Request for Startups 2027: What We Fund
We are a pre-seed fund that invests in B2B software. For a decade that was a good place to stand, and the economics were well understood. Build something embedded and necessary, charge per seat, keep net revenue retention above 100%, exit at a multiple of ARR.
Parts of that model are under real pressure, and I do not think all of it holds for the next five years. That is not the same as saying software stops being a good business. It means what makes a software business durable is changing, and a lot of investors are still underwriting as though it has not.
One thing has not changed, and it is the part we actually compete on. We do not fund categories, we fund founders. We reach them earlier than most funds do, and we spend that time teaching the work rather than only assessing it. Any firm can publish a list of attractive markets. What is harder is knowing whether a specific founder can find the customer, kill their own assumptions quickly, and build a business efficiently once they have.
So this request for startups starts with what we think is happening, then what we think survives, then the list of startups we want to fund (skip to the list here). The thesis is the useful part. The list is downstream of it.
What is actually threatening software
Four things, and only one of them gets discussed much.
The build cost collapsed. If a competent team can rebuild your product in a weekend, the code was never the moat. That was always true in principle. It is now true in practice, and it removes protection that a large share of vertical SaaS was quietly relying on.
Foundation models absorb generic capability. Anything a model does natively gets absorbed into the model. Thin wrappers are one release away from worthless, and everybody knows this, which has not stopped them being funded.
Seats compress. This is the one nobody wants to say out loud. Charge $100 per employee, help your customer do the same work with half the people, and you have halved your own revenue by delivering exactly what you promised. A company can win on every product metric and shrink on the only one that matters. Greylock’s Saam Motamedi has made the sharpest version of this argument: agents break SaaS economics, and force a move from tools priced per seat to agents priced per outcome.
Incumbents ship AI down existing distribution. An incumbent does not have to be better, only adequate and already installed, and that is a real disadvantage to start from. It is not decisive. Incumbents are reliably mediocre at deep, specialized workflows, and plenty of work is specialized enough that adequate does not displace anybody. So the question is not whether you sit beside an incumbent. It is whether the customer could pull you out without breaking how they operate.
Together that is not an argument against software. It is an argument against many of the software models investors have historically rewarded.
What we think survives
Three things a model cannot absorb, and a fourth about how you charge.
Data the workflow itself creates. Not “we have data”. Data that exists because your product ran, that nobody outside the workflow can observe, and that makes the product better the longer it operates. If somebody with a crawler and a budget could reproduce it, it is not an advantage.
Accountability. In healthcare, banking, law and compliance, a wrong answer carries legal consequences, and a model cannot be the accountable party. Companies that take on that liability, with auditability and a defensible standard of accuracy, are protected by the thing that makes them slow to build.
Integration nobody wants to rebuild. Connections into legacy systems, physical equipment, regulators and humans in the loop. Unglamorous, specific, and the reason an incumbent’s AI feature does not actually replace you.
Revenue that scales with work, not headcount. If your pricing is a tax on your customer’s staffing, your own product is the thing that shrinks it. We are most interested in companies that charge for output delivered, and most worried about per-seat pricing on software that eliminates seats.
What that difference looks like in practice:
| Traditional SaaS | AI-era workflow software |
|---|---|
| Helps a person do a job | Completes much of the job |
| Competes for software budget | Can compete for labor or services budget |
| Often charges per seat | Can charge per task, transaction, or outcome |
| Measures productivity | Can measure completed work and business results |
| Automates individual activities | Can coordinate entire workflows |
Where this changes our underwriting
Our stated thesis is embedded, necessary B2B workflow software, with real customer pain, an existing budget, and a capital-efficient path to return. I still believe all of it. Three parts need sharpening, and none of them get dropped.
“Existing budget” was read too narrowly. The test is right and it stays: somebody has to already be paying to get this problem handled. What we read too narrowly is which budget counts. We looked for software budget, which is the smaller and more contested pot. The best AI companies we see are often paid out of labor and outsourced services instead: contractors, offshore teams, headcount already committed. That is an existing budget by any definition, and a far larger one. So this widens the aperture rather than loosening the discipline. We will still ask which line item the money comes out of, and we will be more interested, not less, when the honest answer is payroll, because nobody replaces people on the strength of a demo.
“Embedded and necessary” needs a second question: embedded in what. Embedded across a workflow spanning several parties, where nobody owns the whole chain, is the strongest version, because there is no single incumbent in a position to close it. Embedded inside a platform somebody else owns is harder, but it is not disqualifying, and we have been too quick to treat it that way. Starting inside an incumbent’s ecosystem is often the fastest route to the customer. What decides it is what happens next: whether the company becomes something the customer cannot operate without, or stays a feature the platform owner can ship themselves.
“Traction” is a weaker signal at pre-seed than it was. AI made demos and pilots close to free, and early usage spikes with them. We will weight renewal and expansion inside the same buyer well above logo count, and discount anything whose evidence is a signup number.
One part of the thesis gets stronger. Capital efficiency was always our argument for not needing unicorns, and falling build costs make it more achievable at the application layer, not less. A $20M fund taking meaningful ownership in companies that reach eight figures and exit without a decade of dilution is a better plan in 2027 than it was in 2022.
