Every venture I build at Gigaverse starts from the same observation, and it is not really about AI. It is that an enormous amount of what gets sold as a service is a system with people wedged into the gaps. Compliance filings. Patient scheduling. Catalogue hygiene. Someone is being paid by the hour to move information between systems that could talk to each other, and to apply judgement at maybe four points in a forty-step process.
Close the gaps and the service stops being a service. That is the whole thesis, and the market has now spent a couple of years testing it in public.
The arithmetic that reframes the opportunity
Foundation Capital put the shift most usefully in 2024: AI companies are moving from software-as-a-service to service-as-software, inverting the essence of SaaS. In software, the vendor sells access to a tool and the customer remains responsible for the outcome; in services, responsibility for the outcome sits with the seller. Instead of QuickBooks, you sell tax work done by an AI accountant.1
The arithmetic is what makes it more than a positioning trick. Salesforce — one of the most successful software companies ever built — generates around $35 billion in annual revenue. Global companies spend roughly $1.1 trillion a year on sales and marketing salaries alone.1 Add software engineering, security and HR and you are at about $2.3 trillion in global salaries; add enterprise IT and business-process services and the addressable figure reaches $4.6 trillion.1 A year later the same team put it more bluntly: the prize is not the familiar $200 billion SaaS pool, it is the $4.6 trillion enterprises pour into salaries and outsourced services.2
Other estimates converge on the same order of magnitude by different routes. HFS Research, whose Phil Fersht coined the Services-as-Software label, projects roughly $1.5 trillion by 2035 as software and services converge.3 J.P. Morgan Private Bank frames it as a $3–5 trillion opportunity across business services.3 Madrona reaches the same conclusion from the demand side: the biggest opportunity for AI is targeting labour directly.4
The demand signal is not hypothetical either. HFS research with Publicis Sapient found two-thirds of enterprises planning to replace significant human-led services with AI-led alternatives within three years.3 HCLSoftware's 2026 survey of 173 enterprise leaders has 76% prioritising agents and autonomous systems, with 81% reporting live or pilot initiatives — and governance identified as the missing link for a quarter of them.5
Features stopped being a moat
Here is the part that surprised people who came from SaaS. In the old model you differentiated on product: architecture, UX craft, feature velocity. That advantage has largely evaporated, because AI accelerated what Foundation Capital calls the collapse of software primitives. Decades of enterprise UX and domain logic compressed into a handful of modular capabilities — inbox triage, structured tables, LLM search, workflow routing, generative response — and what used to take months of application-specific design can be assembled in a weekend.2
The consequence is that nearly every AI product now looks the same. A legal ops system resembles a claims processor; a RevOps copilot resembles an underwriting assistant.2 When everyone can ship the same primitives on the same models, what you build stops being the moat and how you integrate, embed and operate becomes the moat.
“In enterprise AI, integration is not a post-sale activity. It is the product surface.”2
If that is true — and everything I have seen operating says it is — then the org chart of a services-as-software company looks nothing like a SaaS org chart. The centre of gravity moves to the people who sit inside the customer's mess.
The forward-deployed engineer stops being a cost centre
In the pre-AI era, forward-deployed engineers were implementation consultants: technical enough to configure, personable enough to survive a go-live. They are now among the most strategic assets in an enterprise AI company, and the reason is that customer environments are irreducibly weird.2
The published examples are instructive because they are so specific. Sierra embeds an engineer inside a Fortune 500 support team for two weeks, shadowing agents across Zendesk, Slack and an in-house escalation spreadsheet — surfacing a hidden rule that VIP customers jump the queue when their last CSAT fell below four, a rule that would have broken the bot's triage logic had it stayed invisible.2 Harvey puts forward-deployed legal engineers inside Am Law 100 firms for weeks to codify how redlines are handled and how decisions escalate, and those implementations become part of its deployment framework.2
What makes this a business rather than a treadmill is that the work compounds. Edge-case rules get encoded as runtime-editable parameters rather than hard-wired logic, so a category manager can move a competitive-RFQ threshold from $10,000 to $7,500 in an admin panel and the sourcing agent adapts on its next run, with the change logged for audit.2 Patterns abstract into modules: a change-data-capture pipeline built for one healthcare client became an ingestion kit offered to every prospect on legacy relational databases, cutting future deployment time by 70%.2 Telemetry closes the loop — a legal team tracking clause-extraction accuracy fed each false negative into a nightly fine-tune and lifted recall on indemnification clauses from 92% to 98% inside a month.2
This is the part I have found hardest to explain to people who want a clean software P&L. The first deployment in a vertical is expensive and looks like consulting. The fifth is fast and looks like software. Whether you are building a business or a boutique depends entirely on whether you engineered for that transfer.
