Enterprise AI · Venture building
Ziplabs explores where AI can change how enterprises work, builds the product and the system around it, and develops the strongest opportunities into new businesses.
Fig. 01The model is one component. The harness around it decides whether the system holds up, and the gold edge is the business it has to fit.
Conceptual
What we do
The gaps are rarely only technical. A product has to fit how people work, the system has to hold up in daily use and the economics have to justify the change. We bring product strategy, engineering and commercial judgment to the same problem, from the first customer conversation to the business case.
01
A customer problem with a real cost, and a reason AI changes what is possible.
02
Who it serves, what it replaces and why someone would change how they work.
03
Where the path stays fixed, where an agent decides and what it may touch.
04
Working software with evaluations, tracing and recovery that hold up in daily use.
05
Evidence from real workflows that the result is worth adopting.
06
Pricing, go-to-market, partnerships and the team the opportunity requires.
Perspective
The model matters. So does everything built around it, and the business it has to serve. These judgments shape what we pursue and how we build it.
The opportunity is a workflow with a real cost and people who would change how they do it. The model, and how much autonomy it gets, follow from that.
Related terms: Workflow redesign · Adoption
Where the path is known, a fixed workflow is usually easier to predict and test. An agent loop earns its place when the next step depends on what the last one found. A second agent needs its own reason, such as work that runs in parallel.
Related terms: Agentic workflows · Orchestration · Multi-agent
The software around the model keeps task state, chooses what each call sees and carries out tool calls. More context is not automatically better.
Related terms: Agent harness · Context engineering · Memory
MCP and A2A make tools, data and other agents easier to reach, and scopes can narrow what a connection may touch. Which business action an agent should take is still a decision about identity, authority and ownership.
Related terms: MCP · A2A · Delegated authority
Evaluations check the state a run leaves behind, on tasks that resemble the real work. Traces explain failures. Idempotent operations keep a retry from repeating an action that matters.
Related terms: Evaluations · Observability · Idempotency
Token prices are the visible cost. The number that matters includes retries, review time, latency and failure, and it moves as models and prices change.
Related terms: Unit economics · Model selection · Latency
AI usage can grow while the business stays the same. Value appears when a workflow is redesigned around what the system does well, people rely on it, and the result is worth what it costs to run. That is also what stays defensible when a better model is available to everyone.
Related terms: Adoption · Workflow redesign · Defensibility
Current work
Two areas are active now. Both sit where agents meet consequential enterprise work: problems with a real cost, where the next step calls for judgment and the result has to be trusted by the people accountable for it.
When an incident unfolds, the evidence arrives in pieces across alerts, logs, tickets and identity records. The hard part is turning those pieces into an understanding coherent enough to act on.
Direction
Investigation that is faster and better grounded, where conclusions stay tied to the evidence behind them and the decision stays with the people accountable for the response.
As agents begin to act across enterprise systems, organizations need to grant useful autonomy and still know what was done, on whose authority and with what result.
Direction
Autonomy that an organization sets deliberately, with an account of the work that keeps permission, action and outcome distinct.
These are the areas in motion today. The approach carries to other consequential enterprise work, and we are glad to hear about problems that fit it.
Insights
Essays on the product, technical and commercial decisions behind enterprise AI.
· Updated · 11 min read
Most useful systems combine fixed steps with steps where a model decides what to do next. The design work is choosing which is which, and building the harness that makes both dependable.
Read the essay· 8 min read
When a model does part of the work, the product is sold on outcomes, its costs rise with use, adoption depends on trust in the output and the advantage moves from features to the workflow around the model.
· 5 min read
Early in a venture, a measure is worth tracking when it would change the next commitment of time or money. Most familiar counts can wait until there is a business to measure.
Sajjad's career spans hands-on engineering in AI and machine learning, distributed systems and databases, and founder, CEO and CPO roles across startups and established organizations. At Ziplabs he brings product, technical and commercial judgment to the same problem.
About Ziplabs and its founderContact
We welcome conversations with enterprise leaders, founders, investors and potential partners working on problems where AI could change how the work gets done.