AI Agent Development Company in Portland

Ranges, because scope drives the number and anyone quoting a fixed price before discovery is guessing. A scoped proof of concept covering one agent, three to five tools, an evaluation suite, an action-log view and a hosted demo behind your SSO typically runs USD 30,000 to 70,000 over six to ten weeks. A production agent for a Portland enterprise, meaning failure-analysis triage for a fab, prior-authorization drafting for a health system, or supply-chain exception handling for a footwear brand, typically runs USD 110,000 to 300,000 including guardrails, a human-in-the-loop approval console, observability and the OCPA and HIPAA documentation pack. A multi-agent platform with role-specialized agents, shared memory, custom evaluators and integration into Epic, SAP, ServiceNow or an MES landscape runs USD 300,000 to 800,000 and up across phases. Portland senior engineering rates sit meaningfully below San Francisco and modestly below Seattle, which is one reason Intel, Nike and AWS all staff engineering here. Air-gapped fab work costs more than cloud work because the GPU capacity, the deployment path and the security review all sit inside the customer's perimeter. We quote fixed fee against a signed statement of work.

Oregon has no comprehensive private-sector AI statute of the Colorado or Texas kind, and any vendor telling you otherwise is selling fear. What Oregon has instead is a set of narrow, targeted statutes plus a privacy law that reaches AI systems directly. SB 1546, the AI companion law, cleared the legislature on March 5 2026, was approved by Governor Kotek on March 31 2026 and takes effect January 1 2027 as chapter 85 of Oregon Laws 2026. Read the definition before you assume it applies: an artificial intelligence companion is a system designed to simulate a sustained human-like platonic, intimate or romantic relationship that does all three of retaining prior-session information to personalize interactions, asking unprompted questions touching emotional topics, and sustaining ongoing dialog about matters personal to the user. Software operating solely for customer service or support, business operations, productivity or information analysis is expressly excluded. Operators in scope must run a suicide and self-harm protocol that refers users to the 988 lifeline, publish that protocol, and post an annual report of referral counts; the machine-identity disclosure, the reminder every three hours to take a break, and the bans on variable-schedule reward loops and guilt-inducing messages when a user tries to leave apply where the operator knows or has reason to believe the user is a minor. An individual who suffers an ascertainable loss or injury in fact may recover the greater of actual damages or 1,000 dollars per violation, plus attorney fees. HB 2748, effective January 1 2026, bars AI and other nonhuman entities from using nursing titles, and HB 2299, also effective January 1 2026, extended Oregon's intimate-image crime to digitally created and manipulated images. The binding day-to-day constraint on most commercial agents remains the Oregon Consumer Privacy Act.

Three provisions of the OCPA shape agent design, and none of them are satisfied by a policy page. Start with scope: the statute applies at 100,000 Oregon consumers, or 25,000 consumers where more than 25 percent of revenue comes from selling personal data, its obligations became operative July 1 2024 under SB 619, and it is enforced solely by the Oregon Attorney General with civil penalties up to 7,500 dollars per violation and no private right of action. The first design constraint is Oregon's list-of-specific-third-parties right, which no other state grants in this form: a consumer may obtain a list of the specific third parties, other than natural persons, that received personal data, and the statute lets the controller choose whether to name the third parties that received that consumer's data or the third parties that received any personal data. Either way you have to be able to produce a real list, so if your agent pipeline sends context to a model provider, a vector database and an observability vendor, all three belong in an answerable register. The second is opt-in consent for sensitive data, which in Oregon includes health condition, genetic and biometric data, precise geolocation, national origin, immigration status, and status as transgender or nonbinary. The third is universal opt-out mechanism recognition, mandatory since January 1 2026. We build deletion that propagates into vector stores, conversation memory and traces, not only the system of record.

Yes, and in the Silicon Forest it is the default rather than the exception. Process recipes, yield data, tool telemetry and design databases are treated as non-exportable to hosted frontier models, so the entire stack runs inside the customer perimeter: open-weight models such as Llama, Mistral or Qwen on on-premise or private GPU capacity, LangGraph orchestration inside an isolated VPC or plant network, retrieval against internal document stores through pgvector or Qdrant, guardrail classifiers running locally, and no outbound calls to any hosted model API. Connectors into MES, SPC, yield-management and tool-log systems are read-only by default. Any write, meaning a lot disposition, a recipe change or a maintenance order, sits behind a mandatory human approval gate with a named approver recorded in the action log. The trade-off is capability headroom against frontier models, which we manage with LoRA and QLoRA fine-tuning on internal corpora, aggressive retrieval grounding, and task decomposition that keeps each agent step inside what a 70B-class open model handles reliably. The same pattern serves Daimler and Precision Castparts plants where customer contracts prohibit external model calls.

For most retail and support agents, very likely nothing, and we would rather say that than sell you a compliance program you do not need. SB 1546 expressly excludes software that operates solely for customer service or support, business operations, productivity or information analysis, no matter how fluent its natural language is. On top of that carve-out the definition is conjunctive: an in-scope system must simulate a sustained platonic, intimate or romantic relationship and retain prior-session information to personalize interactions and ask unprompted questions that concern emotional topics and sustain ongoing dialog on matters personal to the user. A branded assistant that remembers a shopper's size and past orders to resolve service issues faster does not meet that test. The design question worth asking is whether anyone on the product side is pushing the assistant toward companionship as an engagement strategy, because that is the drift that moves a product into scope. So we write the intended scope down during design, keep persistent memory functional rather than relational, disclose machine identity at session start regardless of whether the statute compels it, and revisit the assessment whenever a persona or retention change lands. Where a client genuinely wants a companion-style product, we build the crisis-referral protocol, the published protocol page, the annual referral report, and the minor-triggered disclosure, three-hour break reminders and anti-compulsion safeguards in as product requirements from sprint one.

Every state-changing action is gated by default and we relax that only when a written risk review supports it. Tools sort into four tiers. Read-only tools carry no gate. Low-risk writes, such as drafting an internal note, get post-hoc review with sampling. High-risk writes, such as updating a system of record, sending an external customer message or filing a warranty claim, require a pre-call gate with one named approver. High-impact actions require two approvers under four-eyes review: anything writing to a fab MES or lot disposition, any clinical order or care recommendation at OHSU, Providence, Legacy or an OCHIN member clinic, any financial transaction, any outage or safety communication from a utility, and any action that sells or shares personal data in a way OCPA or HB 2008 constrains. Approvers see a structured payload with the agent's stated intent, the supporting evidence, the predicted impact and the policy citations, plus a one-click approve or reject. The log records approver identity, timestamp, context and outcome, and it is the same artifact your internal audit or the Oregon DOJ would sample.