NowAI does
the work

For decades, software has helped companies track, route and store their work. A claims team could see every claim and a finance team could see every invoice, but a person still read the file, applied the rules and decided what happened next. The software was the cabinet. The clerk did the work.

That boundary is moving. A model can read an email, pull the facts out of a PDF and draft the reply, which lets a system finish work that used to wait for a person. The opportunity is a company that sells the finished file. AI does most of the labor, and people stay close to the parts where judgment and accountability still matter.

Most of the discussion about AI applications stays inside software budgets, which are a relatively small pool of spend. The larger pool sits in the work itself: processing claims, reconciling payments, chasing renewals, preparing compliance files and answering customer requests.

Take an insurance managed general agent with $100M in revenue. It might spend $1M to $2M a year on software, which makes it hard for a software company to build a venture sized business there. Its operating costs tell a different story. Intake, quotes, renewals, customer emails and policy servicing add up to a large share of the cost base. A company that sells the finished workflow sells into that budget.

Over the past decade, companies grew comfortable with distributed teams, offshore operations and outside operating partners. Many functions already run through a patchwork of internal operators, vendors, workflow tools and manual handoffs. The habit of handing work out is already there.

The software shelf is crowded too. A mid size insurer or lender usually owns a core system, a CRM, a ticketing tool, a document store and several workflow products. The bottleneck is the burden of coordinating across all of it. These buyers want lower cost, higher accuracy and faster turnaround, and another application adds to the pile.

A service wrapper fits that reality. The vendor starts in the channels that already exist: the inbox, the portal, the PDF and the spreadsheet. Nobody runs a software implementation or redesigns an operating model first.

That learning curve is the part worth watching. Five measures show whether it is real. Revenue per operator rises. Exception rates fall. Turnaround time comes down. Gross margin improves with volume. New customers in the same workflow onboard faster, because the system has already learned the common paths.

The category is often described as BPO with AI. A traditional BPO takes an existing process and moves the labor to a cheaper or more scalable place. Technology helps, but labor remains the main unit of scale.

An AI native service can look service heavy at the start, because production quality still needs people. Early operators understand the workflow, catch failures and turn each customer's messiness into repeatable steps. Over time that knowledge moves into software, review systems, integrations and evaluation data.

The test is the shape of the line. If the five measures stay flat, the company is a better services firm. That can be a good business, and it is a different one from the business a venture investor is underwriting.

Not every workflow suits this model. The best candidates are frequent, painful, bounded and valuable enough to absorb. The quality bar is clear, the steps follow stable rules, and a person can stay close to the parts that need judgment. Open ended or rare work is better left alone until the models improve or the volume justifies the build.

The obvious places to start are existing services categories: call centers, BPOs, offshore operations teams and back office vendors. The labor is visible and the processes are already outsourced. These markets also carry price benchmarks, set by delivery models that have been tuned for decades, so a new entrant's margin is thinner than it looks.

The more interesting opportunities show up before a formal vendor category exists. Spend a day inside one of these:

A surprising amount of work is repetitive, expensive and still trapped inside the company, because it never cleanly fit into software or outsourcing. The workflow has structure, but the path changes by customer, geography, document type and exception. Nobody could serve it at attractive economics before. It shows up as headcount, backlogs, error rates, slow response times and missed revenue.

Take an MGA, where submissions arrive by email with attachments in every shape. An AI native service can take intake and underwriting preparation as one completed workflow. It receives the submission, extracts the relevant information, checks the file for completeness, enriches it, flags issues and prepares the underwriter's view. Exceptions route to a person, and the system keeps what it learns.

The MGA can pay per file, per policy or per completed workflow. It keeps its core system, manages no additional team and gets faster cycle times and a lower operating load. Pibit runs underwriting workflows for MGAs this way today. The same pattern shows up in healthcare administration, revenue cycle operations, accounting, compliance, logistics, legal operations and customer support.

Pricing follows the unit of finished work. A vendor can charge per transaction, per file, per policy or per completed workflow, and the strongest companies price against outcomes. The buyer pays for work that is done, which keeps the vendor pointed at speed and accuracy and away from hours billed.

This kind of company needs software talent and operating discipline in the same team. The SaaS wave showed that Indian teams can build product companies for global buyers. The services wave before it built deep capability in process design, offshore delivery, quality management and training. AI native services sit where those two histories meet.

The best early customers are small and mid size companies in the US and other developed markets, with real budgets, real backlogs and little appetite for a software transformation. Lower delivery cost helps from day one. The lasting advantage is a team that combines model engineering with disciplined operations and treats delivery as the product.

At ShopOS the work runs in a loop. Research benchmarks a model and turns the result into a reusable workflow. The workflow runs as a deployed brand job. A client or a reviewer makes a decision, the result goes live with its cost trace, and the lesson lands in memory, evaluations, a skill or a template. The next comparable job starts from there.

The loop earns its claim only when the next comparable deployment improves on setup, acceptance, review effort, reruns or reliability. A large archive of prompts and outputs shows activity. The curve shows compounding.

ShopOS sells brand outcomes. A brand asks for a catalog, a campaign or a store page and receives the finished work. A pod of three to five people runs each brand: a creative director, an AI expert, a workflow engineer, QA and last mile, and an account lead.

Every engagement leaves something reusable behind, such as a cleaner input contract, a category template, a rejection rule or a test. We compare the first deployment with the nth one for similar work, and the nth should need less setup, fewer reruns and less review. Today that is a target. Setup time, reruns, human hours per outcome and gross margin need two quarters of tracking before it counts as proof.

Three patterns keep a deployment business stuck as a services business. The first is manual heroics, where the client is happy because one operator remembers everything. The second is a prompt archive, where the team saves text with no stable input, output, test or owner. The third is a flat cost curve, where each new account rebuilds the same labor under a new brand name.

If one of these shows up, the business has revenue and client knowledge and no software yet. Capacity is the other limit, because trained creative people set the pace and hiring takes time. Everything here can look different in six months, so we hold the model loosely and keep the interface simple: a customer can bring in an expert for one task and let them go afterwards.

Teams building in this category run into the same six design choices. Which workflows are frequent, painful, bounded and valuable enough to absorb. Whether to go horizontal across industries or deep into one vertical. How to price: per transaction, per file, per workflow or per outcome. Where the line sits between AI, human review and exception handling. How gross margin should improve as volume grows. And whether to build the operating layer from scratch or buy an existing services business and rebuild it around AI.