The Companies Quietly Changing the Way We Do Business

The companies that dominate the headlines about artificial intelligence are the ones that build models and the ones that sell the chips and data centres to run them. They are enormous, they are covered every day, and they are not, for the most part, the companies that change how a law firm drafts a contract, how a hospital writes a note, how a bank moves money to a supplier in another country or how a warehouse moves a tote from one conveyor to another. Those changes are being made by a different, quieter group of businesses. They rarely have consumer brands. Their customers are professionals and institutions. Their products are tools for doing one job well. And in the first half of 2026 several of them have grown faster than almost any business in the history of software. This study looks at that quieter group. It asks who they are, what they have in common, whether the numbers reported about them can be trusted, what could go wrong, and how a founder or an operator can recognise this kind of company early. I have drawn on company announcements, funding data, named surveys and trade reporting available up to the middle of July 2026, and I have flagged claims that come from promotional sources or from aggregators that repeat one another. Where I give an opinion, I say so.

Start with what I mean by quiet. I do not mean small, since several of these companies are valued in the billions. I mean that their influence comes from sitting inside a workflow, not from being a product that people talk about. A lawyer who uses an AI platform every day to review documents does not tell friends about it at dinner. A finance team that routes supplier payments through a new settlement layer may not know which blockchain is underneath. A warehouse manager who rents a robot by the hour cares about totes per shift, not about the robot's neural network. These businesses succeed by becoming ordinary, and ordinary is exactly what makes them durable. A product that people admire can be replaced by the next admired product. A product that has become part of how a department does its job is harder to remove.

I will group the companies into five families, because the families matter more than any single name. The first is vertical software agents, which take a general model and wrap it in the knowledge, permissions and workflow of one profession: law, medicine, customer service, accounting, insurance. The second is the connective layer that lets software agents work safely inside a company, finding the right information and respecting who is allowed to see it. The third is the plumbing of money, where payments companies are quietly rebuilding how value moves between businesses. The fourth is physical automation sold as a service. The fifth is the set of businesses that make markets: platforms that connect scarce expertise to the companies that need it, or scarce content to the systems that want it. I will take them in turn, and then step back to ask what they share.

Begin with law, which is the most mature of the vertical categories. Harvey, which builds AI for legal work, raised 200 million dollars in March 2026 at an 11 billion dollar valuation, co-led by GIC and Sequoia, up from about 8 billion dollars the previous December. The company said customers were running more than 25,000 custom agents on its platform, and trade accounts describe it as the default platform among the largest law firms. Its reported revenue figures vary by source and by date, with different outlets citing annual recurring revenue of about 190 million, about 195 million, about 300 million and about 350 million dollars at different points in the year. That spread should make any reader cautious, and it is one reason I would not quote a single figure. What is not in doubt is the order of magnitude and the speed. A company founded in 2022 now has a revenue base that most software companies take a decade to reach.

Legal AI also illustrates a pattern that recurs across the quiet companies: the incumbents are not standing still. Euclid Ventures' annual vertical report counted 97 legal deals worth about 3.3 billion dollars, and noted that two of the year's largest vertical rounds were for case management platforms, Clio, which raised 850 million dollars and bought the legal research company vLex for about a billion, and Filevine, which raised 400 million. Neither started as an AI company, but both are rapidly becoming AI powered, which tells you that the established software vendors in each profession have the customer relationships and the data and will not give them up easily. The legal market will therefore not be won by whoever has the best model. It will be won by whoever best combines a model with the files, the integrations, the accuracy standards, the privilege rules and the trust of the profession. Accuracy is a particular issue in law, where a fabricated citation is not a minor error but a professional disaster, and the vendors that survive will be those that make verification part of the product.

