A new kind of company just became possible

Imagine a workflow already costing a business A$30,000 a month. A small technical team delivers the accepted outcome for A$13,500 and prices the spread so both sides win. That opening exists because implementation got cheap first. It narrows as experts learn, vertical vendors package the work, and incumbents adapt.

Code got cheaper. Organisational truth did not. In one dated benchmark a frontier-model task cost US$3.26. Finding the real workflow, making the data usable, defining authority and recovery, and winning adoption and supporting it are all still scarce. That benchmark cost is only the hook; the opportunity is the remaining work.

This is a field note for people who already know the stack. Public benchmarks and papers are sourced. The timing model, example economics and strategy are mine, and illustrative inputs are not forecasts.

The feed saw the opportunity before anyone proved the operating model

Three signals now point in the same direction. None is sufficient alone.

People can feel that the unit of leverage changed.

The screenshots are cultural signals, not proof. They show that investors and operators are already looking beyond software licences.

The models are the second signal. Delegated implementation is now real enough to alter company design: capability moved from suggestion to delegation. Astra and Fable can perform long-running tool work, and Astra scores 57.9% on Terminal-Bench 4.0. Vendor benchmarks need context; the direction still matters.

The third is money. The budget around software may be much larger than the software budget. Sequoia’s services thesis puts it at $1 of software for every $6 of services. That 6× claim is a thesis, not an accounting identity. The useful question is which work dollars can become AI-native delivery.

They all point at the same gap: more code still attenuates before it becomes a release. A 2026 working paper estimates much larger gains at commits than releases (+240 at commits, +30 at releases, estimated). That points at the unautomated work between output and production.

The missing work sits between model output and a business outcome

Generating code, reading documents, using tools and drafting decisions are all cheapening fast. Five layers between that output and a business outcome are not:

1. Truth. What is the real workflow?
2. Authority. Who may decide?
3. Evidence. What proves completion?
4. Adoption. Will anyone change behaviour?
5. Ownership. Who carries failure and support?

My bet is to engineer that gap before vendors package it and experts learn it themselves.

The accepted state is the product

Monday, 8:42 am. A casual teacher starts at 9. This composite school workflow is deliberately ordinary, and that is why the economics matter.

The request arrives late as a PDF, an email, a form and a CSV. The name differs across two forms, working-with-children evidence is missing, and the requested role includes student-data access. The input is not clean enough to automate blindly: four documents disagree, the start time is close, and one permission is consequential.

1. Request, by email and forms. It arrives incomplete and inconsistent, 18 minutes before the start.
2. The agent reconciles and prepares. It builds the case, not the final authority. It matches identity where possible, checks and links sources, drafts accounts, flags missing evidence and prepares the exact approval question.
3. An approver decides the exception. A named person approves the risky part. The system does not infer privileged access from urgency; a responsible approver accepts, narrows or rejects the request, and the system records who decided what and why.
4. Ready, with correct access and evidence. Ready means more than “account created”: correct identity, least privilege, device, MFA, training, payroll state, evidence, and a tested offboarding path. Reversal is tested.

The interface is a component. The accepted state is the product.

No one wins by pretending the model replaces all three roles

The domain operator brings truth, the customer and authority. They know the exceptions, the consequences and what an acceptable outcome looks like. You bring the boundary, the system and the evidence, turning tacit workflow into interfaces, controls, recovery and an inspectable result. Frontier AI brings execution, iteration and scale: it produces, tests and operates more of the implementation loop at marginal cost.

Together they sell an accepted outcome, a result with justified reliance. Your job is to reduce total effort to an accepted result, compared with the buyer’s next-best alternative.

Price the workflow, not the software

The deal works only when the customer keeps value and delivery funds itself. Start with observed time and cost, then expose who keeps the difference.

The test is ΔVyou = V(SME + you + AI) − V(best alternative). If the operator using the same model alone gets the same accepted result at lower total cost, your advantage is not real.

The cheapest call can produce the most expensive result

Route by accepted-result cost, not token price or benchmark rank. In the Artificial Analysis snapshot of 9 September 2026, an Astra task cost US$3.26 and a Fable task US$7.63 on the same aggregate index. These numbers expire, so re-run them before deployment. The cost that matters also counts review, retries and escaped failure.

