Legal AI infrastructure for firms
AI RFP discovery and response drafting
Automatic.coBusiness process automation
Secure AI virtual data rooms
Replacing Manual Back Office Work With AI Agents
Klarna's OpenAI-powered assistant handled 2.3 million customer conversations in its first month, equal to the output of 700 full-time agents, with average resolution falling from 11 minutes to under two. That figure defines the question here: which back office queues clear that economic bar, and which do not.
The math reduces to four variables per workflow: loaded labor cost, exception rate, agent run cost, and payback month. Four functions sit at the center of most buyer conversations right now: accounts payable, vendor onboarding, order-to-cash reconciliation, and HR ticket triage. The answer is different for each.
The Variables That Decide Which Queue to Agent First
Loaded labor cost is base salary multiplied by a burden factor of 1.25 to 1.4, divided by effective working hours. Exception rate is the share of items the agent will hand back to a human, and it drives almost everything downstream. Agent run cost is the per-invocation spend on model calls, tool calls, and orchestration overhead, which for structured back office work typically lands between $0.05 and $0.40 per completion. Payback month is annual labor displaced minus annual run cost minus build cost, divided across months.
The build cost itself is the lever buyers underestimate. Our estimates on a custom agent build run $80k to $250k depending on system reach, and anything touching ERP or HRIS lands in the upper half of that range. McKinsey separately puts banking's generative AI opportunity at $200–$340 billion annually, with roughly 75% concentrated in customer operations, software engineering, and knowledge work. Back office is in that 75%.
Accounts Payable Is the Cleanest Win
A loaded AP clerk at $65k base runs about $84k fully burdened, or roughly $42 per productive hour. A seasoned clerk processes 1,800 to 2,200 invoices per month, averaging three to five minutes per invoice when the data is clean. Agent run cost on a well-structured PO invoice sits near $0.10 using current frontier models. Exception rates land around 15–20% for three-way-match PO invoices and climb to 40% for non-PO expense invoices.
Payback for a mid-market AP queue running 15,000 to 25,000 invoices per month lands at four to six months once the agent is live, including integration to the ERP's invoice table. The build versus buy math tilts toward buy for pure PO matching and toward build when the invoice mix is non-PO heavy, because exception routing becomes the real product.

Vendor Onboarding Pays Back Slower Than You Think
Vendor onboarding looks structured from the outside: W-9, banking details, insurance certificate, OFAC and sanctions screen, diversity attestation. In practice the exception rate runs 35–50%. Tax IDs mismatch, insurance expiration dates fall inside the window, banking changes require out-of-band verification to avoid supplier impersonation fraud. The human in the loop is a senior AP or procurement analyst at $95k to $120k loaded, which raises the ceiling on savings but also raises the stakes on agent errors.
Agent run cost on a multi-step onboarding flow lands closer to $1.20 per vendor because the agent is calling four or five tools and reading documents. The real build cost sits in approvals plumbing, which is why guardrails for approvals, audits, and role-based access dominate the sprint plan. Expected payback: nine to fourteen months for a mid-market program onboarding 400–800 vendors annually.
Order-to-Cash Reconciliation Has a Narrow Fit
Cash application is the queue where buyers most often overestimate agent fit. When remittance advice arrives with the payment, in the ACH CTX envelope or an EDI 820, matching is mechanical and agent run cost can stay under $0.15 per payment. When a buyer sends a wire with no remittance, or a check against three invoices with a short-pay on one, the agent is reading a PDF and emailing the buyer for clarification, which is slow and error-prone.
Loaded cash application analyst cost sits near $72k. For a shared services team processing 20,000 receipts per month with a clean remittance rate above 70%, payback is five to seven months. Below 50% clean remittance, the economics invert and the correct move is to fix inbound data first, often through an integration between the CRM and the ERP so the invoice carries enough metadata for a deterministic match before the agent sees it.
HR Ticket Triage Looks Like AP but Isn't
IBM's AskHR handled more than 16 million messages in 2025, roughly 39,000 interactions per day, and became the single entry point for 85% of upline managers. That is the headline number back office buyers want to copy, and it is also the one that most misleads them. AskHR sits on top of years of structured policy data and a Workday instance IBM controls end-to-end.
A typical enterprise rolling out HR ticket triage faces a long-tail question distribution, policies that vary by country or union, and hard RBAC requirements for anything touching compensation or benefits. Loaded HR generalist cost is $75k to $90k. Agent run cost per ticket is $0.20 to $0.35 when the knowledge base is well-structured, and rises sharply when the agent has to ground responses in scanned PDFs. Payback is six to nine months when the knowledge base is already in shape, and twelve to eighteen months when the first project is actually a knowledge management project in disguise. The architectural pattern, especially for regulated data, usually ends up being an agent fronting a private retrieval system over internal sources, which is what our AI development practice ships most often now.
What the Scaled Programs Have in Common
The World Economic Forum projects data entry clerks, bank tellers, and administrative assistants among the fastest-declining occupations through 2030, with 92 million roles displaced globally against 170 million created. The programs actually posting those numbers share three traits: structured inputs with stable schemas, exception rates under 25%, and enough monthly volume that even a 60% automation lift pays off integration work. They also share a design principle about blast radius, since any agent writing to a financial ledger or HRIS needs the kind of controls described in our notes on agents that touch ERP and CRM.
The queues that fail the test share the opposite pattern: unstructured inputs, exception rates above 40%, and workflows where the human spends most of their time chasing data rather than making decisions. Agenting those queues just moves the chasing from a person to a model, at higher cost per attempt.
Where Agents Earn Their Place
Agent back office work first if the queue has structured inputs with a stable schema, an exception rate below 25%, and at least 1,000 items per month. Agent it second, after a data-cleanup project, if the exception rate is 25–40% and the data quality problem is bounded and fixable. Leave it manual, or outsource it, if the exception rate is above 40%, the inputs are unstructured, or the policy surface changes faster than the knowledge base can be maintained. The payback month, not the demo, decides which bucket a workflow belongs in.