Customer service is the second vertical to watch, partly because it is the first in which the pricing model has visibly changed. Sierra, founded in 2023 by the former Salesforce co-chief executive Bret Taylor and the former Google executive Clay Bavor, reportedly grew annual recurring revenue from about 26 million dollars at the end of 2024 to about 150 million dollars by January 2026 and was reported at around 200 million later in the year, and it raised 950 million dollars in May 2026 at a valuation reported between 15 and 16 billion dollars. Decagon, which builds similar agents, raised 250 million dollars in January at a 4.5 billion dollar valuation after adding more than a hundred enterprise customers in the previous year. Intercom, Zendesk and HubSpot have all introduced per resolution pricing, in which the customer pays when the agent resolves an enquiry. The interesting question about these companies is not whether agents can answer questions, which is established, but whether the economics hold when every competitor cuts prices and the underlying models become cheaper. My view, which I developed in my study of business models, is that the durable players will be those that own the integration with a customer's systems, the measurement of resolution quality and the escalation paths to humans, not those that merely supply the conversation.

Healthcare is the third, and I covered the evidence for ambient clinical scribes in my review of industries. The company most associated with the category, Abridge, raised a large round at a valuation of about 5.3 billion dollars in early 2026, following a 316 million dollar Series E the previous year, and competitors such as Ambience and OpenEvidence have raised nine figure rounds as well. The independent studies show modest time savings and meaningful reductions in reported burnout, and the commercial story rests on that second finding as much as the first. A hospital that keeps clinicians from leaving is saving real money, even if the minutes saved per visit are small. The competitive danger here is visible and worth naming: in January, both OpenAI and Anthropic announced healthcare offerings with the privacy and integration features that hospitals require, according to trade reporting, which shows that model providers do not intend to remain purely horizontal. A vertical company that depends on a model provider for its core capability may find that provider competing with it.

I mention that risk in every vertical, because it is the central strategic question for the whole group. The vertical companies argue that domain depth, proprietary data, regulatory permission and workflow integration give them a defensible position. The model providers argue, implicitly, that a sufficiently capable general system will make much of that depth unnecessary. The evidence so far favours the vertical companies in regulated and relationship driven fields, where trust, liability and integration with existing systems matter as much as raw capability. It is less favourable in fields where the task is simple and the data is public. A founder choosing a vertical should therefore ask whether the barrier to entry is the intelligence, which is becoming a commodity, or something else, such as a licence, a data asset, a distribution channel or a reputation, which is not.

The second family, the connective layer, is easier to overlook and may be among the most important. When a company deploys software agents, it must solve a problem that has nothing to do with intelligence: how does the agent find the right information across dozens of internal systems, and how does the company make sure the agent only sees what the person using it is allowed to see? Glean, which began as enterprise search and is now positioned as a work AI platform, reported surpassing 300 million dollars of annual recurring revenue in May 2026, only fifteen months after crossing 100 million, and said the number of Fortune 500 customers had nearly doubled in a year. Its value lies in the unglamorous work of connecting to hundreds of applications, preserving permissions and keeping data clean, which a company needs before any agent can be trusted. An analysis by Andreessen Horowitz of where enterprises are adopting AI found that coding, customer support and search account for the lion's share of use, which fits this picture: the first value is in finding and using what the company already knows.

The third family, the plumbing of money, is where I think the quietest and most consequential change is happening. In my review of finance I described how stablecoin rules in the United States are moving from proposals to effect. Here I want to look at the company that has done most to turn the technology into infrastructure for ordinary businesses. Stripe, which said it processed 1.9 trillion dollars of payments in 2025, up 34 percent, and was valued at about 159 billion dollars in February, has assembled a stack for moving money as stablecoins. It bought the stablecoin orchestration company Bridge in 2025 for about 1.1 billion dollars, acquired the wallet company Privy, and with the investment firm Paradigm incubated Tempo, a blockchain designed specifically for payments, which raised 500 million dollars at a 5 billion dollar valuation and went live on its main network in March 2026 together with a protocol allowing software agents to pay for services. Stripe's annual letter reported that Bridge's volume more than quadrupled in 2025, and that stablecoin payments globally roughly doubled to about 400 billion dollars with around 60 percent of it between businesses. Mastercard announced in March that it would acquire the stablecoin infrastructure company BVNK for about 1.8 billion dollars.

