From Research to Production: What the Open Roles Show
Truss, founded in 2022 with 11–50 employees, operates a platform that spans the full tax engagement lifecycle: intake, document collection, prep, engagement letters, e-signatures, payments, workpapers, and final delivery. It sits alongside CCH Axcess, UltraTax, Lacerte, ProConnect, ProSystem fx, GoSystem Tax RS, Drake, and ProSeries rather than replacing them. That integration mandate alone changes the engineering profile: you cannot ship a wrapper around someone else's compliance engine without treating reliability as a feature, not an afterthought.
The product's own feature set makes the priority plain: Smart Rename and Smart Split handle document classification at scale; AI Prep and One-Click Prep push returns to review-ready status inside the firm's existing software; Cost & Surge Control gives partners visibility into token spend. These are not research problems. They are systems problems with an LLM in the loop.
Truss describes itself as the first agentic AI tax workflow platform built for modern CPA firms. The word "agentic" carries weight: it implies orchestration, memory, tool use, and error recovery across a multi-step workflow that spans weeks, not a single chat turn. Building that requires engineers who have debugged distributed systems, designed idempotent pipelines, and instrumented observability so that when the model drifts, the team knows before the client does. The company hosts workshops on the historical arc of AI in tax and fields CEO-level conversations about where AI fits, and where it doesn't. That market-facing posture signals a team that sells outcomes, not models.
Customer testimonials on the Truss site and LinkedIn page reinforce the shift. Firm owners cite eliminated back-and-forth emails, automated routine requests, and Mondays that "suck less." The language is operational, not experimental. Partners include Aprio, BDO Alliance, CPAmerica, and the Information Technology Alliance, organizations that buy reliability, not potential. When the IRS issued guidance on AI in tax preparation in late June, Truss responded with a workshop tracing the shift from historical roots to current deployment. The timing suggests a company moving in step with regulatory clarity, not ahead of it.
The hiring signal is consistent: Truss needs builders who understand that tax software fails in specific, expensive ways: missed deadlines, incorrect carryforwards, broken audit trails, and that an LLM introduces new failure modes without removing the old ones.
Tax Fluency Over ML Credentials
Truss's career page emphasizes shipping over process. Candidates who lead with model-tuning publications or prompt-engineering portfolios stall. Candidates who can explain how a 1040 Schedule C flows into a workpaper, why IRS e-file rejection codes matter, or what SOC 2 Type II evidence looks like for a document-collection pipeline move forward.
The product architecture demands this fluency. Truss sits between clients and the firm's tax software, including CCH Axcess, GoSystem Tax, Ultra Tax, Lacerte, ProSeries, ProConnect, CCH ProSystem fx, and Drake, organizing intake, workpaper prep, Q&A, and delivery. It does not prepare returns. It collects documents, structures data, and automates communication. That means every AI-assisted feature, such as document classification, summary generation, and drafted email replies, must preserve the evidentiary chain a CPA signs off on.
Audit readiness is not a compliance checkbox; it's the product's core constraint. Firms using Truss report cutting back-and-forth emails and moving 95 percent of client interaction online. That speed only holds if the AI output is traceable. The engineering team builds observability into every LLM call: structured logs, confidence thresholds, human-in-the-loop gates, and exportable audit packets.
Regulatory alignment compounds the difficulty. The platform handles PII, financial records, and communications that fall under GDPR, state privacy laws, and IRS safeguards. Engineers at Truss write data-flow diagrams before they write model cards. They version-control prompt templates the way other teams version-control model weights.
The hiring signal holds across roles. The four open positions, spanning backend, applied AI, and product engineering, all list tax workflow familiarity or regulated-industry experience as required or strongly preferred. Pure research backgrounds are not disqualified, but they are insufficient. The company's founders built the platform after running tax practices themselves. They know the busy-season cadence: intake floods in January, workpaper review peaks in March, delivery compresses into April. An engineer who has never seen a 1040 organizer cannot anticipate the edge cases that break automation at scale.
This does not mean Truss ignores ML competence. The applied AI role requires production LLM experience: RAG pipelines, evaluation frameworks, latency budgets. But the bar is applied competence, not novel architecture. The team uses off-the-shelf models, fine-tuned on firm-specific document corpuses, wrapped in deterministic control logic. The differentiation lives in the control logic: the routing rules that escalate a flagged W-2 discrepancy to a senior preparer, the validation layer that cross-checks extracted figures against prior-year carryforwards, the communication engine that drafts a client follow-up in the firm's voice and tone.
Candidates who understand that distinction (who can articulate why a low hallucination rate on Schedule E rental income is unacceptable while a higher rate on document categorization is manageable) pass the screen. Candidates who optimize for benchmark leaderboards do not. The market Truss serves, small-to-mid CPA firms drowning in manual intake, buys trust, not benchmarks. The hiring process filters for engineers who build that trust into the system.
The Engineering Bar: Reliability, Observability, Regulatory Alignment
Truss doesn't hire engineers to optimize model weights. It hires them to keep a tax workflow platform from becoming a liability. The company's own site leads with "SOC 2 Type II controls plus continuous hardening" and tells firms to "point auditors to OWASP-aligned, SOC 2 Type II reports and control mappings." That language isn't marketing — it's the contract Truss signs with every CPA firm that puts client data on its rails. Candidates who can't speak to those controls don't clear the screen.
