78% of Clients Want AI, Only 6% Get It

A new kind of company sells the outcome, not the software

78% of corporate buyers of professional services, the people who hire lawyers, auditors, and tax consultants, say it matters or is required that their vendor use AI to raise the quality of the work. That's from Thomson Reuters' "Future of Professionals Report 2026": 1,816 professionals in law, tax, audit, accounting, compliance, and risk, across 62 countries, surveyed in March and April 2026, published in June. Only 6% of those buyers actually get that quality from most or all of their vendors today.

The gap between 78 and 6 is exactly the market a new type of company is chasing. These companies don't sell a tool the client turns on themselves. They sell the finished result: a filed tax return, a signed contract, an issued insurance policy. Behind the scenes, the work runs on agents, AI-powered programs that complete a task on their own without a person guiding each step. The client only ever sees the outcome and its price.

Y Combinator, the startup accelerator that helps launch technology companies, published a breakdown of this model on June 3, 2026, called "How to Build an AI-Native Services Company." The author, Charlie Warren, is a visiting partner at YC. Worth remembering: this is the accelerator's position, not a market measurement. YC is backing a specific kind of company and speaking on its behalf.

The bet behind it is large. Sequoia Capital, a venture firm that has funded dozens of well-known tech companies, put a number on it in Julien Beck's March 5, 2026 piece "Services: The New Software": for every dollar businesses spend on software, six go to paying people to deliver services. General Catalyst, another major fund, reached a similar picture from different numbers in its August 28, 2025 report "The Future of Services": US service industries generate more than $6 trillion in revenue a year, against $370 billion for the entire software market.

General Legal shows what this looks like in practice. It's a law firm from Y Combinator's Winter 2026 batch that specializes in commercial contracts. Co-founder J.P. Moler told Artificial Lawyer on March 31, 2026, that the firm charges a flat $500 per contract. A master service agreement that takes a lawyer at a large firm 8 to 10 hours takes General Legal 2.2 hours. Margin, the share of revenue left over after the direct cost of doing the work, runs 40-50% per contract. One lawyer here handles 300 to 600 contracts a year.

Bloomberg Law added company numbers on July 14, 2026: General Legal has raised $11.5 million, runs at an annualized revenue of $2 million, and employs 14 lawyers, nearly all from large firms, every one of them holding equity.

Which work gets automated first

YC lists four traits of the work that gets taken over first.

The first: it's already outsourced, meaning handed to an outside vendor rather than done by in-house staff. The client in this arrangement is already looking at the result, not at how it got made.

The second: at the level of a single task, almost no human judgment is required. Judgment isn't spread across every step. It's concentrated at a few checkpoints.

The third: the work is genuinely complex overall. A simple tool or template can't cover it. It takes a chain of many steps.

The fourth: the industry is regulated. That sounds like a downside, but YC calls it an advantage. A regulatory bar keeps new competitors at a distance, and it protects whoever has already cleared it.

YC names the markets directly: taxes, audit, insurance, mortgages, parts of healthcare, parts of logistics. The common thread is paperwork and procedural work with clear internal rules, currently done by hand by expensive specialists.

One piece of advice stands apart, and it's blunt: be careful with equipment and on-site work. Wherever a company owns physical things and services them with human hands, the high-margin software math stops working. A truck roll and a repair cost money no matter how much intelligence AI adds to the process.

Four pedestals of different heights, each holding a white office object: a stack of documents, a rubber stamp, a calculator, a paper tray

The model is no longer the edge

Not long ago, access to a stronger model was itself a competitive advantage. Recent data says that's over.

Stanford HAI, a research institute at Stanford that publishes the industry's leading annual AI report, measured how models handle tasks in taxes, mortgage processing, corporate finance, and legal reasoning in its April 2026 "AI Index Report." The result: 60% to 90% correct, depending on the task.

The spread depends on the task. Stanford HAI reports scores of 74.20% to 77.11% for fifteen models on TaxEval v2, a 2.91 percentage-point gap. On CaseLaw v2, scores range from 62.06% to 73.42%, a gap of 11.36 points. Model selection therefore needs testing on the service’s actual tasks.

Stanford HAI also records the flip side: agents still fail roughly one attempt in three. That's not a rounding error. Nobody should build a paid service on a bare model with no layer of checking and retrying.

