“Enterprise architecture just takes too long.” It’s a complaint every architect has heard. Current-state discovery, capability maps, application inventories, gap analysis, roadmaps: producing a full set of deliverables can easily take six months or more. By the time the work is finished, the business has moved on and half of it is already out of date.
Generative AI can change that, but not if you simply ask it to “build our EA.” What makes the difference is a repeatable pattern: you keep the discipline of the TOGAF ADM and hand off the heavy lifting in each phase to AI.
This article covers the core principles of using AI for enterprise architecture, a sample 12-week delivery plan, and ready-to-use prompts for each phase of the ADM.
Why AI for Enterprise Architecture, and Why Now
Expectations of the EA function are shifting fast. Gartner predicts that by 2028, AI agents will help deliver 50% of enterprise architecture outcomes (Gartner Predicts 2026 via Ardoq). It also warns that by 2029, 90% of EA deliverables will need to be built for AI executing agentic workflows, not just for human delivery teams (Nile).
That means EA teams now have two jobs at once:
- Use AI to accelerate EA work: spend less time producing documents and running analysis, and more time on decisions
- Build EA that AI can consume: create a structured EA repository that AI can reference and act on
These aren’t separate efforts. Once you start using AI to speed up the work, your architecture naturally takes on a structure AI can read.
Core Principles: Four Rules for AI-Assisted EA
1. AI Drafts, Architects Review and Decide
Change how deliverables get made: AI produces the first draft, and people review it and make the call. Let AI handle dependency mapping and technical debt analysis so architects can focus on decisions and stakeholder alignment. AI does the legwork. Judgment and consensus-building stay with people.
2. Define the Metamodel AI Will Consume, First
This is the single most important rule. In the Preliminary Phase, define your element types and their relationships: capabilities, applications, data, technology, projects, and so on. Without that foundation, every AI output comes back in a different shape, and you lose the time you saved stitching the pieces together.
3. Start Small and Expand Iteratively
The ADM is meant to be run iteratively, not as a one-way waterfall, and both TOGAF and ArchiMate should be tailored to your organization’s maturity: start simple, then expand (Visual Paradigm). Resisting the urge to cover the whole enterprise up front is the key to showing results quickly.
4. Focus on a Theme That Delivers Value
Pick one theme tied directly to a business priority. Examples include global ERP consolidation, end-to-end visibility from procurement to payment, or shortening the cycle from engineering change to production changeover.
A 12-Week Plan for Rapid EA Delivery
| Weeks | ADM Phase | What AI Does | What People Do |
| 1–2 | Preliminary / A | Drafts EA principles and the architecture vision from strategy documents; proposes a metamodel from the CMDB and master data definitions | Set scope; align with the sponsor |
| 3–5 | B / C / D | Extracts the capability map, application inventory, and data flows from process manuals and system lists | Validate facts through interviews; standardize granularity |
| 6–7 | Gap Analysis | Identifies gaps, redundant applications, and technical debt; compares options | Decide the target state; set priorities |
| 8–10 | E / F | Drafts candidate initiatives, transition architectures, the roadmap, and the WBS | Make investment decisions; finalize the business case |
| 11–12 | G / H | Builds a mechanism to pre-check projects against EA principles | Launch the Architecture Review Board (ARB) |
In one reported case of AI-assisted ADM, manual drafting time fell by as much as 70–80% (Visual Paradigm). Because the figure comes from the tool vendor, treat it as a rough benchmark rather than a guarantee.
How to Use Phase-by-Phase TOGAF ADM Prompts
The prompts below are ready to use. Replace anything in { } with your own organization’s details.
The System Prompt to Start With
Before running any phase-specific prompt, set the stage with the following system prompt. It makes a noticeable difference to output quality.
text
You are a senior enterprise architect with deep expertise in the TOGAF Standard and ArchiMate 3.x.
Assume strong domain knowledge of {industry: automotive Tier 1 / manufacturing}
and {core systems: SAP S/4HANA, PLM, MES}.
Follow these rules:
1. Separate facts stated in the provided materials from your own inferences, and tag inferences as [INFERRED].
2. Cite the source document and location for every item. If there is no source, write “No source.”
3. If information is missing, do not fill the gap with guesses. List it at the end as “Questions to confirm.”
4. Strictly follow the specified table format.
5. End with 3–5 “Decision points for human review.”
Two things matter most here: forcing a clear line between fact and inference, and requiring sources. Those two rules alone cut review time significantly.
Preliminary Phase: Define the Metamodel
text
Analyze the attached CMDB export, system inventory, and master data definitions,
and propose a metamodel for our EA repository.
Requirements:
– Keep element types to a minimum (10–12 types)
– Define mandatory and optional attributes for each element type
– Define relationships between elements using ArchiMate relationship types
– Show in a table how existing CMDB fields map to the metamodel
Phase B: Build the Business Capability Map
text
Using the attached process manuals, segregation-of-duties policy, and org chart,
create a business capability map.
Requirements:
– Three levels (L1: 10–12 capabilities; L2: 3–7 per L1)
– Do not include organization or system names
– Reference standards: APQC PCF, {industry reference model}
– Tag capabilities not found in the materials but commonly required as [INFERRED]
Output: | L1 | L2 | L3 | Definition | Owning Organization | Source |
Keeping organization and system names out of capabilities is essential. If you include them, you’ll be rebuilding the map every time there’s a reorg or a system replacement.
