The AI-Native Services Playbook: An Operator's Guide

Sequoia says the next trillion-dollar company will be "a software company masquerading as a services firm." Emergence wrote a playbook on it. YC built a startup school module around it. a16z and Bessemer are committing billions to the thesis.

Every one of those was written from the cap table. I have spent the last few years on the other side of it.

We started by selling the tool... two years building AI automations for investment banks and private equity funds. We met thousands of them and the adoption was brutal. Not for safety or compliance reasons. Bureaucracy. So we stopped selling the tool and started selling the work. We have now taken on 25 live deals, ended May with $30M+ in pipeline, and closed Europe's first AI-native M&A deal.

The playbooks are good. I send them to people. But they describe the shape of the thing from the outside. This is the view from inside the operation, where you feel the mistakes first-hand.

A quick note before we start. I will use M&A as my running example because it is the operation I actually run, and the internal lens is the only one worth anything here. But almost everything below applies just as well to law, insurance, accounting, tax, healthcare admin or IT services. So I will keep pointing to companies in those spaces too. If you have never touched a deal in your life, the pattern should still be obvious.

You are not building a software company

The cleanest way to understand this market is the difference between a copilot and an autopilot. A copilot sells a tool to a professional and lets them do the work. An autopilot sells the work itself. We learned the difference the hard way, by spending two years as a copilot before we understood we were in the wrong business.

The moment it clicked was an economics question. The M&A tech market is measured in billions. The M&A services market is measured in hundreds of billions. The work budget does not just beat the tool budget… it dwarfs it, something close to a hundred to one. So we asked the obvious question. Why sell the tool when we can do the work instead?

You see the same split everywhere once you look for it. In legal, a copilot sells software to a law firm. Crosby sells the finished NDA to the company that needed it. In insurance, a copilot sells a tool to the broker. WithCoverage sells the cover to the CFO who needed it. In every case the autopilot reaches past the professional and sells the outcome directly, because the outcome is where the real budget has always lived.

That single decision inverts almost every software instinct you have, and if you come from tech you will fight it for months. We did.

Your customer never touches your product. You are the product. The AI is internal leverage. So you stop measuring activation and engagement, because your only users are your own team, and you start measuring throughput, cycle time and variance instead. You stop shipping features and start removing human minutes from a process.

And you accept that delivery is not a support function you will sort out later. Delivery is the company. Our first deal... 15-day from outreach to NBO... was not won by a model. It was won because the entire process... buyer universe, teaser, outreach, deal materials... had been rebuilt so that a tiny team could run it at a speed a traditional boutique simply cannot match. The founders of Hanover Park, an AI-native fund administrator, understood this so well that their CEO physically sleeps at a customer's office during migration. That sounds insane until you realise the hardest part of this model is never the AI. It is the handoff from the legacy process. If you treat delivery as something to figure out after the product, you end up as a slow incumbent with a clever demo. That is the single most expensive mistake available to a software founder here, and it is the default one.

Three ways to start, and the one we actually took

There are really only three ways to arrive at an AI-native services company, and they are not equal.

• You can start as a software provider to the vertical and then become your own client.
• You can start from scratch with deep industry expertise and add the technology.
• You can buy a legacy firm and try to retrofit AI into it.

We took the first, so I will spend most of my time there, but you should understand the trade-offs of all three before you commit. The path you pick decides what you are missing, and therefore what you have to go and acquire.

Path one: sell the tool, then become the customer. This is our story. For two years we built and sold AI software to investment banks and private equity funds. We learned the vertical from the outside in... we could see exactly where the inefficiency sat, because we had built a tool to attack almost every piece of it. What we could not do was make the industry adopt them. The cycles were brutal and the budget never really moved.

That failure turned out to be the most valuable thing that ever happened to us, because it pointed straight at the pivot. If the tool budget will not move, go to where the budget already is, which is the work itself. Sequoia describes copilot companies trying to become autopilots and getting stuck on the innovator's dilemma... selling the work means cutting your own paying customers out of doing it. We did not have that problem. Our customers were slow to buy, so there was little to cannibalise. The slow sales cycles that felt like failures were actually the green light. That is a strange thing to write, but it is true, and if you are a struggling vertical SaaS founder it might be the most useful sentence in this article.