That's where YC's survival test comes from: as models get stronger, your service either gets stronger along with them, or the model wipes out your edge. A company whose only advantage is access to a specific model is doomed, because tomorrow any competitor gets the same model. The advantage has to sit in what's built around the model, not in the model itself.

An even row of identical white cylinders, all the same height, none standing out

The product is the operation

Once a service gets rebuilt around AI, it starts getting managed like a software product instead of a team of people. That changes which numbers a leader checks every day.

Throughput, how much work the system processes in a given period, and cycle time, how long one unit of work takes from start to finish, become product metrics. YC's advice is to watch them as closely as website traffic, daily, not once a quarter at a board meeting.

Cost of service, all the direct expenses of delivering it, breaks into three pieces. First, fees for running the models, which rise with volume. Second, hosting: the servers and infrastructure everything runs on. Third, the people still in the process. Each piece needs its own number, its own trend over time, and its own owner, not one blurry line item labeled "AI costs."

The people still in the loop deserve a separate conversation. If revenue grows in exact lockstep with headcount, the business is built wrong. It means AI hasn't automated much of anything, just added one more tool to manual labor. People need to scale nonlinearly, with revenue growing faster than staff.

YC offers a blunt gut check for any leader running this kind of company: are the people in the process there because the work genuinely needs their judgment, or because they're patching holes in the product? The second answer is a warning sign. It means the marketing promises automation while the service is still being pulled along by hand under an AI label.

There's a mirror-image trap, though, and it hits early, not once a company matures. Pilots, trial projects run with a new client, are easy to sign in bulk at the start. Interest is high, and selling is easy. The problem is servicing them all: with nothing built to handle the load, a company gets stuck in manual labor, patching every pilot by hand instead of building a real system. YC's advice is to cap the first round of pilot clients at a couple, until the process actually holds up.

A yellow conveyor belt carrying identical white cubes, magenta floor, cyan wall

Quality variance is a threat to the business itself

For this new kind of service company, quality variance, the same request sometimes producing a strong result and sometimes a mediocre one, isn't a small flaw. It's a threat to the business surviving at all. Clients fire a vendor over unpredictable output far faster than they fire one over being a bit pricier or a bit slower than competitors.

Stanford HAI's number comes back here: agents today still fail roughly a third of their attempts. For a one-off chat with a model, that's tolerable. A person just asks again. For a paid service running at scale, a third of failures means a third of unhappy clients, unless there's a separate layer catching and fixing those failures before the result ever reaches the customer.

A shelf with an even row of identical white mugs, broken mugs and shards on the floor below

Sell the outcome, not the seats

Pricing matters just as much as process for an AI-run service. YC is direct about it: sell the outcome, not employee seats and not model usage measured in tokens.

Two pricing models actually work. The first is per-unit pricing: per tax return, per customer inquiry, per loan processed. The client pays for each finished item, and the price doesn't depend on how much manual labor sat inside it. The second is pricing the whole outcome: not the contract as a document, but the deal that got signed.

YC flags two pricing models to avoid. Cost-plus, where a company tallies its AI and labor costs and tacks on a percentage, permanently hands the client every bit of savings from cheaper technology: the model gets cheaper, the price follows it down, and margin never grows. Straight-up dumping, deliberately selling below cost to grab market share, makes the entire category of work look cheap to the market and cuts its future price for everyone, including the company doing the dumping.

Panacea, a company from Y Combinator's Spring 2026 batch, shows a third path. It handles regulatory work for biotech companies bringing new drugs to market. It hires former regulatory-agency consultants with years of experience and puts them to work on top of its own AI platform. It charges by project milestone, paid on completion, not by consultant hours.

The incentive difference here is the whole point. An hourly consultant has no reason to move faster with AI, because speed cuts straight into their own revenue: finish sooner, bill fewer hours. With a fixed price paid per milestone, the incentive flips. The faster and cleaner the team closes a milestone, the higher its own margin.

That's the economics these new service companies are betting on. Traditional service firms have spent years capped around a 30% margin. The bet from the companies YC writes about is to push margin closer to software levels, 50% and up, in a market two to three times bigger than software. That's YC's estimate and its bet, not an industry-wide measurement. There are no confirmed numbers for the whole market yet.

Two empty white office chairs on the left, a single white sphere on a magenta pedestal on the right

Bolting AI onto an old firm doesn't work

Given how large this bet is, the obvious temptation is to buy an existing traditional service firm and bolt AI on top, rather than build from scratch. YC calls that almost always a trap.

