automatic.co

Discovery Build and Run Costs for a First Production Agent

A line-item breakdown of what a first production AI agent costs across discovery, build, and run, with the ranges, drivers, and traps that move the number.

Eric Lamanna6 min read
A three-tiered stepped structure representing discovery, build, and run phases of agent construction.

A first production agent's "$50,000 build" line is not a year-one budget. It is the middle slab of a figure that ranges from roughly $10,000 to over $500,000 once discovery and run are added, and the spread is almost entirely tool count, approval depth, and deployment target.

Discovery sits in the low five figures and decides how much of the build gets repriced mid-sprint. Build is the predictable middle, governed by three variables. Run compounds every month the agent is in production, and it holds the line items that get underestimated most often.

Capex versus opex turns on whether the work qualifies as application-development under ASC 350-40. The rest is variance inside each band.

Discovery Prices the Rest of the Build

Discovery on a production agent specifies the workflow down to the tools, approval tiers, data sources, and failure modes the build team will code against. A tight engagement of two to three weeks with a workflow owner, an architect, and a data engineer takes a small share of total project cost, and that small share decides how much of the rest gets repriced during sprints.

The deliverables that earn the fee are specific:

  • A tool inventory with authentication, rate limit, and idempotency notes per endpoint.
  • Approval tier mapping per action class: read, suggest, write with review, write autonomously.
  • Data sources with lineage, freshness SLAs, and the system of record for each field.
  • Evaluation criteria with pass/fail thresholds and a labeled golden set.
  • A failure-mode register with the containment response for each class.

Without those artifacts, build teams rediscover the same decisions at engineering rates and re-architect once the first production incident reveals the missing decision. AI feasibility modeling earns multiples of its fee by closing those decisions before contract, when changes are still ink-level cheap rather than refactor-level expensive.

From Any AI Use to Measurable EBIT Impact
From Any AI Use to Measurable EBIT ImpactRegularly use AI in at least one function: 88%; Report EBIT impact at enterprise level: 39%; High performers (>5% of EBIT from AI): 6%88%Regularly use AI in at le…39%Report EBIT impact at ent…−56% from Regularly use AI…6%High performers (>5% of E…−85% from Report EBIT impac…
Discovery discipline is what separates the 88% who use AI from the 6% capturing material EBIT. Source: McKinsey, 2025

Build Cost Follows Three Variables

Build cost has a predictable shape: a narrow production agent lands in the low end, a RAG-backed agent sits in the middle, and a regulated multi-agent system occupies the top. The spread inside each band is governed by three variables.

  • Tool count. Each integration adds engineering hours for authentication, idempotency, retry logic, and lineage capture. Modern REST systems come in at a few thousand per connector. Legacy ERPs without clean APIs can double or triple that, and some add a middleware purchase on top.
  • Approval depth. A read-only agent needs no human-in-the-loop UI. A three-tier approval stack with action-class routing adds a review queue, notification plumbing, and audit storage. Approval gates are a distinct deliverable, not a free byproduct of the agent loop.
  • Deployment target. SaaS inference is cheapest. VPC inference adds image hardening and network controls. Air-gapped deployment adds model weight custody, on-prem GPU provisioning, and offline update procedures. The cost difference between SaaS and air-gapped on the same workflow often runs 2x to 3x.

Evaluation and testing are the build line most frequently underestimated. Shipping without a baseline eval harness turns every subsequent model or prompt change into an uncontrolled experiment, and drift goes unnoticed until the first wrong write hits a production system.

An engineer arranging interconnected gears representing the tool and integration surface of a production agent.

Run Costs Compound Month by Month

Run cost has three components, each on a different curve.

  • Inference. Token spend scales with action volume and context length. A light single-purpose agent runs in the low hundreds per month. A fully autonomous multi-agent system runs into five figures.
  • Hosting and observability. A VPC deployment with structured logging, distributed tracing, and metric storage carries a steady monthly load. Air-gapped adds amortized GPU licensing and dedicated storage.
  • Maintenance. Prompt updates, model refreshes, integration drift, and eval recalibration consume engineering hours every quarter. Teams that pencil in 5% of build cost for maintenance overrun by Q3.

The underestimated items are almost always on the run side: eval rerun cadence after each prompt or model change, on-call rotation across the first 90 days, and integration-drift hours that accumulate every time an upstream system silently changes a field. PwC's survey of US business executives found that 88% of agent adopters plan to increase AI budgets over the next year, which suggests the growing line on deployed agents is run spend, not build.

