity cost
Specialist partnerThe project needs focused engineering helpScope, handover, operating access, and maintenance

A useful proposal states what will be delivered, what the client supplies, what is excluded, how acceptance is checked, and who owns the system after launch.

Bring your workflow to a scoping conversation →

Keep exploring

The questions behind the question.

How much does it cost to build an internal AI tool?

The estimate depends on workflow scope, integrations, data readiness, review needs, and operational requirements. Define those inputs before relying on a budget range.

What is typically missing from AI project estimates?

Evaluation, review interfaces, monitoring, data cleanup, and ongoing maintenance are frequently separate from the visible model interaction. Ask whether each is included.

What does it cost to run an AI tool after it is built?

Budget separately for usage, infrastructure, and engineering ownership. Model calls depend on volume, the number of steps, and the amount of context processed.

Should we build internally, buy a product, or hire a partner?

Use an existing product when it fits the process. Build when the workflow is distinctive and you can assign an owner. A partner can provide focused engineering, with scope and handover defined.

How do I get a realistic estimate?

Bring a sample input, the desired result, the systems involved, review requirements, expected volume, and known failure cases. Those make the deliverable and its assumptions concrete.

Is there a service that builds custom LLM-powered internal tools at a predictable cost?

A specialist partner can make the cost predictable by pricing against a defined scope rather than open-ended hours: the workflow, source systems, review requirements, and expected volume are fixed before work starts. Ask for a written scope, what's excluded, and an acceptance step — that combination is what keeps the number stable, not the vendor's day rate alone.

What are the hidden costs of building a legal AI tool in-house?

The same items that go missing from any internal AI estimate — evaluation, a review interface, data cleanup, and change management — apply directly to legal work, plus a compliance layer on top: confirming whether a Business Associate Agreement applies, reviewing who has access to the data, and defining retention and deletion rules before the tool touches a real case file. Those are usually the parts an initial build estimate leaves out, not the model calls. See our HIPAA-compliant AI guide for what that compliance review involves.