The reason is simple. Product-market fit, the moment a company's offering genuinely clicks with customers who are willing to pay for it, doesn't come bundled with the acquisition. An old service firm already has its own metrics, its own hiring, its own expectations about results, all shaped over years around a different way of working: hourly, with people at the center. Rewiring those expectations is harder than building a new company from zero. The only good reason to buy an old firm, in YC's view, is to get a regulatory license fast, one that would otherwise take years to obtain from scratch.

Thrive Holdings is the most visible example of exactly this attempt. In August 2026, it raised $2 billion at a $12 billion valuation. Its model is to buy traditional service businesses and rebuild them around AI. TechCrunch reported on August 12, 2026, that the company already has more than 70 businesses on its platform.

The numbers Thrive shares about itself are striking: its accounting arm covers more than 50 firms and more than 2,000 specialists; its tax agent has processed over 7,000 returns at 98% accuracy, and prep time at participating firms has dropped by more than 30%; its IT division closes tickets 36 times faster than before. All of these are the company's own self-reported figures, not an independent check, which is worth keeping in mind while reading them.

General Catalyst, a fund investing in similar deals, sets a concrete target for the service businesses it buys: move growth from single digits to 10-20% a year and double margin, aiming for 30-40%. Compare that with what YC describes for companies built from scratch: 50% and up. A firm that's been bought and retrofitted moves noticeably slower and more cautiously than one designed around AI from day one.

For a leader currently choosing a vendor, or deciding what to do with their own service function, the conclusion follows from two numbers already in this piece. Demand for AI-level quality is real: 78% call it important or required. Supply barely exists: only 6% get it. A third number, from the same Thomson Reuters survey, shows what separates the companies pulling ahead from the ones standing still: where a company has a named AI strategy, 66% of employees say AI meets or beats expectations. Where there's no strategy, only 22% say the same. The difference isn't access to AI. It's whether a strategy has actually been built around it.

An old grey typewriter with magenta robotic arms crudely bolted onto both sides, cables hanging loose

Q&A

Our company doesn't sell services, we buy them. Does this even apply to us?

Yes, on both sides. As a buyer, you're one of the 78% in the Thomson Reuters survey. It's worth asking every outsourced vendor, your lawyer, your auditor, your tax consultant, what exactly they do with AI and how they measure the quality of the result, not just whether they use it. As a company, you almost certainly provide some service to other businesses yourself, and it's worth running it through YC's four traits: is it already outsourced, does it need judgment at every step, how complex is it, how regulated is the industry?

We don't have our own tech team. Are we out of this game?

Not necessarily. Panacea and General Legal show that the core of these companies isn't programmers. It's specialists with real experience in the underlying service who work on top of a ready-made AI platform. The tech layer can be built in-house, bought as a product, or commissioned from a team that specializes in embedding AI systems inside a business. The decision that matters isn't who writes the code. It's how the process, the pricing, and the quality control around the outcome are structured.

We already use AI inside our team. Isn't that enough?

Thomson Reuters found that 74% of professionals use AI several times a week, and 44% several times a day. That's a norm by now, not a point of difference. The deeper problem: 41% of people using AI at work have no access to professional tools built on vetted, industry-specific material, and 34% use tools their organization never approved. Even where professional tools are provided, 18% of employees still don't use them. Employees using AI on their own and AI built into the service itself, as a product with measurable quality, are two different things. The first almost always already exists. The second is rare, and it's the second one clients are actually looking for.

Where do we start if we want to check our own service or vendor right now?

Start with three numbers. First, margin: how much actually remains from revenue after the direct cost of doing the work, including AI, infrastructure, and people, and which way it's trending. Second, cycle time per unit of work, and whether it has changed at all since AI entered the process. Third, the share of requests or tasks where the result has to be reworked by hand, which is exactly the quality variance that drives clients away fastest. If nobody at the company tracks any of these three numbers on a regular basis, that's where to start, before any conversation about models or agents.

Majento builds and deploys AI software and agentic systems for marketing and business processes, using a forward-deployed model: engineers embed inside the client company's own processes, find the bottleneck, build a working solution, and see it through to real use. If you want to run your own service or your vendors through this lens, reach out on Telegram at t.me/shimaoz or by email at hello@majento.ai.

78% of Clients Expect AI From Vendors. Only 6% Get It