The Services-to-Software Flip: How Agencies Are Becoming Agent Companies

Table of Contents

• What the Services-to-Software Flip Actually Means
• Why Now: The Conditions That Make the Flip Possible
• The Unfair Advantage Agencies Have Over Pure Startups
• The Three Migration Paths
• Path 1: The Internal Tool That Escapes
• Path 2: The Hybrid "Services-Wrapped Software"
• Path 3: The Clean-Sheet Spinout
• The Economics: Why the Margins Flip Before the Revenue Does
• What Breaks: The Org Chart, the Comp Plan, the Identity
• How to Tell If Your Agency Can Actually Make the Jump
• Insights Most People Overlook
• References

What the Services-to-Software Flip Actually Means

Strip away the buzzwords and the flip is simple. An agency sells human hours. A software company sells a product that runs without those hours. For thirty years the wall between the two was thick, because automating professional services meant building software that could reliably do judgment work, and that software didn't exist.

Now it sort of does. Agentic AI changed the math. An agent that can read a contract, draft a response, reconcile a ledger, or run an outbound sequence is no longer a research demo. So the question every services firm is quietly asking is: what if the workflow we've been renting out by the hour could be sold as a thing that runs itself?

That's the flip. You take the part of your business that is a repeatable process dressed up as bespoke expertise, you encode it into an agent, and you change the unit you sell. Instead of "we'll assign three analysts at $180 an hour," it becomes "we'll review every contract you send us for $40 a document" or "we'll book the meeting and you pay when it shows up." The work product looks similar. The business underneath it is a different species.

This sits squarely in the vertical-agents conversation, because the flip almost never works horizontally. Agencies don't have a generic edge. They have a specific one: they know how prior-authorization denials actually get overturned, or how a Series B SaaS company actually books outbound meetings, or how a restaurant group actually does its weekly close. That domain specificity is the whole asset, and it's why the vertical-agent moat built on proprietary workflow data matters more here than raw model quality.

Why Now: The Conditions That Make the Flip Possible

People have been predicting the death of the billable hour since the first offshore BPO. It didn't die. So why is this moment different?

Three things converged.

First, the agents got good enough at the specific, narrow tasks that make up the bulk of agency work. Not good enough to replace a senior partner's judgment, but good enough to do the 70% of the work that was always grunt labor masquerading as expertise. a16z has written extensively about how AI is collapsing the line between software and services, and the core insight is that the addressable market isn't the software budget anymore. It's the labor budget, which is roughly ten times larger.

Second, buyers changed. A CFO who would never have bought "AI legal review software" in 2022 will happily buy "contract review, billed per document, with a human backstop" in 2026, because the framing is a service outcome, not a software gamble. The flip works partly because it lets you sell software to people who only know how to buy services.

Third, pricing models matured. Per-seat SaaS pricing made no sense for agentic work, because the value isn't access to a tool, it's completed work. The market has largely settled on per-task and per-outcome pricing, which happens to map perfectly onto how agencies already think. They already priced by deliverable. They just had to swap the labor input for an agent.

The Unfair Advantage Agencies Have Over Pure Startups

Here's the part venture-backed founders hate to admit: the agency that's been grinding in a vertical for eight years often has a better shot at building the winning vertical agent than the ex-FAANG team with a fresh seed round.

The reason is data and distribution, and agencies have both.

On data: a startup building a clinical-documentation agent has to go find examples of good clinical documentation, edge cases, denial patterns, and the weird local conventions of how things actually get done. An agency that's been doing medical billing for a decade has all of that sitting in its case files. That's the proprietary workflow data that becomes the real moat for a vertical agent. It is genuinely hard to replicate and it's already paid for.

On distribution: the agency has customers. It has signed contracts, an account manager who picks up the phone, and trust that took years to build. When it ships an agent, it doesn't start at zero. It upsells a base that already pays it money. A startup spends 18 months and most of its seed round trying to manufacture exactly this.