Top 5 OWASP LLM Application Risks a Production Agent Must Guard Against
Top 5 OWASP LLM Application Risks a Production Agent Must Guard AgainstPrompt Injection: LLM01; Sensitive Information Disclosure: LLM02; Supply Chain: LLM03; Data and Model Poisoning: LLM04; Improper Output Handling: LLM051Improper Output HandlingLLM052Data and Model PoisoningLLM043Supply ChainLLM034Sensitive Information DisclosureLLM025Prompt InjectionLLM01
Each of these drives guardrail, monitoring, or eval-rerun line items on the run side; underscoping them at build time shifts the spend, not the total. Source: OWASP, 2025

Capex and Opex Split by Stage

Discovery and prototyping are opex under most US GAAP policies. Backend engineering, integration code, and QA during the application development stage can be capitalized as internal-use software under ASC 350-40 when the organization has an established capitalization policy and the project meets the specified criteria. Implementation training, data preparation, and post-go-live maintenance are expensed.

A $120,000 first-year agent typically splits roughly as:

  • Discovery, $5,000 to $15,000, opex, period expense.
  • Build engineering, $30,000 to $60,000, potentially capitalized as application development.
  • Data preparation, QA, and training, $15,000 to $25,000, opex.
  • Run cost over year one, $25,000 to $50,000, opex.

The split matters because capitalized agent cost depreciates over its useful life, usually three years for custom-built software, while opex hits the current period in full. That changes EBITDA treatment, and it sometimes changes which approval tier the project crosses. A five-year TCO model is useful for the board conversation but blurs these phase-level line items; procurement wants the line items.

Agent Build Tiers on Cost and Calendar Time
Agent Build Tiers on Cost and Calendar TimePoC / AI pilot: 8; Narrow production agent: 9; Agent with RAG: 21; Advanced agentic system: 43; Regulated multi-agent: 76Relative build cost →Relative timeline →123451PoC / AI pilot2Narrow production agent3Agent with RAG4Advanced agentic system5Regulated multi-agent
All five tiers sit on the same diagonal: there is no fast-but-expensive shortcut. Complexity costs both money and calendar weeks. Source: Master of Code, 2026

Pricing Against the Right Band

Price the agent against the band that matches its production scope, not the band the pilot proved out.

  • Narrow band. One workflow, one or two tool calls, read-only or low-tier approvals, SaaS or VPC deployment. Build in the tens of thousands, run in the low hundreds per month.
  • RAG band. Retrieval over a controlled corpus, three to six tools, tiered approvals, VPC deployment. Build in the mid five figures to low six figures, run in the low thousands per month.
  • Regulated band. Orchestration across agents, legacy ERP integration, full approval stack, VPC or air-gapped. Build starts in six figures and climbs, run in the mid four to low five figures per month.

Pick the narrow band when the workflow has one owner, a bounded tool count, and reversible actions. Pick the RAG band when retrieval quality is the primary technical risk and the approval stack is tiered but shallow. Pick the regulated band only when air-gapped deployment or SOC-grade audit is a procurement requirement, not a preference. A pilot that worked in the narrow band and now adds retrieval or deeper approvals should be repriced against the RAG band from the first day of discovery; the delta between the two is usually 3x, not 1.5x, and the extra 50% rarely survives a second scoping round.

written by
Eric Lamanna
Director of Business Development

Eric Lamanna is a Digital Sales Manager with a strong passion for software and website development, AI, automation, and cybersecurity. With a background in multimedia design and years of hands-on experience in tech-driven sales, Eric thrives at the intersection of innovation and strategy—helping businesses grow through smart, scalable solutions. He specializes in streamlining workflows, improving digital security, and guiding clients through the fast-changing landscape of technology. Known for building strong, lasting relationships, Eric is committed to delivering results that make a meaningful difference. He holds a degree in multimedia design from Olympic College and lives in Denver, Colorado, with his wife and children.

Put an agent to work, the right way.

Start on Automatic and put the workflow you want to automate in front of engineers who have shipped agents in regulated environments.

Explore services
the briefing

Agentic AI, in your inbox.

Occasional, high-signal notes on building and operating AI agents — automation patterns, architecture, and governance. No spam.