AI Engineer Code 2025
AI Copilots for Tech Architecture: The Highest-ROI Use Case You’re Not Building — Boris Bogatin and Toufic Boubez, Catio
Read the talk
AI Copilots for Tech Architecture: The Highest-ROI Use Case You’re Not Building
A live system model, traceable recommendations, and guidance inside developer workflows can connect architectural decisions to the business outcomes faster coding is meant to serve.
From a talk by Boris Bogatin and Toufic Boubez
Before you start: Familiarity with software architecture, cloud services, and developer workflows will help; no experience building AI agents is required.
What happens when faster coding follows the wrong architecture?
During Toufic Boubez’s time at Splunk, Boris Bogatin recalls developers dismissing the possibility that coding copilots could meaningfully supplement their work. Now those tools are part of everyday development. Project management, implementation, and operations already have substantial tooling, including platforms such as Splunk and Datadog. That leaves a question upstream of implementation: does producing more code help if the architectural direction is wrong?
Architecture determines which investments faster development compounds. A sound direction can advance business objectives; a poor one can accelerate rework and technical debt. Boris frames architectural decisions as drivers of nine-figure spending. His case for an architecture copilot begins with the leverage of those decisions, rather than a measured comparison of copilot returns.
Yet the tools supporting that work often remain spreadsheets, tribal knowledge, and gut instinct. CTOs and architects have traditionally supplied the judgment, but more decisions are moving into developers’ hands as organizations shift responsibility left. The opportunity is to equip that distributed decision-making with something more dependable than informal knowledge.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Establish a shared current reality
Toufic groups the recurring problems into three challenges:
- Visibility: As the technology estate grows, teams lose track of what exists and how its parts connect.
- Defensible priorities: Leaders need evidence for where to invest, what to prioritize, and how that work contributes to ROI.
- Scalable guidance: Developers receiving more decision-making authority also need access to architectural expertise.
The missing foundation is a dependable, live map of services, dependencies, and drift. Without that baseline, decisions become slow and defensive, spending becomes redundant, and risks remain poorly understood. A shared architectural reality should play a role analogous to shared code: something teams can consult together instead of reconciling competing opinions about the system.
Toufic describes organizations making multimillion-dollar bets without knowing what they already own. A static diagram cannot resolve that problem for long. The map must update as the system evolves, preserving the connection between a proposed direction and the deployment from which the organization will actually start.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make priorities explainable
Visibility does not decide which project deserves scarce resources. Each team can argue that its project is essential. The useful question is what to do next given the organization’s constraints, existing investments, and strategic goals. Recommendations must expose their consequences for cost, performance, risk, and time to value.
Boris’s conversations at CTO dinners and through Architecture Deconstructed bring this down to a practical concern: leaders are not always asking for technical perfection. They need the business to understand why architectural work matters, allocate budget to it, and see how it advances business needs rather than sending teams in unrelated directions.
An actionable recommendation therefore needs more than a suggested change:
| Question | Required explanation |
|---|---|
| Why is this valid? | Reasoning grounded in the system and its constraints |
| Where did it come from? | Traceable supporting evidence |
| What should it change? | Expected impact on relevant objectives |
| How will we know? | Measurable outcomes |
Together, those explanations support a roadmap whose initiatives are scored for impact, with ROI, business objectives, and best practices considered explicitly.
Only then does accelerating implementation have a clear target. Boris connects justified priorities to the value of additional coding output; Toufic calls the reverse sequence “ready, fire, aim.” The architecture decision supplies the direction that makes implementation speed useful.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Guide developers without becoming the bottleneck
Shifting decisions left does not automatically scale architectural expertise. Developers already make architectural choices, while architecture guilds and enterprise architecture teams struggle to review them all. The challenge is to provide useful guidance without forcing every decision through a scarce reviewer.
Periodic strategy presentations do not close that gap. Boris describes teams presenting their standards every two weeks and getting little engagement. Developers are trying to ship features against immediate business requirements; translating those features into the architecture team’s preferred baseline is another task competing for their attention.
The proposed alternative is ongoing conversational guidance that produces designs suited to a developer’s actual requirements, with policy and architectural guidance already in context. It belongs inside the developer’s workflow. Autonomy needs alignment, but alignment cannot depend entirely on gates. Unconstrained autonomy risks inconsistency; mandatory waiting undermines the productivity that delegation was meant to create. The intended result is compliant, strategically aligned designs without making developers wait for every answer.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Stacks: model the deployment and its objectives
The first pillar of the proposed architecture copilot is stacks, the live visibility layer. It ingests data across clouds, Kubernetes services, and logging platforms, then models dependencies and changes over time. The resulting digital twin represents the deployed architecture: what the organization has, rather than what its wiki says it has.
Deployment data supplies only part of the context. A copilot also needs curated business objectives, requirements, standards, and strategy. Those inputs must connect company priorities to workspace and team objectives so that recommendations fit the particular organization and the people doing the work. A technically plausible recommendation is not enough if it optimizes for the wrong destination.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Reason across interconnected decisions
The recommendation layer has to reason across an interconnected system. Toufic treats this as distributed problem-solving: separate the dependencies and parts of the problem, examine them, then combine the findings into a recommendation that retains the global context. A collection of locally sensible changes is insufficient if their interactions undermine the overall design.
Multi-agent systems offer one way to organize that work. Agents focus on different parts of the problem and collaborate toward a solution. In the approach described here, they rely on LLMs for architectural knowledge. Toufic characterizes that knowledge expansively—as though the models had read practically every architecture book and best practice—but does not identify a model or establish its training coverage. The useful premise is access to broad architectural knowledge, combined with the specific system context.
