#ai-workflow

HR lives in repeatable, deadline-bound lifecycle work: onboarding, offboarding, role changes, policy campaigns, leave and accommodation intake. That is where AI task-generation and document-collection do their best work. The hard line is the hire, fire, and accommodation decision, which stays human, because the EEOC says anti-discrimination law applies to AI hiring tools just as it does to any other employment practice.

Engineering and GTM-ops teams can now wire AI straight into their own workflows: implementation, security questionnaires, incident response, change management, access reviews. The trap is vibe-coded automation with no change-control, which fails the first SOC 2 audit. AI drafts and assembles; a human still approves every production change and control sign-off.

Accounting is already process-shaped: PBC lists, close checklists, review sign-offs. That is exactly why AI helps when it runs inside the process and hurts when it freelances. With the AICPA reporting accounting graduates down 6.6% in 2023-24, the rote work has to be carried by a tighter process rather than more hires. The model extracts and drafts; the preparer and partner own the numbers.

Terse instructions hand every unstated decision to whoever reads them, and an AI agent fills the gaps with guesses. A May 2026 Hacker News thread on reliable prompts converged on one pattern: give context, name the output, list what to avoid, set the tone. Process steps need the same four.

Insurance runs on document-heavy intake and adjudication: underwriting submissions, FNOL, claims handoffs, renewals. That is where AI extraction and triage pay off. But the coverage call and the denial stay human, because the NAIC unfair-claims rules require a reasonable, accurate explanation for any denied claim, and an unowned automated denial cannot give one.

Agencies and consultancies bleed margin on the unbilled coordination between the billable work: onboarding, kickoff, status updates, QA, renewals. AI can draft most of it, but only if there is a defined process for it to slot into. Most firms do not come from a rival tool. They come from spreadsheets, email threads, and one senior person who remembers how it all works.

After lawyers got sanctioned for filing fake AI-generated cases, the lesson was process, not a better model. In Mata v. Avianca a judge fined the attorneys $5,000 for citing non-existent opinions. The firms getting value from legal AI are the ones with a disciplined intake-to-closing process the model plugs into, with a human review gate the work cannot skip.

Companies blame the AI when an agent stalls on their workflow. The real problem is older: the decision logic lives in a few senior heads, not in any document. On Hacker News, builders kept hunting for real examples of agents doing work and mostly found rebranded automation. Document the steps that really mean ask Bob first.

The administrative load is where AI pays off in healthcare, well away from the clinic floor. An AMA survey found 94% of physicians say prior authorization delays care and the average physician handles 43 of them a week. AI can extract, assemble, and draft across that load, while a person with the right role owns every step that touches protected health information.

LangGraph, CrewAI, and AutoGen are developer SDKs for building AI agents, not tools your operations team will ever open. Microsoft moved AutoGen to maintenance mode, which says plenty about framework churn. Here is what each one does, who it serves, and where a business workflow platform fits.

Banks do not lack AI demos. They lack the audit-trailed, multi-approver process to put a model anywhere near a regulated decision. AI fits as a classify-and-draft layer feeding a named human owner, never the decision-maker. The CFPB has already said a black-box model cannot dodge the duty to give specific reasons for a credit denial.

Most SOPs assume the reader already knows the unwritten context. An AI agent does not. AWS shipped Strands Agent SOPs in November 2025 using RFC 2119 keywords like MUST and SHOULD because human-style instructions confuse machines. Here are ten rules for writing a process an agent can actually run.

Companies keep launching a company AI agent that is supposed to do everything, and it ends up mediocre at all of it. A March 2025 Hacker News thread asked for less capability and more reliability. The fix is a different unit: deploy AI inside one named workflow at a time.

Generating a working app from one sentence is easy now. Keeping it useful inside a real business is not. An a16z partner put it bluntly: what AI generates is not enterprise software. A vibe-coded tool with no defined process drifts the moment reality leaves the happy path.