The observability bar reflects what the Big Four have already learned at scale. Deloitte's Zora, EY.ai, KPMG and PwC's agentic platforms all run air-gapped or in private cloud with no internet egress except to whitelisted APIs. Encryption is AES-256 at rest, TLS 1.3 in transit. Agents operate under RBAC tied to the human auditor's credentials. Prompt injection detectors watch for data exfiltration attempts. Deloitte uses a "sandbox and verify" pattern where agents propose actions, a sandbox executes them, and results are validated before touching production. All four firms employ red teams dedicated to agent vulnerabilities. Truss operates in the same regulatory atmosphere: its AI-assisted workpaper prep, document extraction, and client communication features touch the same sensitive financial data. The observability stack has to catch drift before it becomes a restatement.
Modern data pipeline observability breaks into three dimensions that map directly to Truss's workflow: performance monitoring that surfaces bottlenecks in document ingestion and transformation; data quality monitoring that validates freshness, completeness, and consistency across client uploads, extracted fields, and generated workpapers; and lineage tracking that shows how a figure flows from source document through AI extraction to the final return. The fragmentation problem is real: firms typically run multiple tools across ingestion, transformation, orchestration, and consumption. When an error spans systems, the root cause often sits upstream from where symptoms appear. Truss's platform sits in the middle of that fragmentation: it collects client documents, runs AI extraction, manages Q&A, and tracks workflow status.
Error handling in financial data pipelines carries a different weight than in consumer AI. The PCAOB is already updating standards to require firms to document algorithms used for risk scoring and maintain version control of agent models. Automated workpapers must carry source citations that form an immutable chain of evidence. When Truss's AI extracts a figure from a client's uploaded brokerage statement and places it on a workpaper, the lineage has to survive an audit. That means idempotent processing, replayable transformations, and audit logs that satisfy both SOC 2 and IRS e-file requirements. GDPR adds the right to erasure on top: a deletion request has to propagate through every derived artifact without breaking the audit trail.
Regulatory alignment isn't a checklist item at Truss — it's the product architecture. The platform advertises white-label client portals, secure document handling, and audit trails as core features. That means the engineering team builds compliance into the data model, not the perimeter. Encryption keys, access control policies, and retention schedules live in the same version-controlled infrastructure as the application code. Candidates who have shipped SOC 2 Type II evidence packages, mapped OWASP controls to implementation, or navigated IRS e-file requirements signal they've done this work. Pure ML researchers rarely have. The hiring signal is clear: Truss needs builders who have operated in regulated environments where a data quality incident triggers a regulatory finding, not just a PagerDuty alert.
The market validates this bar. Deloitte cut first-year hiring by 30% and created "agent supervisor" roles for experienced staff who oversee AI output. EY retrained 10,000 tax professionals to validate agent conclusions rather than calculate manually. Junior staff now review agent-generated workpapers in month one instead of building them from scratch in year three. The trade-off is real: partners worry the "spidey sense" for fraud atrophies without granular reconciliation experience. Truss's engineering team has to build tooling that preserves that judgment layer: observability surfaces that let reviewers trace AI decisions to source evidence, anomaly detection that flags the exceptions human reviewers should catch, and audit trails that withstand regulatory scrutiny.
Why AI in Tax Workflow Is Scaling Now
The CPA profession is in the middle of a workforce crunch that no amount of overtime can solve. The American Institute of CPAs reports that three-quarters of firms have struggled to hire qualified staff, and employment in accounting and auditing is projected to grow just 4% from 2022 to 2032, barely keeping pace with the overall labor market. Meanwhile, firms posted an average 24% revenue increase over the past 12 months. More work, fewer hands. The math doesn't work.
This pressure shows up in priorities. Thomson Reuters' 2024 State of Tax Professionals Report found that efficiency and automation topped the list for the second year running, followed by talent retention and hiring. Two-thirds of respondents said they'll streamline processes over the next two years, and nearly half are eyeing new technology solutions, up from 41% in 2023. Midsize firms are already automating 26–50% of their workflows at a rate of 25%, with another 20% pushing past the halfway mark. Large firms trail only slightly at 27% and 15% respectively.
Yet AI adoption remains thin. Only one in four firms uses AI at all, and most of that is for research. Just 7% call it a top investment priority. But 35% say they'll invest in some form of AI over the next two years. That gap — low current use, high stated intent — is where companies like Truss operate. The market isn't asking for research prototypes. It's asking for tools that plug into existing workflows, handle regulated data, and survive audit scrutiny.
The pricing shift sharpens the demand. Value-based billing (charging for expertise and results rather than hours) jumped to the No. 3 priority in 2024, with 48% of firms already offering it on some services and another 18% planning to by year-end. When revenue decouples from time spent, every manual hour becomes a margin leak. Nearly half of firms said they'd increase automation if they can't fill roles. Another third would turn to freelancers, 29% to outsourcing. The labor shortage has a technical escape hatch, and firms are preparing to use it.
Regulatory complexity compounds the problem. Firms cite licensing limits, uncertainty around advisory boundaries, and the sheer time cost of compliance as reasons they can't expand services — even as 74% already offer tax strategy and 66% report strong client demand for business advice. The opportunity is there. The capacity isn't.
Truss's four open roles — focused on production engineering, reliability, and tax-domain fluency — map directly to this moment. The buyers aren't looking for model architects. They're looking for systems that reduce manual labor without creating new compliance risk. The 57% of firms planning growth through technology and automation need vendors who understand SOC 2, IRS e-file standards, and the difference between a hallucination and a liability. That's a builder's market. Truss is hiring for it.
The IRS guidance workshop in June 2026 was a working session for firms already using Truss to map the new rules onto their existing workflows. The engineers who built that mapping — prompt templates version-controlled, validation rules codified, audit packets auto-generated, are the same ones now screening the next four hires. They know exactly what the screen tests for because they built the system that passes it.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.