Talk to us

Perspective October 2026

The SDLC was built for humans. AI changes the equation.

Most enterprises have added AI to one step of a process designed, end to end, around humans doing almost all of the work. Output rises. Delivery does not.

Get the full paper

What is inside

  1. Accelerating one stage does not accelerate the pipeline. It moves the constraint to a review gate still staffed for the era when people wrote every line.

  2. Three properties never get faster: correctness, conformance and continuity. Each one fails at a different point in the lifecycle.

  3. Machine error does not look like error. It is confident, fluent, and formatted exactly like success.

  4. Quality becomes an evaluation problem. Define the expected behavior and what counts as proof, then generate, score, refine and score again.

Murty Chennubhotla

Chief Architect, Scorpiosys

The observation

Is AI actually making delivery faster?

The productivity gain inside one activity is real, and it is large. What has not changed, in most enterprises, is everything around it. We have dropped a radically different engineering capability into a single step of a process designed, end to end, around humans doing almost all of the work.

  1. Discovery unchanged
  2. Design unchanged
  3. Develop measurably faster
  4. QA unchanged
  5. Release unchanged
  6. Support unchanged

Developer productivity < Delivery productivity

Accelerating one stage of a pipeline does not accelerate the pipeline. It moves the constraint. More code arrives at review, more generated output waits to be validated, more change queues for production.

The SDLC did not fail. It was engineered around an assumption that was true, that humans perform the work, and that assumption has just stopped holding.

The gap

Three things that do not get faster

Human error tends to be local and visible. Machine error is confident, fluent, and formatted exactly like success. When AI is added to a lifecycle built for human throughput, output rises and three properties stay exactly where they were.

Where correctness, conformance and continuity fail across the lifecycle
DiscoveryDesignDevelop and QASupport
Correctness Does it do what was intended? Intent lives in a deck. The business sees the result at UAT. No map of what already exists. Schema reinvented. Stubbed output reads as done. Structural checks pass it. Wrong data found by users. No ground truth to re-test.
Conformance Did it stay inside its boundaries? Scope agreed verbally. No risk or policy owner named. Policy sits outside the design. No acceptance criteria. The agent edits beyond the ask. Review was sized for humans. Actions undeclared. Nothing revocable without a deploy.
Continuity Does it hold, and inside a budget? No budget envelope. No owner for AI spend. Decisions never locked. No record to point back to. Re-architecture mid-sprint. Rework paid for twice. Spend seen on the invoice. Drift found by incident.
Where each property fails first: correctness at the very beginning, conformance in design, continuity after release.

These are not coding problems, which is why no coding tool addresses them. They are questions about intent, authority and evidence, and the lifecycle is where those are decided.

The answer became Intelligence Studio

If the lifecycle were designed today, what would it look like?

An AI-native delivery framework that brings human judgment and machine capability into one governed lifecycle: four phases, on a foundation established once and versioned as it evolves, rather than rebuilt project by project.

  1. 01 Figure Intent What the business actually means, captured as context an agent can act on.
  2. 02 Frame Decisions Architecture, data and policy settled and signed before generation starts.
  3. 03 Forge Build and evaluate Agents generate inside the boundary; output is scored against the signed criteria.
  4. 04 Field Operate Behavior, drift and spend watched after release, against the same evidence.

Foundation TechnicalAgenticComplianceEconomic

The last phase feeds the first. The lifecycle is a loop, not a line.

The control model

Machine speed, enterprise control

Intelligence Studio is not a collection of coding agents. It is a control model for how AI is allowed to participate in software delivery. Four elements carry it.

  1. 01

    Agent

    AI participates in the work: interpreting, synthesizing, evaluating and monitoring, not only writing code.

  2. 02

    Artifact

    Decisions of consequence become durable context that both people and agents work from.

    Without it, context is lost.

  3. 03

    Guardrail

    Architecture, data, security and governance define the boundaries of execution.

    Without it, the agent exceeds the ask.

  4. 04

    Human gate

    Consequential decisions stay owned and signed by an accountable person.

    Without it, nobody owns the outcome.

The objective is not autonomous software development. It is governed software development at AI speed.

Where this comes from

Built from enterprise delivery, not a laboratory experiment

  • AI-scored investor relationships

    Conviction Engine reads the conversations a firm is already having, extracts the signals and scores each relationship inside the firm’s own CRM. Nothing leaves their tenant.

  • AI-mapped legacy systems

    An agent read an undocumented codebase of 2,200 files and reconstructed its data model, the relationships between its parts, and 131 gaps nobody had written down, in under a minute.

  • AI-enabled patient journeys

    Automation through a care journey with enterprise controls, where a named person signs before anything irreversible happens.

None of these started with a model. Each started with a decision about what the system was allowed to do, who owned that decision, and what evidence would show it had held.

What would your SDLC look like if it were designed today?

Start with an AIDLC readiness conversation: one session, your delivery lifecycle, and an honest read on where AI would actually change the outcome.

Start with a 30-minute conversation Get the full paper