The catch is that these advantages are wasting assets. The proprietary data edge shrinks every quarter as horizontal model providers and well-funded startups close the gap. The distribution edge only matters if you move before a horizontal platform eats the vertical you sit in. Agencies are usually slow, and slowness is the one thing this transition does not forgive.

The Three Migration Paths

There's no single way to do the flip. In practice, firms take one of three routes, and the route mostly determines whether they end up with a real software company or an expensive internal tool.

Path 1: The Internal Tool That Escapes

The most common and most accidental path. An agency builds an agent to make its own delivery cheaper. The motivation is margin, not product. The ops team automates the boring 70% so the same headcount can serve more clients.

This is a great way to fatten your services margins and a terrible way to build a software company by default, because internal tools are built for your own quirks, not for a market. But occasionally a client sees it, asks to use it directly, and the firm realizes the tool is the product. The flip happens almost against the firm's will. The danger is that you wake up two years in with a tool deeply wired to your internal processes and impossible to sell to anyone who doesn't already work like you.

Path 2: The Hybrid "Services-Wrapped Software"

The pragmatic middle. The firm ships an agent but keeps humans visibly in the loop, often calling it "managed" or "expert-verified." Customers buy an outcome. Behind the curtain, an agent does most of the work and a human reviews the edge cases.

This is where most successful flips actually live for the first few years, and it's underrated. It lets you charge software-like prices, capture the trust premium of human accountability, and quietly raise the percentage of work the agent handles over time. The human-in-the-loop layer is also where you handle the agent reliability and liability problems that pure-software competitors gloss over, which matters enormously in regulated verticals. The risk is that you never fully flip. The services gravity is strong, and "we'll just keep a human on it" becomes a permanent crutch that caps your margins and your valuation multiple.

Path 3: The Clean-Sheet Spinout

The hardest and highest-upside path. The firm spins up a separate entity, a separate team, often a separate brand, and builds the agent product as a real company with the agency as its first customer and data partner.

This is the only path that reliably produces a venture-scale software business, because it escapes the cultural and financial gravity of the services parent. It's also the path most agencies can't stomach, because it requires cannibalizing your own billable revenue and explaining to your partners why you're funding a thing designed to make their hourly work obsolete. The spinouts that work usually have a founder who came up through delivery, understands the domain in their bones, and is willing to be the bad guy who automates their own colleagues.

The Economics: Why the Margins Flip Before the Revenue Does

The seductive math is the gross margin curve. A services firm runs at 30 to 50% gross margins because labor is the cost of goods. A SaaS company runs at 70 to 85% because the marginal cost of serving one more customer is nearly zero. The dream is to swap the former for the latter.

What actually happens in the early years is messier and worth being honest about. Agent-delivered services don't have zero marginal cost. Every task burns tokens, and inference is a real, variable, per-unit cost in a way that classic SaaS hosting never was. McKinsey's analysis of the economic potential of generative AI is bullish on the productivity gains, but the unit economics of a per-task agent business look more like a hybrid of software and a usage-metered utility than pure SaaS.

So the realistic trajectory is: your margins improve immediately on the work you automate, because an agent doing 70% of a task is cheaper than a human doing 100%. But you don't hit true software margins until inference costs fall, your agent handles a higher share of tasks unsupervised, and you stop needing a human reviewer on every output. That last condition is the hard one, and it's tied directly to how you price industry-specific value capture, because if you've priced per outcome, every reliability improvement drops straight to your bottom line.

The valuation argument is real, though. Revenue that looks like software trades at a multiple of revenue that looks like services. A flip that's even partially credible can re-rate a business in the eyes of acquirers and investors. That re-rating, more than the operating margins, is what's driving a lot of the flips you're seeing announced.

What Breaks: The Org Chart, the Comp Plan, the Identity

The technical work is the easy part. The thing that kills most flips is human.

Start with the comp plan. Agencies pay people, and often partners, on billable hours or utilization. An agent that does the work in a tenth of the time torpedoes utilization. You are asking your best people to enthusiastically build the thing that shrinks their bonus. That's not a tooling problem. It's an incentive problem, and if you don't redesign comp before you build, the org will quietly strangle the agent in code review.