The head start this path gives you is real, and it is specifically two things: model fluency and engineering DNA. We already knew what the models could and could not do in M&A, and we already had a team that could build. Most people starting one of these firms spend their first year acquiring exactly that. We had it on day one.

But do not mistake the head start for the business, because two hard truths hit immediately. First, the software you built as a vendor is mostly the wrong software. We had built for sellability… polish, configurability, an interface a banker would tolerate. Internal software has completely different priorities: throughput, cycle time, the relentless removal of human minutes. We rebuilt large parts of it for ourselves, because now the only user we cared about was our own deal team. Second, and harder, we had no domain credibility and we could not actually deliver the service. Watching bankers work for two years is not the same as having closed fifty deals. So we did the obvious thing and brought in a senior banker with exactly that track record to lead the work. We became, structurally, a services firm with an unusually strong engineering core, rather than a software firm dabbling in services. That distinction is the whole pivot. Make it halfway and you end up as a SaaS company moonlighting as an advisor, which is the worst of both worlds.

Path two: start with the expertise, add the technology. This is the more common route and it is the mirror image of ours. The lawyer who starts the AI-native law firm. The insurance lifer who starts the brokerage. The banker who starts the advisory. Crosby's founder was a top practising lawyer. Harper's founders grew up in insurance families. They begin with the very thing we had to go and buy… credibility, and an intimate, inside-out feel for the operation. Buyers trust them on day one.

What they have to acquire is the part we already had, and it is the harder half to fake. Engineering DNA is not a hire, it is a culture, and the classic failure mode of the domain-expert founder is building an engineering department instead… a separate function shipping dashboards while the firm keeps running the way it always has. The second failure mode is subtler. Domain experts tend to rebuild exactly what already exists, because they know it so well, when the entire point is to reimagine the work for a world where AI does most of the intelligence. If you take this path, your hardest job is giving your engineers the standing and the room to challenge how the work is done, not merely to automate the current version of it.

Path three: buy your way in. Acquire a legacy firm, inherit its customers and its licences, and layer AI on top. On paper it short-circuits everything… instant revenue, instant domain people, instant distribution. In practice it almost never works as a way in, and both YC and Emergence are right to warn against it. You cannot acquire product-market fit. A legacy firm arrives with legacy expectations on metrics, hiring, pace and performance, and AI bolted on top changes none of those realities. You spend your energy fighting the culture you bought instead of building the one you wanted.

There are two honest exceptions. One is when you need a regulatory licence quickly and acquiring is genuinely the fastest legal route to it. The other is timing. Once your platform and your AI-first culture are already solid, an acquisition stops being your identity and becomes an accelerant... a way to bolt on revenue and relationships that now conform to your way of working rather than overwrite it. The order matters enormously. Buy in order to become an AI-native firm and you will fail. Buy because you already are one, and you can move very fast.

The point that sits underneath all three: every path hands you something and leaves you short of something else. We started with the technology and had to go and earn the trust. The domain expert starts with the trust and has to build the engineering culture. The acquirer starts with revenue and has to overcome the culture they inherited. Know exactly which one you are, and be honest and aggressive about closing the gap you started with. The gap is where these companies die.

Picking the market

There are a handful of traits people now list for a good AI-native services market. Work that is already outsourced. Judgement concentrated in a few places rather than smeared across every task. A genuinely hard problem underneath. Regulation that raises the bar rather than blocks you. I agree with all of them, and M&A happens to score well on each.

But after three years.. two building the tool, one running the firm... I would compress the whole thing into one test. As the models get better, does your service get stronger, or do the models commoditise you?

This is the test I run on everything. M&A passes it cleanly. Every model improvement lets us build sharper buyer universes, draft better materials, structure data faster, support diligence more deeply. None of that touches the senior banker on the call closing the deal. The model makes the human more valuable, not less. The same is true of medical coding, where the work is translating clinical notes into tens of thousands of standardised codes... complex, but rules-based, so a better model just does it better and cheaper. A worse market is one where a better model replaces the thing you sell outright. If you cannot answer this test honestly in your own category, stop and pick a different one.

