AI Agent Development · Los Angeles

AI agent development in Los Angeles.

Most LA agent conversations come down to a spreadsheet. Manila BPO at $8 to $12 an hour fully loaded. Salesforce Agentforce seat licensing. Intercom Fin per-resolution pricing. A Bangalore-shoring contract with the Big Four. A custom agent billed at $0.05 to $0.20 a ticket once it is in production. The right answer depends on volume, deterministic share, and the regulatory posture you live under.

We do the cost math honestly. For a DoorDash-scale support queue handling several million tickets a month, a custom agent running tier-1 deflection is a seven-figure annual saving versus BPO. For a Riot Games player-support stack at Santa Monica, the picture is different because of the toxicity-appeal workflows that no off-the-shelf product handles well. For a SpaceX supply-chain workflow under ITAR, BPO is not even an option and the comparison is custom-agent vs current-state.

Fixed-scope projects are quoted per engagement based on scope. Discovery first, fixed quote after, four to eight weeks of build with weekly demos on Pacific time.

How we work

Run the cost math with us.

What an agent actually costs vs the alternatives.

Start with the BPO baseline. A tier-1 support seat at Concentrix, Teleperformance, or TaskUs in Manila or Cebu bills out at roughly $8 to $12 fully loaded per hour. A trained seat handles 6 to 10 tickets per hour depending on the channel and the complexity. That puts the per-ticket cost at roughly $0.80 to $2.00. Add the contract minimums, the management fee, and the QA overhead, and the all-in number rises another 15 to 25%.

Now the agent. Token cost for a tier-1 ticket handled by Claude Sonnet 5 or GPT-5, with a system prompt, a few retrieval results, two or three tool calls, and a structured response, runs roughly $0.02 to $0.10 in model spend. Add infrastructure (compute, vector store, observability) and the all-in marginal cost lands at $0.05 to $0.20 per ticket. The fixed cost is the build (quoted per engagement) plus ongoing eval maintenance and model-spend at the margin.

On a 50,000 ticket-per-month queue, BPO at $1.20 average is $720K a year. Agent at $0.12 average is $72K a year. The agent does not handle every ticket cleanly: figure 30 to 50% of tickets escalate to a human in year one, dropping as the eval set matures. Even with that escalation, the math works above roughly 25,000 tickets a month. Below that, BPO often still wins on pure cost.

The off-the-shelf agent products are a different comparison. Salesforce Agentforce charges per conversation or per resolution depending on the SKU; at $2 per resolution and a 50% resolution rate, the per-ticket effective cost lands near BPO territory. Intercom Fin charges $0.99 per resolution; same shape. Zendesk AI is bundled with seat licensing. These work when the vendor schema matches your data and you are happy to live inside their platform. They stop working when you need to call a content-ID system, a player-account database, or an ITAR-bounded supplier file.

The cases where the custom agent is clearly right: high volume (50K+ tickets/month), unique tool integrations (a content workflow, a dispatcher tool, an internal pricing engine), data that cannot leave your VPC, or a regulatory posture that excludes offshoring. The cases where BPO or off-the-shelf is clearly right: low volume, standard CRM schema, no custom tools, no compliance perimeter.

We model your specific numbers in the discovery call. If the math does not work, we say so. There is no point shipping an agent that costs more than the BPO it replaces.

What we build for Los Angeles clients.

Six components that show up in every agent we ship for media, gaming, aerospace, and delivery-ops teams in Southern California.

Custom agent loop

LangGraph state machines for plan-execute or ReAct flows. Transitions are explicit. The graph is reviewable and testable in CI, not a 300-line prompt that nobody owns.

Tool integrations to LA-specific systems

Salesforce, Zendesk, Snowflake, Stripe, Adyen, content-ID systems (Disney, NBCU patterns), game-account databases, DoorDash-style dispatcher APIs, aerospace supplier portals, AS9100 quality systems.

Cost telemetry per workflow

Every agent run reports token cost, tool-call cost, and downstream system cost. Dashboards by feature, channel, and customer cohort. You can see what each part of the agent actually costs without rebuilding accounting.

Human-in-the-loop gates

Refunds above a threshold, ban appeals, content takedowns, ITAR-controlled approvals, supplier deviation decisions. Any irreversible action routes to a human with the full reasoning trace and assembled case file.

Eval harness with cost gating

Task completion rate, hallucination rate, tool-call accuracy, AND cost-per-resolution. CI fails if cost-per-resolution drifts above a threshold on the eval set. Cheaper-model fallback for prompts where Claude or GPT-5 is overkill.

Compliance perimeter for regulated work

ITAR-bounded deployments on US-only infrastructure, PII tokenization for media post-production workflows, SOC 2 audit logging, and dedicated tenancy for clients who cannot share inference infrastructure.