Then the org chart. A software company needs product managers, ML engineers, and a different sales motion. An agency is organized around delivery pods and account management. These are different animals with different rhythms, and bolting one onto the other usually produces a software team that's perpetually starved of resources because the services side, which actually pays the bills today, always wins the budget fight.

Finally, identity. "We're a firm of experts" and "we're a software company that's automating those experts" are contradictory self-images, and the people who joined for the first one rarely want the second. Harvard Business Review's work on why companies struggle to escape the services trap keeps surfacing the same lesson: strategic transitions fail not on the strategy but on the operating model and the people, and the services-to-software flip is one of the most violent operating-model changes a firm can attempt.

How to Tell If Your Agency Can Actually Make the Jump

Not every agency should attempt this, and the honest screening questions are unglamorous.

Is your work genuinely repeatable, or is the repeatability a story you tell clients? If every engagement is truly bespoke, there's no workflow to encode and the flip is a fantasy. The firms that succeed usually deliver the same core process over and over with cosmetic variation.

Do you own your workflow data, contractually and technically? If your client agreements give the client all rights to everything you produce, your supposed data moat belongs to someone else. Check the contracts before you check the model.

Can you survive cannibalizing your own revenue? The flip means selling fewer hours on purpose. If your firm can't take a near-term revenue hit to build a more valuable long-term business, you'll flinch at exactly the wrong moment.

And the quietest question: are you early enough? The window to flip a given vertical is open while domain data is scarce and before a horizontal platform or a well-funded vertical startup commoditizes it. If you're reading think-pieces about your vertical being disrupted, the easy window may already be closing, and you're now racing rather than strolling.

Insights Most People Overlook

The flip is often a defensive move dressed as offense. Most coverage frames agencies becoming agent companies as ambitious empire-building. In reality, a lot of these flips are agencies trying to disrupt themselves before a startup does it to them. If your core service is automatable, someone will automate it. Doing it yourself is the only way to capture the upside instead of just absorbing the downside. The aggressive-sounding flip is frequently a controlled retreat.

Your best margin opportunity isn't selling the agent, it's keeping the savings. Many agencies rush to externalize their agent into a product because product revenue sounds sexy. But the quietly enormous play is to deploy the agent internally, keep your prices flat, and pocket the widened margin. You don't have to become a software company to benefit from the flip. Plenty of firms will make more money never selling the agent at all and simply doing the same work for half the cost. The flip-to-product path is glamorous; the flip-to-margin path is more reliable.

The human-in-the-loop layer is a moat, not a transitional embarrassment. Software-first competitors treat human review as a cost to engineer away as fast as possible. Agencies that flip have something those competitors can't easily buy: a trained workforce that already knows the edge cases, plus the accountability customers in regulated verticals demand. Far from being a weakness, "we have humans who catch what the agent misses" is exactly the trust differentiator that wins healthcare, legal, and financial buyers. The agencies that win may be the ones who keep humans in the loop longest, not the ones who remove them fastest.

Whoever controls the data-labeling pipeline controls the company. In an agency-turned-agent-company, the people generating and correcting the agent's training and feedback data are your former delivery staff. Their corrections are the asset that improves the agent. If you don't formalize that feedback loop, making the reviewers' edits flow back into the system, you're burning your single most valuable, hardest-to-replicate input. The firms that win treat every human correction as labeled training data, not as a one-off fix.

The valuation re-rate can happen faster than the operational reality, and that's a trap. A firm can announce a flip, get investor and acquirer attention, and re-rate its multiple long before the agent actually does most of the work. That gap is dangerous. If you raise or sell on a software story your operations haven't earned yet, you've written a check the org has to cash under pressure, and the services gravity that pulls every flip back toward billable hours is brutal when you've already promised the market you've escaped it.

References

• a16z, AI Will Bring the Services Economy Into Software
• McKinsey, The Economic Potential of Generative AI: The Next Productivity Frontier
• Harvard Business Review, What It Takes to Make Strategy Work