AI Writes the Code, So Why Is the Back Office Still Manual?

The Problem

Prentis is an AI lab founded in April by Ritankar Das with Reid Hoffman and Mark Pincus. It is in talks to raise $100 million at a $1 billion valuation, has signed contracts worth up to $50 million with customers including a healthcare management services organization and several manufacturers, runs on a team of about 25, and projects a $75 million annualized run rate for the third quarter. The pricing model is the part worth staring at: the company takes 20% of realized savings (TechCrunch). Its thesis fits in one line. Automating everyday office tasks will soon overtake coding as AI’s largest use case.

The thesis is probably right. The gap sits between that sentence and what happens on a Tuesday afternoon in an operations team.

OSWorld 2.0 measures computer-use agents on 108 real workflows, each taking a human a median of about 1.6 hours. The results are sobering. Under a 500-step budget, the best system was Claude Opus 4.8 with maximum thinking at 20.6% strict end-to-end completion. Partial credit reaches 54.8%, but finishing alone happens in roughly one run out of five. GPT-5.5 plateaued near 13% (arXiv). The failure modes the authors identify are not GUI control or coding ability. They are constraint tracking, handling information that surfaces mid-task, self-verification, and recovering hidden state, all of which describe how a person holds a long job together.

So two questions land on the buyer’s desk. When four runs out of five stall halfway, who cleans up the wreckage? And in “20% of realized savings,” who measures the savings?

Today both are handled in a slide. A consultant estimates minutes-per-case before rollout, multiplies by loaded labor cost, and that number goes into the contract. Nothing is measured afterward. When an agent stops mid-run, an analyst opens the screen, guesses how far it got, and takes over. No record survives of what completed and what did not, which means there is no way to audit next month’s invoice either.

Why Now

Outcome pricing is spreading fast. Billing on savings is not a Prentis quirk; it is what a market does when seat-based pricing cannot clear a CFO’s approval. That structure only holds if someone measures. A vendor computing its own savings and sending its own invoice is a temporary arrangement. Finance asks the obvious question at the second renewal, and the RPA generation already taught enterprises to distrust unaudited automation ROI claims after years of UiPath-era bot counts that never showed up in headcount.

The technical preconditions are in place. Agents already emit screenshots, click coordinates, DOM access and tool calls as logs. Systems of record expose case status through APIs. What is missing is not data but the layer that reconciles the two streams and rules on whether a case actually closed.

Hiring signals point the same direction. Scan public job boards at AI infrastructure companies and the roles growing fastest are solutions architects and security engineers rather than model researchers. The bottleneck has moved from model quality to deployment and auditability.

There is also a clock. Pilots signed in the first half of this year hit their renewal conversations two or three quarters out. The tool that produces the evidence for those meetings has to exist before them.

How to Build It

flowchart LR
    A[Human screen logs] --> C[Step map]
    B[Agent execution logs] --> C
    C --> D[Per-step pass checks]
    D -->|pass| E[Closure record]
    D -->|fail| F[Escalate with state and evidence]
    E --> G[Monthly measured-savings report]
    F --> G

Keep the MVP narrow: one workflow family, one customer. Claims intake, customs duty refunds, invoice reconciliation, anything that takes a person 30 to 90 minutes and repeats. Start by instrumenting the human version, splitting it into steps and timing each one. Then overlay agent execution logs on the same step map. Attach a pass condition to every step, and insist those conditions are checkable against a system rather than judged by a model: record created, amount matches the ledger, document filed. When a step fails its check, bundle the state and the evidence and push it to a human queue. At month end, multiply closed cases by measured time saved and hand the customer an invoice they can verify.

The discipline is to not build the agent. Prentis, Anthropic, OpenAI and Thinking Machines are all building those, and that race needs nine figures. A layer that scores any vendor’s agent the same way and hands off failures the same way gets more valuable the more neutral it stays. The stack is unremarkable: event log collection, a step map editor, a small DSL for pass conditions, one human queue. The hard part is authoring the step maps, and that is also the moat, because each new customer deepens the library.

Success Criteria

The first assumption is that buyers actually want savings verified. Ask finance leads at five companies already on outcome-based automation contracts what they used to check last quarter’s invoice. Two weeks is enough. If three of them say they forwarded the vendor’s own report, the market is real.

The second assumption is that agent vendors tolerate a scorekeeper. Some will not, which is why this sells to the buyer, not the vendor. The twist is that a vendor billing on outcomes eventually needs the same evidence to defend a renewal, so the ones pushing hardest against it early are plausible inbound leads a year later. How fast that turn happens decides the pace of the business.