What is ContextOps?

The new hire test

Imagine handing a new hire one document on their first day and expecting them to do the job. They wouldn’t get far, because most of what makes someone good at their job was never written down. Yet that’s what most companies expect of their AI. We call it the new hire test: if a new hire couldn’t do the job from a document on day one, neither can AI.

Defining ContextOps

ContextOps is the discipline of turning operational evidence and expert judgment into tested, approved instructions, and maintaining them as the business changes.

Put simply: your company needs infrastructure for its people to teach its AI. ContextOps makes that teaching fast and measurable.

Think of it like DevOps for company knowledge. DevOps turned shipping software into a continuous discipline; ContextOps does the same for how a business runs. It’s never “done,” it’s operated. Like code, every instruction gets an owner, a version history, review and testing before it goes live, and a way to roll back if something breaks.

AI isn’t failing on intelligence. It’s failing on context.

When AI is missing context, it doesn’t flag it. A broken integration throws an error. An agent missing context just gives a confident wrong answer, and often nobody notices until a customer does.

Models are now highly advanced, and getting smarter every day. Yet most enterprise AI implementations still fall short, because they’re missing company-specific context: how your team handles a refund exception, which customers get escalated, which workaround actually fixes a recurring issue.

Why temporary fixes fall short

Companies usually try one of these fixes:

• Record retrieval. AI can search docs and tickets for similar cases, or follow links between past records and decisions. Either way, it only finds what’s already there: docs that are incomplete or out of date, and tickets that mix good calls with mistakes. Nothing turns them into a clear rule.
• Documentation projects. An SOP rewrite or knowledge base overhaul captures how the business worked at one moment, and starts going stale the day it’s finished.
• Fine-tuning. Training a model on last year’s tickets bakes in old habits, good and bad, and every change means retraining. A written instruction can be edited in minutes; a learned behavior can’t.

None of these keeps context accurate over time. Businesses change every day, with new products, policies and edge cases, and context that isn’t maintained decays.

How ContextOps works

ContextOps runs as a continuous loop with four steps.

The roles are clear. The people who do the work decide what the instructions say, AI does the heavy lifting of drafting and updating them, and engineering keeps the underlying systems running.

1. Observe

Start with how work gets done today: the tickets, emails, chats, call transcripts and docs a company already has, plus the processes people follow across their tools. No single person knows every step, but together they show how the work really happens. Once agents are live, you also observe how they perform. Every escalation, edit and miss shows where the instructions fall short. Not every miss is an instruction problem, though; sometimes it’s tone, a missing system permission or a customer preference.

2. Synthesize

Turn thousands of individual cases into draft instructions: the general rule, the exceptions, and where human judgment is still needed. This is the gap retrieval can’t close. Retrieval can only find what’s already there. Synthesis writes down what never was, and flags where past cases contradict each other.

3. Govern

The people who own the work stay in charge. Every proposed instruction comes with the cases behind it, is tested before it goes live, and takes seconds to approve. Owners can also change a policy directly and test it before it ships. This matters because history isn’t best practice. Past tickets mix mistakes with legitimate exceptions, and only the people who own the work can tell which is which.

4. Deploy

Put approved instructions to work inside the tools teams already use, like email drafts and chat, instead of asking anyone to switch. Then the loop starts again: what agents run into becomes new evidence to observe, and coverage grows with every cycle.

The ContextOps loop in practice

Say your support team refunds shipping on late orders for repeat customers, but that rule isn’t written down anywhere. ContextOps spots the pattern in past tickets and drafts it as an instruction. The support lead approves it with one exception, and the agent starts applying it. When new edge cases start coming up, the agent escalates them, and an update goes back to the lead for a quick yes or no.

Because context lives in one place, every fix pays off everywhere. The same approved instructions can power a support agent, a Slack Q&A bot, an email assistant or a ChatGPT integration, and they stay with the company as it switches models or adds agents. ContextOps matters most in recurring work, like support, IT incidents, claims and internal requests, where teams handle similar cases every day.

Measuring ContextOps: coverage

The core metric is coverage: the share of real cases AI can handle correctly using only approved instructions, with no one filling in missing judgment. It shows how well you’re running ContextOps today, and it’s the leading indicator of how much autonomy you can safely give your agents. Business results like deflections and fewer escalations follow from it. Every turn of the loop should push coverage up.

ContextOps vs. context engineering

Context engineering is the delivery system. ContextOps makes sure what it delivers is true.

You need both, but good engineering can’t fix wrong context. The same goes for most context tools, from context layers to observability: they move information to AI or watch what it does. ContextOps makes sure that information is right.

Context decides whether AI works

The model is no longer the bottleneck. Context is. The best way to start is narrow: point AI where your instructions are already solid, track coverage, and expand as it learns. That’s how AI passes the new hire test.

That’s why we built Edra: to give the people who know the work the fastest way to teach it to AI, and to keep teaching it as the business changes.

See ContextOps in action: book a demo.