The routing rule follows: use the cheapest model that reaches acceptance after human review and failure cost.

“Build the app” is the wrong contract

A strong model will efficiently satisfy the wrong definition of done. The same staff-access job looks very different depending on where the agent is told to stop.

More conversations, fewer obligations

Parallelise discovery. Serialise responsibility. A suggested discovery funnel runs from forty operator conversations to ten evidenced workflow maps, three paid bounded pilots and one continuing venture.

Cofounder or another idea?

The partner filter scores a potential partner from 0 to 5 on domain depth, customer access, ownership of commercial work, evidence access and delivery reliability. Customer access and commercial ownership each carry 25% of the weighted fit, domain depth 20%, and evidence access and reliability 15% each. The relationship follows from the first rule that applies:

A partner rated 4, 4, 3, 4 and 3 scores 3.6: a paid pilot before founder talks.

About 50% can make sense for half the continuing founder responsibility. It is not the automatic price of version one. Define duties, vesting, governance and departure before the percentage, and move from trial to paid pilot to continuing duties, vesting and governance in that order. Discovery can run in parallel; responsibility should not, until capacity is proven.

Your evidence must compound faster than the market catches up

Models improve. Experts learn. Vertical tools package the work. Incumbents adapt.

I use a scenario model to make that race explicit; it is not a forecast. The advantage A(t) = G₀·e−kt·D(t)·E(t) starts at today’s gap, decays with market catch-up, grows with your learning and reuse, and is scaled by your delivery evidence. With a gap of 55 today, 40% catch-up a year, 20% learning and reuse a year and 60% delivery evidence, the gap has an illustrative half-life of 1.7 years and a capture index today of 33.0.

Under those inputs your delivery capability roughly tracks market catch-up, so the reading is to run a paid test: make a small, reversible investment and measure the result. If catch-up outruns your learning by 22 points or more, move and stay narrow; run a paid test, not a portfolio announcement. Delivery evidence below 40% means earning evidence before equity. If learning and reuse outpace catch-up with at least 70% evidence, you may be building an accumulating asset, and the next question is whether a second delivery is actually cheaper.

The temporary advantage is only useful if you convert it into customers, evidence, reusable systems and ownership.

One workflow, one buyer, one bill

The next evidence should come from a user, a payment or a failed acceptance test. A 30-day version:

1. Days 1–5: find the ugly workflow. Baseline time, error, delay and consequence. Output: a signed problem boundary.
2. Days 6–10: sell the diagnostic. Charge to map acceptance before software. Output: a paid decision contract.
3. Days 11–24: ship one complete slice, from real input to accepted state. Output: accepted or rejected evidence.
4. Days 25–30: count the delta. Compare SME + AI against SME + you + AI. Output: a fee, a founder talk, a revision or a stop.

Keep score on time (human hours per accepted result), quality (representative cases and critical failures), burden (review, rework and support), value (customer gain after switching) and reuse (whether the second build is actually cheaper).

I wouldn’t start ten companies. I’d start forty conversations, find ten real workflows, sell three bounded pilots and let one earn the right to become a company.

Five ways this dies

• The expert learns. Experts using frontier models alone reach the same accepted outcome at lower total cost. Then my intermediary role is weak. I should narrow into specialist engineering, own distribution, or stop pretending orchestration is scarce.
• The model absorbs it. The base model absorbs my orchestration, evaluation and recovery patterns. Then prompts and scaffolding were never the moat. I need domain data, distribution, reusable operations or a different role.
• Support wins. Delivery looks profitable until exceptions and support consume the margin. Then the service is underpriced or badly chosen. Count total burden, reprice, narrow the accepted state or stop.
• Nobody pays. Operators praise the demo but nobody pays to change the workflow. Then the pain, buyer, timing or offer is wrong. Do not convert compliments into market evidence.
• Equity traps me. Paper ownership accumulates while cash and attention disappear. Then the portfolio is an obligation portfolio. Require funding, vesting, continuing duties and explicit stop conditions.

The model bill collapsed; the company bill did not. That mismatch could let one engineer and one operator build what once needed a team. It could also disappear as experts learn and vendors package the workflow. So I’d move now, stay narrow and measure everything. Everyone has a thesis. None of them proved how.

The companion operating note is Your AI can build the app. Make it prove the outcome.