Why do I call this quiet? Because the end customer will not know it is happening. A business that pays a contractor abroad through a Stripe powered platform will see a fast, cheap payment and will not care what moved underneath. That is how infrastructure changes win: they do not ask the user to change behaviour, they make the old path cheaper until it stops making sense. The strategic point is that the company already holding the merchant relationships, which Stripe does, can add a new settlement layer with far less friction than a start-up could. It is a good example of how incumbents absorb a technology rather than being displaced by it. For small businesses elsewhere, the lesson is that the most important stablecoin products are likely to be features inside tools they already use, not a separate crypto account they must learn to manage.

The fourth family is physical automation sold as a service, which turns a risky technology into something a cautious customer will try. Agility Robotics, whose humanoid robot Digit moves containers in warehouses, has reported more than a hundred thousand totes moved for a logistics customer under a robots as a service contract and signed its first commercial manufacturing agreement with Toyota in Canada in February 2026 after a year long pilot. Figure's robots contributed to the production of more than thirty thousand vehicles during an eleven month deployment at a BMW plant in South Carolina, and BMW extended its humanoid programme to a German plant in March. As with the other categories, I would treat the reported results with care, since they come largely from the companies and their partners. But the commercial structure is worth noticing. When the vendor charges per task and carries the risk of downtime, the customer's decision becomes an operational trial and not a capital bet, which is how new industrial technologies get into factories. I expect the financing of such fleets, and the insurance and maintenance around them, to become businesses of their own.

The fifth family is the market makers, businesses that connect scarce things. The most striking example in 2026 is the set of companies that supply human expertise to train and evaluate AI systems, with one such platform reported to have reached a revenue run rate in the high hundreds of millions of dollars within roughly a year and a half. I would treat the precise figures with real caution, as they come from growth trackers and press reports and not from audited statements. But the existence of the category is significant. As systems become more capable, the scarce input is no longer raw data but expert judgement, lawyers who can grade legal reasoning, doctors who can assess clinical advice, engineers who can review code. A marketplace that finds, vets and pays those experts is a classic two sided business, and it benefits from the same dynamics that made earlier marketplaces durable. It also raises questions about the conditions of the people doing the work, and about how dependent the model builders are on a small number of suppliers, which I will leave to others to explore.

Now to the question of what these companies share. I see six common features. First, each owns a workflow, not a feature. They are used daily, by named people, for a defined purpose, and they are hard to remove. Second, each has a measurable outcome. A document reviewed, a conversation resolved, a note written, a payment settled, a tote moved. Because the outcome can be counted, the customer can justify the spend and the vendor can price on it. Third, each deals with trust in a field where trust is scarce: privilege in law, privacy in medicine, authorisation in payments, safety on a factory floor. Fourth, each depends on data and integration that take time to build and cannot be copied by writing a prompt. Fifth, each reached scale through distribution into a particular profession or institution, through bar associations, hospital partnerships, merchant networks or enterprise sales, not through consumer marketing. And sixth, each is exposed to the same two threats: model providers moving into their territory, and competitors cutting prices as the cost of intelligence falls.

It is worth pausing on the quality of the numbers, because the quiet companies attract so much attention that the reporting around them is full of noise. Reported revenue for the same company can differ by almost a factor of two depending on the outlet and the date. Valuation multiples quoted in aggregator posts are often asserted without a source. Lists of the fastest growing companies rank them by metrics defined differently from one list to the next, and several of the revenue claims come from the companies themselves, which have every reason to choose the most flattering measure, whether annual recurring revenue, annualised run rate or gross bookings. A founder or investor reading these figures should ask three questions: is the number recurring or one off, is it gross or net, and is it from the company or from an independent source? Using that filter, I have kept to the figures that appear in company announcements or multiple independent reports and avoided the more extravagant claims.