Why this work fits Los Angeles.

LA stacks four industries where the agent vs alternative math is the central conversation. Media and entertainment (Disney, NBCU, Warner Bros. Discovery, Netflix, Snap) runs content-ID, royalty-claim review, and post-production workflows that no BPO can touch because the IP cannot leave the firm. Gaming (Riot in Santa Monica, Activision in Santa Monica, Naughty Dog, Insomniac) runs player support at consumer-platform scale with the platform and ESRB constraints that off-the-shelf products do not handle.

Aerospace (SpaceX in Hawthorne, Northrop Grumman in El Segundo, Boeing in Long Beach, Aerospace Corporation) runs supplier qualification and quality-record assembly under ITAR and AS9100 constraints that make offshoring legally impossible. Delivery ops (DoorDash dispatcher escalation, food-delivery dispute handling) runs the most volume-heavy queues in the city, where tier-1 deflection economics tilt strongly toward custom agents.

We work remotely with LA clients on Pacific time. Discovery calls fit between 8am and 6pm PT, demos happen weekly, and the codebase ships to a repository you own. No agency-style retainer that locks you into our infrastructure forever.

Industries where we see strongest fit: media post-production ops, gaming player support and trust-and-safety, aerospace supplier and quality workflows, food-delivery and gig-economy dispatcher ops, and high-volume e-commerce support where BPO has been the default.

Questions from Los Angeles clients.

How does an agent compare to a Manila BPO contract for tier-1 support?: Two ways to look at it. On unit cost, a Manila BPO agent at a major operator (Concentrix, Teleperformance, TaskUs) bills out at roughly $8 to $12 fully loaded per hour. A tier-1 support seat handles 6 to 10 tickets per hour depending on complexity, putting the cost at roughly $0.80 to $2.00 per ticket. A well-built agent runs $0.05 to $0.20 per ticket in token and infrastructure cost for the same tier-1 scope. On softer factors, BPO carries hiring lag, attrition (60 to 100% annual at most centers), training time, and time-zone constraints. The agent carries model drift risk, eval maintenance cost, and a higher ceiling on the share of tickets it cannot handle. We model both honestly in the discovery call. The agent is usually right when ticket volume exceeds 50,000 a month and the deterministic path is more than 40% of the queue. Below that, BPO often still wins.
Why build a custom agent instead of using Salesforce Agentforce, Intercom Fin, or Zendesk AI?: Off-the-shelf agent products work when your workflow fits their integration surface and your data lives where they expect it. Intercom Fin is excellent for self-serve deflection inside an Intercom-native support stack. Salesforce Agentforce works when the agent only needs Salesforce objects and your team will operate it inside the Salesforce platform. Zendesk AI is competent for response drafting and macro suggestion. They become a constraint when the workflow has to call out to a custom internal API (a content-ID system at Netflix, a player-account database at Riot, a dispatcher tool at DoorDash), when the data model does not map to the vendor schema, or when you need observability and eval gating that the vendor does not expose. We have replaced and complemented all three in different deployments. The vendor product is rarely strictly cheaper once you factor in seat licensing at scale.
What does an aerospace supply-chain agent look like under ITAR constraints?: ITAR (22 CFR Parts 120-130) restricts technical-data transfer to non-US persons and to non-US infrastructure. The agent runs entirely on US-based, US-person-operated infrastructure: Azure OpenAI in a US-only region with the appropriate compliance posture, or a self-hosted Llama or Qwen model on hardware inside your data center. PII and technical data never leave the boundary. The agent assists with supplier-qualification document review, deviation-request triage, and quality-record assembly under AS9100 and the prime-contractor program requirements. The export-control officer reviews every assembled package before submission. We have scoped this kind of work for SpaceX-adjacent suppliers in the South Bay and Long Beach; the architecture is well-understood once the compliance perimeter is drawn.
Can an agent handle gaming player support at Riot or Activision scale?: Yes, with realistic scope. Gaming player support combines high-volume tier-1 queries (account recovery, payment disputes, in-game item issues) with a long tail of complex cases (account bans, toxicity reports, anti-cheat appeals, refund disputes). An agent owns the tier-1 layer with read access to the player account database, payment processor, and game-state systems. Common ban appeals and refund requests route to a human reviewer with the agent's assembled case file (account history, prior actions, evidence). The platform-specific rules (PlayStation, Xbox, Steam refund windows; ESRB constraints) are encoded in the agent's decision rubric. For toxicity and anti-cheat appeals, the agent assembles evidence but never auto-decides because the player-experience cost of a wrong call is high. We run the deflection rate against the eval set monthly so you can see what the agent is and is not handling.

