How Endava builds an agentic organization with Codex

Enterprise software vendors have spent the last eighteen months racing to embed AI agents into their platforms. The pitch is familiar: automate workflows, accelerate delivery, reduce headcount. What they leave out is that most organizations attempting this shift will fail—not because the technology does not work, but because they are building agents into structures designed for human bottlenecks.

This piece is for the VP of IT or COO who has been asked to evaluate “agentic AI” for their organization and needs to understand what actually changes when you deploy AI that acts, not just suggests.

What Agentic Actually Means

The term gets thrown around loosely, so let us be precise. An agentic system is AI that takes multi-step actions toward a goal with limited human intervention between steps. It is not a chatbot that answers questions. It is not a copilot that suggests code. It is software that reads a requirements document, identifies gaps, queries stakeholders for clarification, and produces a specification—while you are in another meeting.

A recent case study from a global software consultancy describes reducing requirements analysis from weeks to hours using this approach. That number is real, but it obscures the harder question: what happens to the three weeks of human review, negotiation, and organizational alignment that used to happen alongside that analysis?

The answer, in most failed implementations, is that those three weeks simply shift somewhere else—usually to the validation and rework phase after the agent delivers output that was technically correct but organizationally naive.

Where the Real Cost Hides

Vendor proposals for agentic AI typically include licensing, integration, and a pilot timeline. What they omit:

• Process redesign labor—someone has to map every decision point where humans currently intervene and determine which ones the agent can handle, which require escalation triggers, and which need to be eliminated entirely
• Trust calibration—teams need 60–90 days to learn when to override agent output and when to let it run, a period during which productivity often drops below baseline
• Exception handling infrastructure—agents work well on the 80% case but require human fallback systems for the 20%, systems that did not exist before because the human handled everything
• Output validation tooling—when an agent produces a 40-page requirements document in two hours, you need new processes to verify it, or you inherit errors at scale

In most engagements we see, these hidden costs add 40–60% to the total investment beyond the vendor quote. Organizations that budget only for the software and a pilot discover this in month four, when they are too committed to turn back but too underfunded to finish properly.

The Two Paths Forward

Organizations adopting agentic AI are splitting into two camps, and the results diverge sharply.

The Automation Path

The first camp treats agents as faster humans. They map existing workflows, identify slow steps, and deploy agents to accelerate those steps. This is the path most vendors sell because it is easy to scope and easy to demo.

The problem: you inherit every inefficiency baked into the original process. If your requirements process took three weeks because of organizational politics, approval bottlenecks, and unclear ownership, an agent that produces requirements in two hours does not solve the problem. It just moves the bottleneck downstream, where validation and approval now take three weeks instead.

Typical outcome on this path: 20–30% efficiency gain in the automated step, but minimal impact on end-to-end cycle time. Six months in, executives ask why the expensive AI initiative has not moved the needle.

The Redesign Path

The second camp uses agent deployment as a forcing function to redesign workflows entirely. They ask: if this step can now happen in hours instead of weeks, what else changes? Who needs to be involved differently? What approvals can we eliminate? What handoffs become unnecessary?

This path is harder to sell because it requires organizational change, not just software deployment. But it produces different results. Organizations on this path report 50–70% reductions in end-to-end cycle time for specific workflows—not because the agent is faster, but because the agent’s speed exposed how much of the original process was waiting, not working.

What to Assess Before You Start

Before evaluating any agentic AI solution, answer these questions honestly:

• Can you document the current process in enough detail for an agent to follow it? If no, you are not ready—you are asking AI to automate something you do not fully understand
• Who currently makes judgment calls in this workflow, and are they willing to cede those decisions to an agent? If no, you will spend your budget on change management, not technology
• What is your tolerance for agent errors at the volume the agent will produce? A human makes one mistake per day; an agent might make one mistake per hundred outputs, but it produces a thousand outputs per day—so your error volume goes up, not down
• Do you have the infrastructure to validate agent output, or will you be trusting it by default within 90 days because checking it is too slow?

Organizations that answer these questions before selecting a vendor save months of wasted pilots. Organizations that skip them discover the answers the hard way.

The 12-Month Outlook

The agentic AI market is moving fast, but the implementation challenges are not changing. Over the next year, expect three things:

First, vendor consolidation. The standalone agent startups will get acquired or struggle against platform vendors (Microsoft, Salesforce, ServiceNow) bundling agent capabilities into existing subscriptions. This is good for pricing, bad for innovation.

Second, a backlash cycle. Organizations that deployed agents without process redesign will start reporting disappointment. The “AI failed us” narrative will pick up in trade press, even though the failure was implementation, not technology. Use this window to learn from others’ mistakes cheaply.

Third, a capabilities gap. The organizations that invested in the redesign path in 2024 will pull ahead in operational efficiency by late 2025. The gap will be visible in delivery timelines and cost structures. Late movers will have more mature technology to deploy, but less organizational readiness to absorb it.

The organizations succeeding with agentic AI are not the ones with the best technology or the biggest budgets. They are the ones willing to ask an uncomfortable question: if this process can now happen in hours, should it happen the same way at all? The agent is not the strategy. The redesign is the strategy. The agent just makes the redesign non-optional.

Start with one workflow where you have executive authority to change the process, not just accelerate it. Prove the model there. Then scale.