The last mile is where the margin goes
The reliability bar in enterprise services is nothing like the bar for a consumer chatbot. A legal assistant flagging compliance risk needs 99%-plus accuracy, not helpfulness.2 And that gap between plausible and dependable is where budgets die. Foundation Capital's read on 2026 states it cleanly: you can reach 80% with 20% of the effort — enough to close a pilot — but production demands 99% or more, and that last stretch can take 100 times more work.6
That reliability bar rewrote the go-to-market motion too. Because AI performance depends on data quality, workflow integration and domain tuning, customers cannot meaningfully evaluate these systems outside their own environment — so they now expect to experience functionality, integration and outcome before signing.2 Clean-sample demos fail on real purchase orders full of legacy formatting and industry jargon.2
The result is a cost-of-sale crisis. AI proofs of concept require data ingestion, orchestration, prompt tuning and live validation; the cost of evaluating a bad-fit customer is measured in headcount hours rather than product clicks.2 And it does not end at go-live — the old deploy-and-maintain model becomes deploy-and-continuously-re-engineer as customers add product lines and workflows.2 Churn therefore costs more than lost revenue: you forfeit weeks of implementation work, and modest churn erodes margin once engineering effort and non-trivial token costs are counted.2
Sensible teams have responded by qualifying harder rather than selling harder. One healthcare company sorts deals into "below the line" — standard workflows, templated integration — and "above the line," which triggers gating and manual technical scoping.2 That kind of triage is unglamorous and it is the difference between a services-as-software company and an agency with a model in the loop.
Pricing follows responsibility
If you have taken on the outcome, seat pricing stops making sense. Foundation Capital describes a spectrum rather than a switch: access-based, then usage-based, then workflow-based, then outcome-based.2
The middle of that spectrum is where most real businesses land. Harvey straddles it: firms pay roughly a thousand dollars per lawyer per year, but renewal conversations turn on hours saved rather than seats deployed.2 Pure outcome pricing breaks down when results depend more on the customer's business than on your software — which is exactly why AI SDR tools, despite promising measurable sales outcomes, mostly price on tasks or usage.2 Usage pricing has its own trap: voice platforms billing per minute spoken find that every optimisation cutting call length also cuts revenue.2
What this means if you are building one
The version of this I would give a founder in three sentences. Pick a vertical where the service is mostly a system with people in the gaps, because that is where the work compresses. Expect the first deployment to look like consulting and engineer it so the fifth does not. And do not sell an outcome you cannot instrument, because the moment you take responsibility for a result, measurement stops being reporting and becomes the product.
The structural reason I think this favours studios over single companies is that the expensive parts — the ingestion kits, the telemetry, the deployment scaffolding, the senior team that has done it before — are the parts that transfer between verticals. The domain knowledge does not transfer. The machinery does. That is the bet behind running Gigabit, GigaCommerce, Top Dentistry and FormBridge off one senior team rather than four.
One honest caveat. The market has grown sceptical for good reason: customers now demand proof on real data before commitment, having watched a lot of transformative promises fail to survive contact with their own purchase orders.2 Foundation Capital's warning against "vibe revenue" is the right one — early sign-ups are easy, and durable growth comes only from speed-to-value and renewals that grow because the software keeps doing more of the work, more reliably, each quarter.2
The incumbents are, ironically, helping. As Salesforce, ServiceNow and Microsoft push their own agentic offerings they legitimise the category and make buyers more willing to bet on startups who move faster.6 The category is no longer the hard part. The last mile still is.