Validate it before you pitch it ToolSuite walks an idea through customer discovery and scores what comes back. 30 free credits, no card. Try it free →We fund founders, not categories
The list below is the easy part. On its own it is worth very little, because picking a category is not the same as having a company.
What we underwrite is the founder. Specifically: whether they find out what the customer needs rather than assume it, kill their own assumptions quickly instead of defending them, reach revenue without spending a round to learn something a week of conversations would have taught them, and build a business rather than a product.
That matters more on this list than it would on most. A workflow spanning four organizations means running customer discovery in four places. Selling against payroll means persuading somebody to stop paying people. Taking on accountability in a regulated workflow means being right in a way a demo never has to be. None of these are won by shipping faster.
It is also not a judgment you can make across a pitch meeting, so we try not to make it that way. We have put more than 1,000 ventures through the bootcamp since 2015 and built the same method into ToolSuite. That gives us a sourcing advantage, but sourcing is not mainly what it is for. It is where founders get better at the part that decides everything else: the interviews they have not run, the assumption they have not tested, the business model they have not priced.
About two thirds of the portfolio has been through the bootcamp, not as a route in but because the founders we are seriously considering get invited to spend three days with us. We would rather watch somebody run customer discovery for three days than read their deck twice, and they leave further along than they arrived.
Plenty of those still end in a pass. Finding that out before the money moves is the entire point.
Startups We Want to Fund
Filtered through all of the above, and shorter than most lists on purpose. Last year’s list ran to six broad themes and no thesis; this one is the other way round. If you want ideas rather than a filter, that is a different list.
Workflows that span organizations
Nobody owns the whole chain, which is why no incumbent can simply close the gap.
-
Vertical AI coordinating multiple parties. The valuable workflows are rarely one person doing one task. They are a contractor, an insurer, an inspector and a homeowner waiting on each other. Moving the chain is worth multiples of speeding up one link.
-
Procurement and supplier onboarding. Risk review, approvals and source-to-pay across departments that do not share a system.
-
Permitting, inspections and licensing. Local government workflow is slow, universally disliked, and sits on real budget. YC is pointed here too, which we read as confirmation rather than competition.
-
Eligibility and benefits verification. Infrastructure that checks income and entitlement across fragmented programs. Deeply unglamorous, structurally defensible.
Work that used to be labor
This is where the budget actually is.
-
Voice agents running inbound operations. Not voice as an input method. Voice handling the call, the scheduling and the dispatch for businesses that currently pay people to do it.
-
AI workforces for field and home services. The back office of the trades, where the alternative is hiring and the comparison is a wage, not a software license.
-
Generalists doing specialist work. AI that lets someone without the credential perform the task competently, under supervision. The clearest example of software priced against labor.
-
Productizing what is currently consulting. Services businesses that industrialize delivery into a repeatable product. Unfashionable, capital efficient, and often a better business than the venture-backed alternative.
Workflows where being wrong has consequences
A model can be confidently wrong. In these workflows somebody still has to sign.
-
Compliance infrastructure. Consolidating fragmented tools, giving finance and risk teams real-time visibility, reducing the specialist headcount needed to stay legal.
-
Identity risk and fraud operations. For banks and fintechs, where the accuracy standard is set by a regulator rather than a product manager.
-
Clinical documentation inside the record. Integrated into the EHR workflow rather than sitting beside it. The integration is the moat.
-
AI for air-gapped environments. Defense, finance and healthcare running frontier capability inside secure, disconnected systems.
The layer underneath the agents
Less fashionable than the agents themselves, and much harder to pull out once it is installed.
-
Governing unstructured enterprise data. Lineage, permissions and quality where decades of documents meet a model. Not another vector database.
-
Proprietary models on proprietary data. Helping companies own their intelligence rather than rent it.
-
Security alert triage and response. Products that close the loop rather than summarize it, with an explicit answer on what they will never do without a human.
-
Infrastructure for agent-scale workloads. Rate limiting, observability, cost control and failure handling, rebuilt for a caller that never gets tired.
Industries nobody writes about
The least crowded part of this list, and usually the part with the oldest incumbents.
-
Applied AI in construction, agriculture, logistics and the trades. Specific problems, old incumbents, and founders who usually already worked in the industry.
-
Industrial data capture. Enormous quantities of machine data generated and discarded. Capturing it is step one. Knowing what it is worth is the business.
What to send us
An idea from this list does not get you funded. Eighteen founders could pick the same number and seventeen of them would be guessing.
What gets our attention is the founder who has already done the unglamorous part: talked to the people with the problem, found the specific version worth solving, and got somebody to commit money or time before the product existed. Ten honest customer conversations beat a polished deck, every time.
When you pitch, tell us three things. The problem and who has it. Which budget line the money comes out of. And what would prove you wrong.
If you are earlier than that, still deciding whether the idea holds, that is a shorter piece of work. Start here.
Startup Ignition Ventures
$20M pre-seed fund. First checks $100K to $1M. Real feedback in hours.
Pitch SIV →