Section 1
What leading means when the ground moves every quarter
In a mature category, the leader is the company with the largest installed base and the most complete product. In this one, the installed base is young and the product boundary keeps moving, so market share is a lagging and fairly weak signal. Better signals are structural. Who controls a scarce input, such as compute, distribution, or a body of proprietary operational data. Who sits inside the system where the work already happens, which is where automation gets adopted rather than admired. Who has a reason to still be here in five years, meaning a business model that works at today's prices rather than one that assumes costs keep falling. Anything that leads on demonstrations rather than on deployments is a marketing position, not a market position.
Section 2
Four categories, and what each one is optimizing for
Model providers optimize for capability and for developers building on top of them. They set the ceiling of what is possible and change it without asking you. Workflow and orchestration platforms optimize for breadth of connectors and speed of assembly. They win when a business needs many small automations quickly, and they charge in a way that scales with your volume. Incumbent software vendors embedding AI into the CRM, the ledger, or the helpdesk optimize for retention. Their advantage is that the data and the users are already there. Their weakness is pace, and a tendency to price the AI features as an upgrade tier. Integrators and consultancies optimize for billable delivery. They are genuinely useful when the constraint is process design rather than technology, and expensive when they are hired to choose tools for you. Knowing which one you are talking to explains most of what they will recommend.
Section 3
How to evaluate a leader you plan to depend on
Innovation you can read about is less important than the answers to five procurement questions: what happens to your data, what the price does at ten times your volume, how you would export and rebuild elsewhere, what the deprecation history looks like, and who is accountable when the output is wrong. The model below turns those into a scorable comparison. Then pressure-test it with [How to Evaluate AI Vendors and Partners](/blog/how-to-evaluate-ai-vendors-and-partners).
Section 4
Borrowing from leaders without copying them
The practices that travel from a large, well-resourced company to a small one are the cheap ones. Keeping an evaluation set of examples with known correct answers. Logging every automated decision. Putting a human review step in front of anything customer-facing. Retiring automations that stop earning their keep. The practices that do not travel are the expensive ones: dedicated platform teams, custom model training, elaborate internal tooling. Copying those is how a fifteen-person company acquires a fifty-person company's overheads. Read case studies for the sequence and the failure modes, not the destination. The destination was reached with a budget you do not have.
Section 5
Concentration risk, and where the vendor makes money
Depending on one provider is normal and often correct. Depending on one provider without knowing your exit cost is the mistake. Before you build a workflow that matters, establish where the data lives, whether you can export it in a usable form, and how long a rebuild would take on a different stack. Be specific about the commercial model. Usage-based pricing means your success raises your bill. Per-seat pricing means growth raises it. Free tiers that carry your data are paid for somewhere. None of this is sinister, but a leader's incentives should be visible to you before you are locked into them. The operator communities are usually more honest about this than any analyst note, which is the case for [AI Automation Communities and Learning Resources](/blog/ai-automation-communities-and-learning-resources).
Section 6
Signals that separate a leader from a loud vendor
Named customers running in production rather than in pilot. Published limitations, and a support channel that acknowledges them. Version and deprecation notes you can actually find. Pricing that survives a volume projection. An honest answer to the question of what the product does badly. And one negative signal worth taking seriously: a vendor who cannot describe a customer they would decline. Someone selling to everyone has not decided what they are for, which means their roadmap will be pulled around by whoever pays the most next quarter.