The deeper reason I chose this specific corner of M&A is structural. The economics of that work were broken long before AI arrived. A proper sell-side process is hundreds of hours of work, and the fee on a small deal barely covers the team needed to do it well. So the work either does not get done, or it gets done badly with corners cut. Buyer outreach shrinks to the handful of contacts the lead banker already knows. The process drags. Competitive tension is weak.

You find that same broken-economics signature across services. US accounting has lost roughly 340,000 accountants in five years while demand has risen, and around three-quarters of the CPAs who remain are nearing retirement. Insurance claims adjusters are retiring faster than new ones can be trained to replace them. In each case the old model cannot serve the work at the price the work can bear, and that gap is the opening. That is the kind of void to look for. Not "an industry AI could help." A specific slice where the old economics were already broken, so your leverage does not just make you faster… it makes a previously unservable customer servable.

For us, the wave is succession. In Europe, a third of entrepreneurs plan to step away from their businesses this decade, putting roughly 7 million companies and 30 million jobs in motion. In Germany alone, 626,000 SMEs plan ownership transfers by 2027. And this is not a regional bet. Roughly 6 million companies are expected to transition in the US. In Japan, Shunsaku Sagami built M&A Research Institute on the same crisis, compressed average deal timelines from over 12 months to about 6, and became a billionaire doing it. Same structural void, three continents. When a pattern wins independently in three markets, it is not a fluke.

Deciding what you actually sell

This is the decision the playbooks skip, and it is the one I underestimated most. Picking M&A was not the hard part. M&A is enormous.

The first principle I would borrow here, because it is true and it is what made our model work, is to start where the work is already outsourced. Nobody sells their company alone. They hire an M&A advisory firm. That matters more than it sounds, because it means Eilla is a vendor swap, not a reorg. The budget line already exists. The owner is already buying an outcome rather than a tool. We are not asking anyone to change their behaviour, only to change who they hire. This is exactly why Crosby started with NDAs and not some grander legal ambition... NDAs were already sent to outside counsel, so the budget and the habit were already there to be taken. Replacing a vendor is easy. Asking a company to dismantle an internal team is a war.

The second principle is to split the work into intelligence and judgement, and be honest about the ratio. The intelligence in M&A - the buyer universe, the materials, the data, a large share of the communication and diligence support… is most of the hours, and AI does it. The judgement… the calls, the dinners, the negotiation, the read on a room, the relationship… is a small share of the hours and almost all of the value, and it stays human. Recruitment is the same shape: screening and matching candidates is intelligence and automates well, but closing a candidate and reading culture fit is judgement built on years of pattern recognition. A market where every step needs human judgement cannot scale. You want heavy intelligence with judgement concentrated in a few decisive places.

Then comes the narrowing, and this is where most founders go wide too early and productise nothing. We made a deliberately narrow call. Sell-side, founder and owner-led businesses, broadly $5M to $50M, success-fee-only, no retainers.

What we said no to matters as much as what we said yes to. We said no to retainers, because a retainer is the painkiller that lets a mediocre advisor survive without ever closing a deal, and our entire thesis is that if we cannot close, we die. We didn't focus on large sponsor processes early, because the workflows are bespoke and would have dragged our engineering off the repeatable core. We said no to buy-side mandates, because the data and the incentives are different enough that it is really a second company. Strala did the same thing in insurance by refusing to be a general claims platform and focusing only on claims processing for a specific set of customers. You expand later, from strength. You do not expand because a tempting deal walked through the door.

Building the team

The line that travelled furthest from my last article was that having engineers is not the same as having engineering DNA. I will defend it harder here, because the team is where these companies live or die, and where the software instinct fails most expensively.

You need three things, in this order.

Real domain credibility, because you are selling trust before you have delivered anything. Our processes are led by a senior banker who has closed more than fifty deals across a career at firms the buyer immediately recognises. When that person is on the call, the conversation is different. The owner is handing over their life's work, and a slick interface does not earn that. This is universal in services. Harper's founders grew up in insurance families and carried that credibility from day one. Panacea, an AI-native FDA regulatory firm, hires experienced FDA consultants rather than pretending software alone clears a drug. Mechanical Orchard's credibility came from a CEO who had already run Pivotal. You cannot fake this, you cannot hire it junior, and it is what wins the competitive pitch before the technology gets a chance. If the founders do not carry it, your first serious hire buys it.

