AI solutions architect role: RFPs, designs and delivery trade-offs
In this article (10 sections)
- The architect's real output is a decision someone can fund
- The artefact: an RFP compliance matrix
- Options, not a single answer
- How the week differs by employer
- Win rates, delivery surprises and design reviews
- Presales heads, delivery leads and client CTO offices
- Confused with solutions engineer and enterprise architect
- RFP, reference architecture and hyperscaler signals
- Shadow three bids, then own one design
- Principal architect, practice head or CTO-office roles
Related programme: the FDE for Professionals course (16 weeks (proposed)). See also open Forward Deployed Engineer jobs in India, updated daily.
An AI solutions architect designs how an AI system will fit into a client's or company's existing technology, data and constraints, and makes the trade-offs explicit: which models and services to use, where data lives, how it is secured, what it will cost and what can go wrong. In Indian IT services firms much of the role sits in pre-sales, answering RFPs and shaping deals; in product companies and GCCs it is closer to technical leadership on delivery.
The architect's real output is a decision someone can fund
Diagrams are the visible part. The substance is a set of decisions with reasons, costs and risks attached, written so that a client's CIO, a delivery manager and an engineer can each find what they need. An architect who produces a beautiful diagram without a cost estimate or a list of assumptions has not finished the job.
The artefact: an RFP compliance matrix
Services architects spend a lot of time responding to requests for proposal. The compliance matrix is the document evaluators read first. Here is an excerpt for a hypothetical RFP from an Indian bank for an internal knowledge assistant for branch staff.
| Req. ID | Requirement (summarised) | Compliance | Our approach | Assumptions and risks |
|---|---|---|---|---|
| F-04 | Answer staff queries from circulars and SOPs with source references | Full | Retrieval over the document repository; every answer cites document and section | Circulars available in machine-readable form; scanned PDFs need OCR, estimated separately |
| F-09 | Support Hindi and English queries | Partial | English and Hindi supported; accuracy on mixed Hindi-English to be measured in pilot | Evaluation set of real staff queries to be provided by bank |
| NF-02 | All data to remain within India | Full | Deployment in an India cloud region; model hosted in-region | Depends on availability of the chosen model in-region at contract signing |
| NF-05 | 99.5% monthly availability | Full | Two availability zones; documented fallback to search-only mode | Excludes bank network and identity provider outages |
| NF-11 | Role-based access: staff see only documents for their role | Full | Access filters applied at retrieval time from the bank's directory groups | Bank maintains accurate group membership |
| S-03 | Model must not be trained on bank data | Full | Contractual and technical controls; no fine-tuning in phase 1 | To be confirmed against provider terms |
Two habits make this matrix credible. "Partial" is used honestly, with a plan to measure, instead of claiming full compliance on something nobody has tested. And each assumption names who is responsible for it, so that when the scanned PDFs turn out to be a large share of the repository, the scope conversation starts from something written down.
Options, not a single answer
For any significant design question the architect presents options. For the same bank:
- Hosted model API in-region: fastest to deliver, pay per use, depends on the provider's regional availability and terms.
- Open-weight model self-hosted on the bank's cloud tenancy: more control over data and versions, but GPU capacity, patching and evaluation become the bank's or the vendor's ongoing cost.
- Search-first with limited generation: lower risk and cost, less fluent answers.
Each option gets a rough cost range, a delivery timeline, the main risks and what would change the recommendation. Where costs are uncertain, the architect says what they depend on, such as the number of daily queries, instead of presenting a false precise figure.
How the week differs by employer
In a services firm: two RFP responses in progress, solution review boards, client orals (presentations to the evaluation panel), and estimates negotiated with delivery leads who will have to live with them. In a product company or GCC: design reviews for new features, guiding engineering teams through trade-offs, and resolving cross-team questions about shared platforms.
Win rates, delivery surprises and design reviews
In pre-sales: win rate, deal size and the gap between estimated and actual delivery effort on won deals. That last one is the honest measure of an architect's work. In delivery-side roles: whether systems meet their stated reliability, security and cost targets, and how many architecture decisions had to be reversed.
Presales heads, delivery leads and client CTO offices
Pre-sales architects usually report to a head of solutions or practice lead, and work with sales, delivery managers, security, legal (for data terms) and partner teams at cloud providers. Delivery architects report to engineering leadership and work with tech leads, security and product.
Confused with solutions engineer and enterprise architect
A solutions engineer is focused on demos and technical selling for one product. An enterprise architect governs technology standards across a whole organisation. The AI solutions architect designs specific AI solutions, often across several products and the client's systems.
RFP, reference architecture and hyperscaler signals
- "RFP/RFI responses, solutioning, effort estimation" — services pre-sales.
- "Reference architectures for GenAI on AWS/Azure/GCP" — cloud-partner heavy.
- "Hands-on: build POCs in Python" — expect to write code; ask how much.
- "Responsible AI, governance, data residency" — regulated clients; compliance knowledge matters.
- "Pre-sales" alongside "quota" — commercially measured.
Shadow three bids, then own one design
Learn the firm's delivery capacity and past estimates first; architects who do not know what their teams actually delivered produce estimates that lose money. Then own one RFP response end to end, and follow a won deal into delivery to see where your assumptions held.
Principal architect, practice head or CTO-office roles
Principal architect, practice head, CTO of a business unit, or field CTO roles. The FDE for Professionals programme, on a proposed schedule for experienced engineers, covers enterprise AI architecture, data boundaries and commercial judgement of the kind this role demands.
Pankit Kumar has 10 years in Data Science & AI, building and shipping production systems in regulated pharma and clinical environments. He is a freelance trainer at Boston Institute of Analytics, AnalytixLabs and Scaler, and has taught this material to thousands of working professionals.
This article is part of our FDE for Professionals programme — 16 weeks (proposed). An accelerated advanced pathway for IT professionals ready to own enterprise AI delivery.
Explore FDE for Professionals