The argument in brief
Purohit's starting point is that frontier models keep improving, infrastructure keeps getting better, and AI coding tools have cut the time it takes to build software. She says a feature that once took a team months can sometimes be copied in days. Her conclusion is that the same forces helping founders build are helping their rivals.
She separates two questions founders tend to blur:
- What helps you get ahead? Speed. A startup can re-architect around a new AI capability quickly. A large company has existing customers, migrations, roadmap commitments, security and legal reviews, and internal processes to work through.
- What helps you stay ahead? Speed fades once everyone sees the same opportunity. In her words, if your edge is shipping first, assume someone ships second.
She also says she increasingly hears early-stage investors ask for a "12-month moat". That is her anecdote. It is not a survey result.
Three places defensibility can compound
Purohit sees these three areas across the B2B AI startups she works with.
1. Workflow
A thin AI feature can be useful and still be easy to replace. Her argument is that a vendor becomes harder to remove once it is part of how the work gets done. She points to the resurgence of forward-deployed engineering, where engineers sit with customers and map the real process. The questions she lists:
- What happens when a request arrives?
- Who approves it?
- Which system holds the authoritative data?
- Where do exceptions occur?
- Which decisions are written down, and which live only in someone's head?
- What happens when the AI is wrong?
The test she offers is whether each implementation stays bespoke forever, or whether it teaches the product how a whole category of customers operates. If it's the latter, she calls implementation "product discovery at unusually high resolution" rather than services work. As her example she uses Maven AGI, which she says started in customer support and expanded across channels and teams as it embedded deeper.
2. Proprietary data
Purohit is cautious about "data moat" claims. Her warnings:
- Public information doesn't become proprietary because it sits in a vector database.
- A customer uploading documents doesn't automatically create defensibility.
- Having a lot of data isn't a moat.
What may count is operational data: messy records, historical decisions, human corrections, exceptions, internal terminology, industry rules and tacit knowledge. Usage can also generate signals, such as a user correcting an output, an operator approving or rejecting a recommendation, or a workflow failing. She says those can improve retrieval, evaluation, routing, workflows, context and model performance.
She then adds four tests for any feedback loop:
- Does the data materially improve the product?
- Do you have the rights and architecture to use it appropriately?
- Does the advantage compound fast enough to matter?
- Could a competitor reproduce it?
Her example here is dSilo, an agent-native procurement platform. She describes it connecting ERP, procurement, contract, finance and legacy systems into a shared context layer for domain-specific agents.
3. Distribution
Her case is that AI has lowered the cost of building software but not the cost of earning trust. Enterprise buyers still ask who introduced a vendor, how it fits their stack, whether security teams trust it, and whether the vendor will still exist in three years. So she recommends entering through ecosystems that already hold customer relationships, procurement paths and credibility.
Her Microsoft-specific advice is to use Microsoft Marketplace as a procurement and deployment channel. The idea is that customers can buy using their existing Azure spend commitments, and startups can unlock co-selling with Microsoft's sales teams.
The flywheel
Purohit links the three into a loop:
- Embedding in the workflow gives you data and feedback competitors lack.
- That makes the system perform better.
- Better outcomes build trust.
- Trust and partnerships create distribution into the next customers.
She adds that AI capability is advancing faster than enterprises can absorb it, because they don't reorganize workflows, data, security frameworks, governance and operating models with each model release. She argues that this gap is where application-layer category leaders will be built.
What the essay leaves out: co-sell has conditions
The essay says Marketplace and co-selling are benefits of the program. Microsoft Learn's documentation on Azure IP Co-sell Acceleration is more specific. It was last updated May 7, 2026, so check the live page before acting. It describes Azure IP Co-sell as one of Microsoft's top go-to-market benefits for startups and partners building on Azure. Its stated effect is that the solution counts toward customers' Azure spend commitments, so a Marketplace purchase can draw down budget the customer already has to spend. The page also says a startup that achieves the status can submit co-sell opportunities, and Microsoft sellers then work alongside it.
Reaching that status through the acceleration route requires all of the following:
- Active Microsoft for Startups membership.
- At least $8,000 in monthly Azure consumption for three consecutive months. Paid usage or Azure credit consumption can count.
- A live, transactable SaaS offer in Microsoft Marketplace that has passed Microsoft's SaaS technical validation.
- At least three customers in the past 12 months, shown by three referenceable stories or three Marketplace purchases.
- A completed Co-sell Solutions page in Partner Center, including a one-pager and pitch deck.
- A qualified pipeline: either co-sell referrals totaling $1M or more in estimated contract value, or one qualified opportunity worth $500K or more in Marketplace-billed sales.
The 120-day clock
The designation comes with a 120-day validation period. To keep it, a startup must reach either $100,000 in Marketplace billed sales or $100,000 in paid Azure consumption over the trailing 12 months. If neither happens, the designation is removed until the standard co-sell requirements are met. The acceleration benefit can be used only once.
Microsoft advises activating it when deals are progressing and expected to close soon, because the window is time-bound. Review of a submitted request may take 7 to 10 business days.
The documentation says the startup team monitors participants that have a transactable offer and meet the consumption threshold, and contacts them. Startups that think they qualify but haven't heard anything are told to contact their Startup Advisor or file a support ticket.
Does co-sell actually help?
Microsoft's own Tech Community post claims co-sell deals are about 30% larger and close up to twice as fast. That is Microsoft's number, from its own post, with no independent verification. The post also says Microsoft sellers are incentivized to prioritize Azure IP co-sell eligible solutions. Third-party partner guides describe the same mechanics, including seller quota credit. Several of those guides are published by companies that sell Marketplace onboarding services, so read them with that in mind.
A founder's checklist
The essay's questions work as a self-audit. The data-rights and architecture question may matter most to IT and security reviewers evaluating AI vendors.
- Workflow: Does the product sit in a critical process, or beside it? Do lessons from one deployment carry over to the next customer?
- Data: Do corrections and approvals measurably improve results? Do you have the rights to use them, and is the architecture built for that?
- Distribution: Can a Microsoft seller or implementation partner easily bring you into an account? Does each customer make the next one cheaper to win?
Caveats
- This is advocacy. The author works for the company whose Marketplace and co-sell programs she recommends. The essay includes no independent study, quantified comparison or market forecast.
- Examples are illustrative. Maven and dSilo are described by Microsoft. Their official sites show the scope of their products, but that doesn't prove their moats are durable. The "few hundred thousand interactions" figure is a hypothetical, not a measured threshold.
- Credits have terms. The page advertises up to $150,000 in startup credits without eligibility details on that page. Don't assume every applicant receives that amount.
- Counterpoint. A model advantage can still matter in some domains, and a Marketplace listing doesn't guarantee discovery, sales or trust. The essay doesn't claim otherwise, but its tone could leave that impression.
Bottom line
The essay's thesis is that customer position matters more than model access in enterprise AI. That is a strategic argument and not a proven law. The practical value for Azure-based ISVs is in the co-sell conditions: the $8,000 monthly consumption floor, the transactable SaaS offer, the three customers, the pipeline thresholds, and the 120-day, $100,000 proof point. Anyone planning around Microsoft's channel should read those before betting the go-to-market plan on it.