Engineering DNA, which is not the same as an engineering department. This is the trap incumbents will fall into and you must not. We have met banks with large engineering teams that are years behind, because every decision routes through a committee that does not understand the technology, and the engineers sit in a separate function shipping dashboards nobody asked for. Engineering DNA means your engineers sit in the room where the work happens. At Eilla, the gap between a banker saying "this is broken" and an engineer fixing it is measured in hours, not quarters. Crosby runs the same way in law: their lawyers sit beside their engineers and give feedback every few hours, and the system improves on that loop. If your engineers have never watched a live process run, you do not have it yet.

A product leader earlier than feels reasonable. This is the one I delayed and regret. Because the customer never touches the software, it feels safe to push product leadership down the road. It is not. Without someone deliberately owning the bridge between the delivery team and the engineers, your roadmap gets set by whoever shouted loudest on the last engagement, and you end up automating noise instead of the things that compound.

On the ratio question everyone asks me: yes, the balance of engineers to bankers should be flipping, and the deal team of the future is one exceptional senior banker, one exceptional junior, and a lot of AI. But do not over-rotate early. You need enough domain weight in the room to win the trust first, then you compound the engineering advantage with every deal.

Building the difference

Tools are not a moat. Every firm in your space will eventually have engineers and a frontier model. So where does a durable edge actually come from?

The data flywheel, built in from day one. Every mandate we run generates data on which buyers respond, which fits convert, which framing lands, which processes stall and why. That makes the next mandate sharper. Harper does the identical thing in insurance: every lead, call, email and policy feeds the system and improves the matching between businesses and underwriters, so conversion climbs over time. The unglamorous operator detail behind this: your engagement letters have to give you the right to learn from the work. If they do not, you are quietly building someone else's flywheel and you will not notice until it is too late.

Proprietary workflow, not proprietary features. The edge is not one clever tool. It is that the entire process has been rebuilt with AI assumed at every step, by a team whose firm dies if the tech fails. An external vendor builds for a problem they observe from the outside. We build for one where the feedback loop is existential, and that loop produces better technology than any SaaS vendor selling the same tool to all our competitors ever could.

Speed, because time kills deals. Our process can go from first call to buyer outreach in days. On a live mandate that means a teaser built, 300+ high-fit buyers identified, a CIM and financial model produced inside a week, NDAs and management meetings inside two, and first non-binding offers in under a month. Speed is not a marketing line, it is the moat. It manufactures real competitive tension for the seller, and it makes us the firm buyers prioritise, because we send them more quality deal flow than any boutique does. Volume of quality builds relationships faster than the occasional phone call ever did.

Brand, borrowed then built. It matters more here than in software, for a structural reason. A SaaS buyer can trial the product before committing. Yours cannot test what they are buying... they are trusting you with an outcome that does not exist yet, so reputation does the job the free trial would. Early on you have no brand, so you borrow it... the senior banker's reputation, the logos on a CV the buyer trusts. Over time, the speed and quality of the work earns you your own. There is no shortcut, but there is a sequence, and getting the sequence right is what lets you start before the brand exists.

The economics nobody warns you about

Obsess over cost of goods sold from day one. Model spend, hosting and any human in the loop all belong in COGS, not buried in operating expenses where they flatter your margins. If your gross margin is not expanding as revenue grows, the AI is not pulling its weight, and you have what Emergence aptly calls Mirage PMF… real revenue growth that is actually powered by labour wearing an AI costume. Be ruthlessly honest about this, because it is very easy to lie to yourself here and feel successful while doing it.

But here is what the financial frameworks underplay, and what genuinely tested my nerve in the early months.

Revenue can be brutally lumpy, and the wrong pricing makes it worse. Software has the decency of recurring revenue. We have months of work and then a single closing that pays for all of it. On success-fee-only with no retainers, we carry every pound of the timing risk. You can do everything right on a mandate and still post a quarter with nothing in the bank. Not every services business has this problem... fund administration and accounting come with multi-year relationships baked in, which is part of why they are attractive. But if your model is lumpy, you need a deliberate mechanism for it. For us, that mechanism is deal volume. We target the SME segment, which is traditionally underserved and lightly contested, and once you are running a large number of deals the revenue smooths out and becomes far more predictable. Build your version of that before you need it, not in the month you cannot make payroll.

