Customer Zero
Customer Zero: we run on what we sell
Before we run AI agents for clients, we run them on ourselves. Our finance, vendor, service and security workflows run on the same managed intelligence stack we deploy for customers, under the same rules.
Tekscape is its own Customer Zero. Our finance, vendor, service-desk, briefing, security and content workflows run on AI agents built on the managed intelligence stack we sell: hardware we own, a managed harness, a semantic model of our business, and cloud models where they fit better. Those agents work under the governance clients get: human approval, audit trails, fail-loud monitoring and change control in Git.
The idea
What Customer Zero means
Pax8's Managed Intelligence Provider Playbook uses the term "Customer Zero" for a simple idea: prove the model inside your own business first. We credit Pax8 for defining the Managed Intelligence Provider (MIP) category. This page shows how Tekscape applies the idea in its own operations. It describes our own practice, not a Pax8 designation.
No independent certification body or standard exists for the Managed Intelligence Provider label. What separates providers is whether they actually operate agents in production. So we show how we run our own company on agents, what rules we hold ourselves to, and what went wrong along the way.
The stack is the same one clients get. Agents run on hardware we own. A managed harness gives each agent its tools, memory, guardrails and logs. A semantic model of our business defines our terms, rules and metrics. For each task we choose between local and cloud models based on cost, data sensitivity and the quality the job needs.
Where agents work
The internal workflows we run on agents
We describe categories, not counts. We don't publish agent counts or internal dollar figures, and we won't until a figure is measured, approved and sourced.
| Workflow | What the agents do | Where a person decides |
|---|---|---|
| Finance close support and reconciliation | Pull ledger and billing data, match records, flag breaks and draft reconciliation notes | Finance reviews every break and posts every entry |
| Vendor invoice and contract reconciliation | Compare vendor invoices to signed agreements and billing records, and surface overcharges and gaps | A person approves every dispute or payment change |
| Service-desk triage and routing | Read incoming tickets, classify them, add context and route to the right queue | Engineers own the fix and the client conversation |
| Executive briefings | Assemble a daily brief from connected systems, with each figure traced to its source | Leadership decides what to act on |
| Security and configuration monitoring | Watch configuration drift, expiring credentials and failed jobs, and raise alerts | Engineers approve changes to production and client systems |
| Content QA gates | Check drafts against brand rules, sourced-claim rules and banned terms before review | A person approves anything that is published |
Governance
The rules we hold ourselves to
These are the same controls we put around client agents. If a rule is too slow for us, it is too slow for a client, and we fix the process, not the rule.
Human approval before anything goes outside
No agent emails a client, vendor or anyone outside the company, publishes content, or makes a financial commitment on its own. It drafts, a person reads the full message, and only an explicit approval releases it.
- Draft, show, approve, then send
- Financial, legal and client-facing actions need a person's approval
Audit trails
Agent runs leave a record: inputs, the sources read, what was produced and who approved it. Figures in a report trace back to a system of record, not to the model's memory.
- Timestamped logs for agent runs
- Numbers trace to their source
Fail-loud monitoring
A job that fails must say so. An empty result is never treated as a clean result. A failure is the first line of the report, and a stalled job trips a watchdog instead of going quiet.
- Failures surface first, not last
- A detector must prove it can fire before its silence counts
Change control in Git
Agent prompts, configuration, schedules and guardrails live in version control. If a change is not in Git, it did not happen. Changes can be reviewed and rolled back.
- Reviewable history for changes
- Rollback without guesswork
The stack
The same managed harness, pointed at our own business
Our internal agents sit inside the same harness we run for clients: tools, memory, guardrails, human review and audit logs around every agent. The semantic model underneath is ours, with our vendors, our billing streams and our definitions.
Model
Open-weight on your hardware, or a cloud model reached through routing.
Tools
The systems and APIs the agent may call.
Memory
Context it keeps between tasks.
Guardrails
What it must never do, enforced in code.
Human review
Approval before anything leaves your boundary.
Evals
Tests run before every change ships.
Observability + audit
A log of every action and why.
Cost routing
Your deployment by default, other models when cheaper.
Customer Zero tests the harness and the governance. It is not a benchmark, and it does not predict results in another business.
What we learned
What running on agents taught us
Much of what we now build for clients came from mistakes we made on ourselves first.
Agents need an ontology
Early agents were fluent and wrong. They mixed up entities with similar names, applied one billing rule to every stream, and used different definitions for the same metric. What helped was writing down how our business works: entities, rules, metrics and where each fact lives. That semantic model is now the first thing we build for a client.
Verification beats confidence
A confident answer is not a correct one. A job that exits cleanly can still hand back stale or placeholder data. We now check outputs against an independent source, test that every check can actually fail, and treat an unverified result as unverified, even when it looks right.
Start narrow
The agents that stuck did one job with a clear owner and a clear pass line. Broad "do everything" agents were hard to test and harder to trust. We start clients the same way: one workflow, measured, then expand.
Honest limits
Why this page has no numbers
We don't publish agent counts, hours saved or dollar figures from our own operations. Internal figures are hard to verify from outside and easy to inflate. When we publish a Customer Zero figure, it will be measured, approved and listed with its source. Until then, we would rather show you how the system works than quote a number you can't check.
Frequently asked questions
Sources
- Pax8, The Managed Intelligence Provider PlaybookSource of the "Customer Zero" idea: prove the model inside your own business first.
- Pax8, The Rise of the Managed Intelligence ProviderPax8 defines the Managed Intelligence Provider category (blog, October 2025).
- CIT Solutions, What is a Managed Intelligence ProviderNo independent certification body or standard exists for the label.
See the stack we run on
Walk through the harness, the governance and one workflow that fits your business, with the people who run it for Tekscape.
Talk to us