The risks deserve a careful statement, because the enthusiasm in this field is easy to over-read. The first risk is valuation. When companies are valued at many multiples of their revenue, any slowdown in growth can cause a sharp repricing, and the funding data I reviewed in my broader work show how concentrated and fast moving the capital is. The second is competition from the model providers, which have both the technology and the customer access to enter the most attractive verticals, as the healthcare announcements show. The third is the price compression that comes from falling inference costs and aggressive competitors, which may squeeze the margins of companies that sell tasks at prices set when intelligence was expensive. The fourth is accuracy and liability. In law, medicine, finance and any regulated field, an error costs money or harms people, and a high profile failure could prompt regulators to impose rules that slow adoption. The fifth is dependence on a few suppliers, whether for models, for computing or for expertise. None of these risks is fatal, but together they mean that today's leaders are not guaranteed to be tomorrow's.

How can a founder or operator recognise a quiet company early, before it appears on lists? I would look for five signals. The first is a customer who uses the product every working day without being asked to, which is visible in retention and in how quickly the product spreads inside an organisation. The second is a price that tracks a measurable outcome the customer already cares about. The third is a moat made of something other than technology, such as a licence, a dataset, a distribution channel or a reputation. The fourth is an unglamorous problem: reconciliation, scheduling, prior authorisation, claims intake, compliance filing, supplier onboarding. The loudest markets attract the most competition, and the quiet ones often have fewer capable players. The fifth is a founder who knows the profession from the inside, who has sat in the room where the work is done and understands which part of it is truly painful. In my experience this last signal is the most reliable, because it is the hardest to fake and the hardest to acquire from outside.

That leads me to what I think is overlooked. The categories drawing the most attention are law, health, customer service and coding, because they are large and visible. Beneath them are dozens of narrower niches where the same logic applies and the competition is thin: construction scheduling and estimating, insurance claims intake, freight documentation, accounting close and reconciliation, procurement, utilities, agriculture advisory, education administration, municipal services. Venture data from the first half of the year shows deal flow in vertical AI spread across many categories, with legal, insurance, construction and healthcare capturing the largest share of the dollars, and the point I draw is that there is room beneath the giants. A company that wins a narrow niche in a market too small for a large vendor to bother with can be profitable with a modest team, and the reduced cost of building software has made such businesses far more viable than they were a few years ago.

Turn now to Bangladesh, because the question of who changes how business is done is as live here as anywhere. The conditions that create quiet companies elsewhere exist here: large numbers of small and medium businesses running on paper, spreadsheets and phone calls; sectors such as garments, education, healthcare and agriculture with specific, well understood pain points; mobile money that already makes digital payment normal; and a pool of capable engineers. What is missing is the product built by someone who knows the sector well enough to solve one problem completely. I can think of many candidates. A system that handles fee collection, attendance and results for schools, which I have worked on, is one. Factory compliance and audit documentation for exporters is another. So are clinic scheduling and records for small practices, reconciliation for businesses that accept payments through several wallets and gateways, and procurement tools for small manufacturers. None of these would make a headline, and several could become sound, profitable businesses that change how an industry operates.

There are particular cautions for builders here. Most of the quiet companies I have described sell in dollars to buyers with large budgets, and a Bangladeshi company will often sell in taka to buyers with small ones, so unit economics must be proven at lower prices and with lower support costs. Trust matters as much as it does elsewhere, and a small firm entering sectors such as health or finance will have to earn it patiently, through pilots, references and visible care with data. Regulation is real, as the recent attention to the governance of mobile financial services shows. And dependence on foreign platforms for models and computing is a risk that should be managed by keeping systems portable and costs measurable. The advantage is knowledge. A founder who understands how a school actually collects fees, or how a factory actually passes an audit, holds information that no global company will acquire quickly.

Let me offer predictions, with the usual caution that they are judgements and not forecasts. Over the next year I expect at least one of the leading vertical AI companies to face a serious accuracy or liability controversy, because the volume of work now flowing through these systems makes errors statistically certain, and I expect the response to be a stronger emphasis on verification and audit features. I expect the model providers to continue entering specific verticals, and I expect some vertical start-ups to be acquired by larger software vendors or by the model providers themselves, as the economics of independent growth tighten. I expect per outcome and hybrid pricing to spread from customer service into other categories where results can be measured. I expect the stablecoin infrastructure that payments companies are building to be adopted first for cross border business payments and agent initiated payments, and to remain invisible to most consumers. And I expect the physical automation vendors who offer service contracts to find that financing and maintenance, more than robotics technology, determine who scales.