The real bottleneck is not the AI. It is senior human capacity. This is the metric I watch above all others, and I rarely see it in a playbook: mandates per senior banker. AI lets us take on ten times the mandates. It does not let one senior banker sit in ten deal rooms at once. The judgement work does not scale linearly, and it never will. And capacity is only half of it. Once the AI absorbs the grunt work, only two scarce inputs remain... deal flow coming in, and senior judgement to convert it. Starve either and you stall. Unlimited sourcing is worthless without the seniors to execute it, and infinite senior time is worthless with nothing to work on. Every AI-native services firm needs its own version of this number. Crosby's is the human review time left on a document after the AI has run... as that approaches zero, their margins approach software margins. Yours might be claims closed per adjuster or returns filed per accountant. Whatever it is, the failure mode is the same: intake outpacing the human capacity to execute it well.

On the target you are aiming at: traditional services firms top out around 30% margins. The bet is that AI operating leverage carries you toward software-like margins, north of 50%, on a market several times larger than software. You do not need to be there on day one. The line just has to be visibly bending, and you have to be honest with yourself about whether it actually is.

The traps

1. Mirage PMF. Strong revenue and happy logos that are actually powered by humans, not AI. The tells: flat or falling gross margin as you grow, revenue per head that is not improving, delivery headcount rising one-to-one with customers. If you cannot name the single number that captures how much of the work the AI is genuinely doing, assume you have this.

2. The early demand trap. When you launch, signing pilots is easy and intoxicating. Sign too many and you drown serving them with humans and never build the product to scale. We keep our live mandate load matched to what our senior capacity can actually execute well, rather than chasing every signature. Saying no here feels insane and is correct.

3. Bespoke creep. Every client asks for something slightly custom. Say yes too often and you are a custom shop, not a product. Until you have enough volume to know which requests are rules and which are exceptions, bias hard toward the standard process. This is exactly why we, and firms like Strala, said no to the broad ambition early and went narrow.

4. Hiring to mask product gaps. The most seductive trap for a founder under delivery pressure. Throwing a person at a broken step feels like progress and quietly becomes your business model. Automate the task, do not hire around the hole.

5. Intake outpacing senior capacity. The constraint that only appears once you are winning, which is exactly when it is most dangerous. The whole point of the autopilot model is that intake is cheap. The discipline is in not letting it run ahead of the humans who carry the judgement.

6. Undercutting on price. Straight-line undercutting signals cheap and low quality, and caps your upside permanently. Price on the value of the outcome, the way Panacea prices on a completed study rather than by the hour. We compete on speed, tension and execution, not on being the cheapest line on the page.

7. Buying your way in too early. I covered this as a founding path above, but it earns a place on the trap list too. Acquire a legacy firm before your own platform and culture are solid, and you will spend your energy fighting the culture you bought instead of building the one you wanted. You cannot acquire product-market fit. Buy later, as an accelerant, never as a way in.

How you actually win

Speed compounds. The firms that move first do not just get a head start… they build data flywheels, proprietary workflows and network effects that get stronger with every single deal. That is a far stickier moat than a legacy brand built on a model the market is leaving behind.

And you do not win alone. The whole professional services chain is accelerating at once… AI-native law firms like Lawhive compressing legal timelines, buyers running AI diligence, accountants automating prep. The fast firms will gravitate toward each other and form ecosystems that are quick end to end. When we run a deal now, the bottleneck is increasingly everyone else, and that is changing fast. Position yourself to be one of the fast nodes, because the slow ones are about to become the rate limiter nobody wants to work with.

The honest summary is this. The investor thesis is right. AI-native services will be one of the defining business models of this era. But the thesis is the easy part. The hard part is the operation, and in this model the operation is the product. Start where the work is already outsourced. Pick the slice where the old economics were broken. Build engineering DNA, not an engineering department. Be ruthless about what you do not do. Watch your senior capacity like a hawk. And be honest, every single month, about whether the AI is actually doing the work or just wearing the costume.

The future M&A firm runs a deal team of one super-human senior banker, one super-human junior, and AI. Not in ten years. Not in five. And the same will be true of the firm that closes your books, files your claims, and clears your contracts.

We are building it now. If you are operating in AI-native services, or seriously thinking about it, I would genuinely like to talk.