Phase C: Assess Applications with TIME Analysis
text
Perform a TIME analysis on the application inventory
(Tolerate / Invest / Migrate / Eliminate).
Criteria: business fit (1–5), technical health (1–5), cost efficiency
Output: | App ID | Business Fit | Technical Health | TIME Category | Rationale | Target Candidate |
Finally, list groups of applications with overlapping functionality
and estimate the benefit of consolidating them.
Gap Analysis: Compare Architecture Options
text
Compare the options for “{decision: e.g., a single global instance vs. regional instances}.”
Criteria: alignment with EA principles, process standardization, cost, risk,
implementation time, scalability, regulatory compliance
Explain the rationale behind each score, give a recommendation,
and finish with “Conditions that would reverse this decision.”
Asking for the conditions that would reverse the decision is the trick here. It prepares you for the question executives always ask: “What happens if that assumption doesn’t hold?”
Phase G: Assess Custom Extensions Against Clean Core
text
Assess the list of custom development and extension requests using Clean Core principles.
Assessment levels:
1. Can be met with standard functionality (change the process)
2. Can be met through configuration
3. Use an approved extension pattern (key user extensibility, side-by-side on SAP BTP)
4. Requires core modification (submit as an exception)
Output: | Request ID | Description | Business Reason | Assessment | Rationale | Alternative |
The Full Prompt Library at a Glance
These are the prompts worth keeping on hand for each phase:
| Phase | Key Prompts |
| Preliminary | EA principles, maturity assessment, metamodel, governance structure and ARB operating rules |
| A: Architecture Vision | Stakeholder map, goals and KPIs, vision document, Statement of Architecture Work |
| B: Business Architecture | Capability map, heat map, value streams, current-state process |
| C: Information Systems | Application inventory, TIME analysis, capability-to-application mapping, interfaces, master data governance |
| D: Technology Architecture | Technology standards, target technology architecture, AI inventory |
| Gap Analysis | Gap table, option comparison, technical debt and dependencies |
| E: Opportunities & Solutions | Candidate initiatives, prioritization, business case outline |
| F: Migration Planning | Transition architectures, roadmap and WBS, migration risks |
| G: Implementation Governance | EA compliance review, Clean Core assessment, Architecture Contract |
| H: Change Management | Change impact analysis, technology trend impact assessment |
| Cross-Phase | Requirements extraction, decisions and action items from meeting minutes, quality review |
Risks of Using AI in EA, and How to Manage Them
Research has found that generative AI in enterprise architecture carries risks such as opacity and bias, contextually incorrect outputs that lead to rework, privacy and compliance concerns, and social loafing, where people lean too heavily on the tool (arXiv). Here’s how to address them:
| Risk | Mitigation |
| Plausible but wrong output | Always require sources; have a person verify anything tagged [INFERRED] |
| Output that misses the context | Provide the system prompt and internal documents every time; build on industry reference models |
| Leakage of confidential information | Use only company-approved AI environments; mask names and figures when using external services |
| Over-reliance on AI | Final decisions rest with business owners and the ARB; record every decision |
One more point is easy to overlook: AI itself needs to be managed as part of the architecture. EA teams are extending the Technology Architecture domain (Phase D) and their repository metamodel to capture model provenance, training data lineage, deployment context, and risk and compliance controls (Orbus Software).
Practical Tips for Making It Work
Run the prompts in sequence
Work through them in order: metamodel, capability map, application inventory, then the mapping between them. Each step improves the accuracy of the next.
Link everything with IDs
Assign IDs to each output and feed them into the next prompt as input. Once capabilities and applications are linked by ID, you can hand off impact analysis to AI as well.
Take output as tables and refine in Excel
Have AI return everything in table format, refine it in Excel, and then load it into your EA tool. That becomes the source data for your repository.
Set a standard unit of work
For each deliverable, use a simple rhythm: AI draft (30 minutes), human review (1–2 hours), and one round of stakeholder validation. It makes planning far more predictable.
Get a second opinion from a separate chat
Open a new chat and ask AI to act as a tough reviewer or to argue the opposite position. It surfaces contradictions and gaps you’re likely to miss on your own.
Summary
AI can help you deliver enterprise architecture faster and more efficiently. But the results depend less on the AI model and more on having the right pattern for using it.
- AI drafts; architects review and decide
- Define the metamodel first so AI output stays consistent
- Focus on a single theme and complete your first ADM cycle in about 12 weeks
- Use a system prompt that always separates fact from inference and requires sources
- Keep final decisions with people, and manage AI itself as part of your architecture
Start with one theme: have AI draft your metamodel and capability map. You’ll quickly see how much it changes the way you deliver EA.
References
- Gartner Predicts 2026: Enterprise Architecture Enables Resilient AI-Powered Business Value (Ardoq)
- Gartner’s AI Prediction for Enterprise Architecture (Nile)
- TOGAF ADM Guide-Through: A Case Study on AI-Powered TOGAF ADM Implementation (Visual Paradigm)
- From Vision to Value: Our Hands-On Journey with AI-Enhanced ArchiMate Modeling (Visual Paradigm)
- Impact and Implications of Generative AI for Enterprise Architecture (arXiv)
- TOGAF Enterprise Architecture: The Complete Guide (Orbus Software)
Related article: AI × TOGAF® ADM: From Descriptive EA to Prescriptive EA
Disclaimer
Parts of this article were developed with reference to generative AI suggestions and were reviewed, refined, and supplemented based on the author’s professional expertise and judgment.
Leave a Reply