His proposed evolution moves from language models toward specialized architectural models, then toward system-behavior modeling and simulation. That would let teams explore scenarios and inspect their likely effects before committing to a decision. Toufic explicitly places this simulation capability in the future, rather than presenting it as an available feature.
The organizational analogy is a design review. Boris sees value in the human process already used by architecture teams; the limitation is how much data and computation it can bring to bear. Collaborating agents would extend that process with more context and continuous availability, rather than discard the idea of reviewing a design from multiple perspectives.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Go beyond isolated best practices
The desired output is an explainable, ROI-ranked recommendation set that understands both the stack and its objectives. It should expose trade-offs across cost, performance, risk, and time so that leaders can prioritize the whole roadmap.
Boris distinguishes two levels of advice:
| Example | Architectural scope |
|---|---|
Migrate Amazon EBS gp2 volumes to gp3 | Apply a comparatively straightforward infrastructure best practice |
| Streamline a data pipeline for reuse across applications | Understand shared dependencies and redesign how multiple applications use the pipeline |
The second example requires knowledge of the overall architecture. Its potential value comes from changing a shared pattern, not merely improving an individual component. Boris presents this as the kind of improvement customers expect to matter most; he supplies no quantified savings result for either example.
Traceability makes those recommendations usable beyond the engineering team. A proposed improvement tied to expected impact and ROI gives leaders something concrete to discuss with executives or a board, including why the change deserves investment.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Bring architectural guidance into feature design
The third pillar is a conversational architectural agent. It gives developers, architects, and other users a way to ask questions about their architecture and obtain advice on optimization and refactoring. The same system knowledge supporting portfolio decisions becomes accessible during everyday work.
The proposed next step is design generation from requirements such as a product requirements document, or PRD. The agent would combine those requirements with the architecture team’s governance, controls, and guidance. The intended behavior is to generate designs that follow the guidance from the outset; supplying guidance alone does not establish guaranteed compliance.
That connects estate-level visibility and strategic roadmaps to the developers who ultimately shape the estate. Instead of depending primarily on a guild review every two weeks, architectural advice becomes proactive. Leadership imperatives shape the AI’s context, training, and narrative, while each answer addresses the developer’s particular situation. Architecture teams scale their influence through the guidance they give the AI.
The intended change in the reviewer’s role follows from that arrangement. If routine standards conformance is addressed during design, architects can spend less time correcting familiar mismatches and more time on strategy and difficult problems alongside development teams. The goal is to increase what the architecture function can accomplish, not simply move its existing review queue into a chat interface.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Close the loop with measured outcomes
Toufic assembles the pillars into four steps:
- Ingest and understand. Normalize data from messy systems into a live digital twin that users can inspect and navigate.
- Align and advise. Combine goals, requirements, and company context—including industry and growth stage—to produce ranked recommendations with projected impact on the metrics that matter to that company.
- Guide inside the workflow. Generate designs, answer what-if questions, and incorporate standards without requiring developers to leave their normal workflow for a separate advisory process.
- Track and improve. Record decisions, verify outcomes, and use what happened to improve subsequent advice.
The last step distinguishes projected benefit from achieved benefit. Recommendations start with expectations about cost, performance, or ROI; tracking outcomes is how the organization determines whether those expectations were justified.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connect architectural direction to coding execution
The next connection is between architectural advice and implementation. Returning to the aiming analogy, Boris places the architecture copilot before the coding copilot: decide the direction, then accelerate the work. Toufic imagines architecture agents communicating with coding agents while people provide directives, monitor their behavior, and correct course. He frames that communication as a coming capability in the talk.
This would make architecture a decision hub within the software development lifecycle, spanning how companies plan, build, and evolve their technology estate before executing software changes. The intended benefits are organizational clarity, shorter decision cycles, roadmaps tied to business impact, and expertise available where developers make decisions. Architecture and coding copilots become complementary strategic tools: one helps choose the work, and the other helps carry it out.
Boris predicts that organizations slow to adopt these capabilities will accumulate legacy systems and debt. Toufic grounds their enthusiasm for coding copilots in Catio’s own use, which he says the team has discussed in LinkedIn articles and blog posts. That is an account of their internal experience, separate from demonstrating the return on an architecture copilot.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Start with one portfolio area and one team
The adoption advice is deliberately narrower than the full vision:
- Choose one portfolio area. Establish visibility and build a digital twin for that bounded part of the estate.
- Generate recommendations for specific outcomes. Tie the proposed changes to business needs in that area.
- Pilot autonomous guidance with one team. Learn how the advice fits actual development work before introducing it throughout the company.
- Expand after proving ROI. Scale gradually toward the broader decision hub once the initial adoption has demonstrated value.
Toufic expects skepticism from architects, CTOs, and developers. The response is to establish evidence in a small scope before asking the rest of the organization to change how it works.
Boris closes by framing adoption as a question of timing—early or late—and offers to explore what a copilot would look like on a listener’s own stack. That discussion can support an independent implementation or work with Catio. The practical starting point remains a bounded architecture problem with an outcome the organization can verify.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Catio’s current architecture decision platform, with product information and routes to getting started.
AWS’s historical explanation of gp3 pricing, independent performance provisioning, and migration from gp2.
Further reading
Toufic Boubez explains specialist architecture agents, context retrieval, orchestration, and recommendation traceability.
Catio’s account of AI-assisted engineering across planning, integrations, UI development, and code review.
Updates since the talk
A July 2026 announcement connecting architecture reasoning and blueprints to coding assistants through MCP.
Instructions for connecting a coding assistant to a Catio workspace, including authentication requirements.
Full transcript
Loading transcript…