Run the cost math with us.

Tell us your ticket volume, your current BPO or off-the-shelf cost, and the workflow shape. We'll reply within one business day with a modeled cost comparison and a price range. If the math does not work, we will say so.

AI agent development in Los Angeles.

Fixed-scope projects are quoted per engagement based on scope. Discovery first, fixed quote after, four to eight weeks of build with weekly demos on Pacific time.

What an agent actually costs vs the alternatives.

We model your specific numbers in the discovery call. If the math does not work, we say so. There is no point shipping an agent that costs more than the BPO it replaces.

What we build for Los Angeles clients.

Six components that show up in every agent we ship for media, gaming, aerospace, and delivery-ops teams in Southern California.

Custom agent loop

LangGraph state machines for plan-execute or ReAct flows. Transitions are explicit. The graph is reviewable and testable in CI, not a 300-line prompt that nobody owns.

Tool integrations to LA-specific systems

Salesforce, Zendesk, Snowflake, Stripe, Adyen, content-ID systems (Disney, NBCU patterns), game-account databases, DoorDash-style dispatcher APIs, aerospace supplier portals, AS9100 quality systems.

Cost telemetry per workflow

Human-in-the-loop gates

Eval harness with cost gating

Compliance perimeter for regulated work

Why this work fits Los Angeles.

Questions from Los Angeles clients.

How does an agent compare to a Manila BPO contract for tier-1 support?

Two ways to look at it. On unit cost, a Manila BPO agent at a major operator (Concentrix, Teleperformance, TaskUs) bills out at roughly $8 to $12 fully loaded per hour. A tier-1 support seat handles 6 to 10 tickets per hour depending on complexity, putting the cost at roughly $0.80 to $2.00 per ticket. A well-built agent runs $0.05 to $0.20 per ticket in token and infrastructure cost for the same tier-1 scope. On softer factors, BPO carries hiring lag, attrition (60 to 100% annual at most centers), training time, and time-zone constraints. The agent carries model drift risk, eval maintenance cost, and a higher ceiling on the share of tickets it cannot handle. We model both honestly in the discovery call. The agent is usually right when ticket volume exceeds 50,000 a month and the deterministic path is more than 40% of the queue. Below that, BPO often still wins.

Why build a custom agent instead of using Salesforce Agentforce, Intercom Fin, or Zendesk AI?

Off-the-shelf agent products work when your workflow fits their integration surface and your data lives where they expect it. Intercom Fin is excellent for self-serve deflection inside an Intercom-native support stack. Salesforce Agentforce works when the agent only needs Salesforce objects and your team will operate it inside the Salesforce platform. Zendesk AI is competent for response drafting and macro suggestion. They become a constraint when the workflow has to call out to a custom internal API (a content-ID system at Netflix, a player-account database at Riot, a dispatcher tool at DoorDash), when the data model does not map to the vendor schema, or when you need observability and eval gating that the vendor does not expose. We have replaced and complemented all three in different deployments. The vendor product is rarely strictly cheaper once you factor in seat licensing at scale.

What does an aerospace supply-chain agent look like under ITAR constraints?

ITAR (22 CFR Parts 120-130) restricts technical-data transfer to non-US persons and to non-US infrastructure. The agent runs entirely on US-based, US-person-operated infrastructure: Azure OpenAI in a US-only region with the appropriate compliance posture, or a self-hosted Llama or Qwen model on hardware inside your data center. PII and technical data never leave the boundary. The agent assists with supplier-qualification document review, deviation-request triage, and quality-record assembly under AS9100 and the prime-contractor program requirements. The export-control officer reviews every assembled package before submission. We have scoped this kind of work for SpaceX-adjacent suppliers in the South Bay and Long Beach; the architecture is well-understood once the compliance perimeter is drawn.

Can an agent handle gaming player support at Riot or Activision scale?

Yes, with realistic scope. Gaming player support combines high-volume tier-1 queries (account recovery, payment disputes, in-game item issues) with a long tail of complex cases (account bans, toxicity reports, anti-cheat appeals, refund disputes). An agent owns the tier-1 layer with read access to the player account database, payment processor, and game-state systems. Common ban appeals and refund requests route to a human reviewer with the agent's assembled case file (account history, prior actions, evidence). The platform-specific rules (PlayStation, Xbox, Steam refund windows; ESRB constraints) are encoded in the agent's decision rubric. For toxicity and anti-cheat appeals, the agent assembles evidence but never auto-decides because the player-experience cost of a wrong call is high. We run the deflection rate against the eval set monthly so you can see what the agent is and is not handling.

Run the cost math with us.