One category,
two growth engines

Start with the distinction most growth analysis skips. The mechanism producing this quarter's number is often not the asset producing the company's durable advantage. They run on different clocks, they need different budgets, and reading the revenue line tells you almost nothing about the state of the slow one.

Which means the phrase growth engine is doing too much work. An engine is a mechanism where each completed cycle makes the next one cheaper. Fuel is anything you buy to turn the wheel until the engine catches: a sales team, an ad budget, a Series A. Both produce growth. Only one of them keeps producing it after you stop paying.

An engine makes each cycle cheaper than the last. Fuel just turns the wheel until the engine catches.

There is a Capital loop in the taxonomy, and money from investors is not it. A Capital loop means margin from one cohort financing acquisition of the next at improving economics. Nearly every funded company clears the low bar of having capital. Almost none clears that one, and conflating the two is how a funding announcement gets read as a compounding advantage.

The five loop types, briefly

They classify by what compounds and what fuel turns the wheel, rather than by acquisition channel. That is why Trust and Skill have no counterpart in the usual taxonomies, and why those two carry most of B2B.

Harvey and Sandstone make the difference legible, because they sit in one category, sell to adjacent buyers, and are compounding in completely different places.

Trust loop · compounds over years

What compounds

Accepted proof. One consequential matter completed, reviewed, and accepted inside a firm that peers watch.

What the customer keeps

Confidence under scrutiny.

What it closed to get there

Cross-account learning, by contract.

What is producing the number

$1.5B raised, deployed into implementation.

Content cycle · compounds continuously

What compounds

Relationship context. One resolved request with its rationale.

What the customer keeps

Faster, clearer guidance.

What it closed to get there

User mastery, by design.

What is producing the number

$40M raised. No loop yet demonstrated.

What both of them closed, and what they did not

The obvious answer to what compounds in legal AI is client work. Every matter teaches the system something, and that knowledge pools into a product that improves for everyone.

Neither company does it, and both closed the door themselves.

Harvey does not use customer data or content to train models or improve its products. Its model providers are contractually barred from training on it and required to retain nothing. Data is separated per customer. Customers own their inputs and outputs and can export or delete. Sandstone's security page says the same in fewer words: no training on customer data, model providers under zero data retention, you set retention, you can export.

Two companies at wildly different stages arriving independently at identical commitments, because in a category where legal, security, privacy, and risk teams examine the data boundary before permitting deployment, those commitments are not optional constraints. They are the entry ticket.

Harvey's strongest visible proof network begins with major law firms, while Sandstone is oriented more directly toward in-house legal operations. They occupy the same broad legal-AI category, but not the same buying system, which makes the contrast useful rather than perfectly symmetrical.

Harvey compounds proof

What accumulates at Harvey is not a data asset. It is an accumulating record of high-consequence legal work performed, reviewed, and accepted, concentrated in the firms whose adoption other firms treat as information. Eighty percent of the top hundred law firms, by the company's own count, and more than three thousand organisations, up from roughly thirteen hundred earlier this year.

That is a formidable stock, and it is also close to a ceiling worth naming. A Trust loop runs on deposits that peers can see, and in the population where that visibility is highest there are twenty firms left. The next increment of proof has to land in in-house teams, international markets and professional services, where the reference network is looser and a managing partner's word carries less. The stock keeps growing. What its level predicts starts to change.

That is a Trust loop, and proof is what it looks like in this category. The trigger is boring reliability, repeated inside a review hierarchy that catches errors cheaply. The loop is fueled by repeated demonstrations of competence. What the customer keeps is not speed but the confidence that what comes out will hold up.

Why that loop and not another. A partner is not buying productivity alone. They are buying confidence that faster work will still withstand partner review, client scrutiny, and professional obligations. The private question is not merely whether the tool is faster. It is whether its work will hold up, and who will have to explain it if it does not. Only proof answers that, and proof compounds one way: uneventfully, repeatedly, over years.

And here is the trade. The no-training guarantee is what makes the proof deposit possible. It is also the commitment that closes the obvious route to a data asset. Harvey bought its Trust loop with its Content loop, and it is not a reversible decision. A firm that spent two weeks satisfying itself that its client files are not training anyone's model does not un-ask the question.

Anyone who stops the analysis there, as I did, will miss what happened next.

Then Harvey built the other loop anyway

On 20 August, Harvey published Tenet as a research preview: a model post-trained with Fireworks from the open-weight Kimi K3 base across roughly 1,750 agentic legal task environments. The training corpus was synthetic data, publicly available legal data, and work produced by human experts. Harvey states no customer data was used. It is a preview rather than a shipped product, and that distinction matters for how much weight the next few paragraphs can carry.

Three weeks later it raised $550 million at a $15.5 billion valuation, past $1.5 billion in total funding, on more than $400 million of annual recurring revenue. The company frames the round as helping legal teams own their intelligence. Bloomberg reported the capital is earmarked for building proprietary models.

Read those together and the move is plain. The constraint was never that Harvey cannot build a compounding intelligence asset. The constraint was that it cannot build one out of privileged client work. So it manufactured the training material instead, and no client agreement governs material a company makes itself.

The constraint was never that Harvey cannot build a compounding asset. It was that it cannot build one out of privileged client work.

The benchmark is the more interesting half. Harvey open-sourced LAB, its Legal Agent Benchmark, on 6 May: more than twelve hundred agent tasks across twenty-four practice areas, graded against more than seventy-five thousand expert-written rubric criteria, on an all-pass basis where a task counts only if every criterion passes. Harvey reports that outside researchers now use it.