I would also make a prediction about the shape of the next wave of quiet companies. As the cost of building software falls, the next group will be smaller, more specialised and less dependent on venture capital than the current leaders. Many will serve a single profession in a single country or language, and many will be run by people with deep experience in the field who learned to use the new tools. They will rarely be mentioned in the technology press, and they will not need to be. If I am right, the most important business story of the late 2020s will be not a handful of giants but thousands of modest companies that quietly make particular kinds of work cheaper, faster and more reliable.

There are limits to this analysis that I should state plainly. The companies discussed are those with the most public information, which means they are the successful ones, and a survey of leaders cannot tell us how many similar companies failed. Many figures come from company statements, funding announcements and trade reporting that cannot be audited. The field is moving so fast that a ranking written today will be obsolete in months. And some of the claims I have mentioned, particularly about market makers for expertise and about the revenue of specific start-ups, rest on thinner evidence than I would like. I have tried to flag these cases, to avoid the most extravagant statistics and to prefer events that were announced publicly, but readers should check any figure they plan to rely on against the original source and its date.

To gather the threads: the businesses changing how work and money move in 2026 are mostly not the ones in the headlines. They are vertical agents that have become part of how lawyers, doctors and support teams do their jobs; the connective layer that lets agents work safely inside organisations; the payments infrastructure that is turning stablecoins into a feature of ordinary business tools; the robots sold by the task; and the marketplaces that connect scarce expertise to the systems that need it. They share a habit of owning a workflow, delivering a measurable outcome and earning trust in a field where trust is hard, and they share two threats, model providers moving into their territory and prices falling beneath them. For a founder, the practical conclusion is encouraging: the opportunity lies less in building something dazzling than in understanding one kind of work so well that you can make it measurably better. The companies that do this are the ones that will still be around when the headlines have moved on, and they are the ones most likely to change the way the rest of us do business without our noticing.

Sources: Landbase, Sequoia-backed companies in 2026 (https://www.landbase.com/blog/fastest-growing-companies-backed-sequoia-capital); SaaS Mag, vertical AI agents (https://www.saasmag.com/vertical-ai-agents-eating-horizontal-saas/); Euclid Ventures, The Vertical Report 2026 (https://insights.euclid.vc/p/the-vertical-report-2026-full-version); a16z, Where enterprises are actually adopting AI (https://a16z.com/where-enterprises-are-actually-adopting-ai/); Perplexity AI Magazine on leading AI start-ups (https://perplexityaimagazine.com/blog/hottest-ai-startups-silicon-valley-2026/); CoinDesk, Stripe annual letter and Bridge (https://www.coindesk.com/business/2026/02/24/stripe-s-bridge-sees-stablecoin-volume-quadruple-as-utility-insulates-from-crypto-winter); CoinDesk, Tempo mainnet launch (https://www.coindesk.com/tech/2026/03/18/stripe-led-payments-blockchain-tempo-goes-live-with-protocol-for-ai-agents); Stripe, Sessions 2026 announcements (https://stripe.com/blog/everything-we-announced-at-sessions-2026); Solid Market Research on humanoid robot deployments (https://www.solidmarketresearch.com/post/humanoid-robots-cross-the-pilot-threshold-where-factory-deployment-actually-stands-in-2026); Capgemini findings summarised by Theresa Robot That (https://theresarobotforthat.com/blog/humanoid-robots-manufacturing-readiness-2026/); Build MVP Fast on vertical AI and model-provider competition (https://www.buildmvpfast.com/blog/vertical-ai-eating-horizontal-saas-2026); Pickaxe on AI agent pricing models (https://pickaxe.co/post/ai-agent-pricing-models).