Publishing the measuring stick is not a data play. It is the most aggressive proof play available. If the category comes to measure legal agents on Harvey's ruler, Harvey's definition of good legal work becomes the target everyone optimises toward, and every competitor's progress gets scored in Harvey's units. That is a Trust loop extended from reference selling into standard-setting, which is a considerably larger asset than a customer list.

Two things keep that from being a victory lap.

The first is arithmetic. Seventy-five thousand expert-written rubric criteria are expensive. If Harvey funds every future one, output stays proportional to input, and an asset whose growth is bought one unit at a time is a well-financed stockpile rather than a loop. Whether anyone outside Harvey has contributed tasks or criteria, rather than merely run against them, is the fact that decides it, and it is not public.

The second is that Harvey wrote the benchmark and then reported its own model's score on that benchmark's hold-out set. An independent July evaluation from Artificial Analysis produced a useful caution. On its related task set, the base Kimi K3 model ranked highest. The comparison is not like-for-like and does not refute Harvey's result. But it illustrates the risk of a company defining a benchmark and then reporting its own performance against it, because a Trust loop reprices on the worst case rather than the average, and a self-graded claim that gets challenged in public is the classic worst case.

The strongest counter-evidence is Harvey's own: Tenet's gains transferred to Mercor's APEX Agents and Crosby's Redline Bench, neither seen during training, one run in a different harness entirely. Transfer to somebody else's test is the part that is hard to fake.

Sandstone has traction. Whether it has a loop is unverified

Sandstone sells context that compounds. Requests arrive from Slack, Outlook, Jira, ServiceNow, Asana and Ironclad across more than fifty integrations, and the platform catches them, classifies intent, gathers counterparty history and prior decisions from across the business, routes, and applies the playbook. In its own words, playbooks learn with each use so positions stay consistent across the organisation.

The traction is real and should not be waved away. Wayfair, Grindr, Mercury, MasterClass, Cox Media, ElevenLabs, Hypertherm. Ten million from Sequoia in January, thirty million from Lightspeed in June. Revenue growth of more than 40x, which is the company's own figure, reported as ninety days in its Series A announcement and as a hundred and forty-seven days in a vendor case study. Same multiple, two denominators, which is a reason to use it as evidence of momentum rather than as a rate.

What none of that establishes is a self-reinforcing mechanism. And there is a structural reason to think the celebrated one is not it.

The context graph accumulates inside a single tenant, and by Sandstone's own commitments the customer sets retention and can export. That is correct for the buyer and it means the stock improves that account without lowering the cost of winning the next one. An asset that accumulates cleanly, improves an existing customer's next cycle, and does nothing for acquisition is a stockpile. Stockpiles are worth having. They buy retention. They are not the same thing as an asset whose growth feeds its own inflow, and founders routinely invest as though only the first kind exists.

A stockpile buys retention. It is not the same thing as an asset whose growth feeds its own inflow.

So what is visibly turning the wheel is forty million dollars of outside capital, deployed into product, implementation and a demo-gated sales motion with no published pricing. That can produce extraordinary growth. It is fuel, not an engine, and the distinction matters because fuel stops when you stop buying it.

The engine Sandstone is actually building shows up elsewhere, and mostly not in the marketing. Twenty-plus general counsel invested in the seed round. The product's central credential is that GCs built it. SOC 2 Type 2 and a security review pack published so a buyer can forward it without asking. In-house legal is small, referential, and unusually mobile between employers.

That is a Trust loop under construction, and it comes with a complication worth naming. Proof offered by people holding equity is not the same instrument as proof offered by people who have nothing to gain. It is not disqualifying, and it is the reason the loop needs a second proof surface built from customers who are not investors.

One more thing Sandstone closed, deliberately. Fifty integrations exist so the rest of the business does not change how it works, and the agent layer exists so lawyers do not have to become good at prompting. That is minimising the mastery a user builds, which is exactly right for adoption and forecloses a Skill loop. Both companies made an expensive, irreversible choice. Most companies never make one.

The loop neither of them is building

Both sell into a profession with extraordinary lateral mobility, and both are leaving the same acquisition mechanism on the floor.

A lawyer fluent in Harvey who moves firms arrives pre-trained, carrying demand into a new account at zero acquisition cost. A legal operations leader who ran Sandstone at one employer arrives at the next one pre-sold. A person leaves and the demand goes with them. That is a User loop sitting in plain sight, in a category where people change employers constantly.

Neither company has a visible instrument for capturing it. No alumni motion, no identity that travels, no way for a former user to raise their hand at a new employer. In a category this well capitalised, with this much attention on it, the cheapest available loop is the one nobody has built. The first instrument for it is a field in the CRM.

What this is actually about

The asset each company is compounding is largely a function of what it could afford to wait for.

Trust compounds over years. A Content loop compounds continuously, but only inside the account it sits in. Capital compounds quarterly, and outside capital is not a Capital loop at all: it stops the moment you stop raising it. Founders tend to choose by aspiration, and they aspire to the loop that produces category ownership, which is almost always the slowest. The real constraint is patience: how long you can keep feeding something that has not paid back yet. Choose an asset whose payback is longer than your runway and you will abandon it in month nine and call it a pivot.

Three questions, whatever category you are in.

Loops are the engine. Moats are the result.