RAG Is Not Dead: The SkyDraft Evidence Playbook

By Karl Lehnert, Director, DevProStudio

The argument that retrieval-augmented generation—RAG—is dead makes a good headline. It is also the wrong question for an Australian business trying to produce a defensible proposal, statement of work or client report.

The useful question is simpler: can the drafting system show that it used the right source material, identify what is missing and keep unsupported claims out of the final document?

That debate is current. A September 2026 GitHub Blog discussion explicitly asked whether RAG is dead as part of a wider conversation about AI development patterns. New models can carry more context and follow instructions more reliably than earlier systems. None of that turns a model into your firm's source of truth.

For document-heavy SMEs, RAG still has a job. It is not to make prose sound clever. It is to bring approved evidence into a controlled draft, at the point where it is needed.

RAG is a retrieval pattern, not a truth machine

In plain English, RAG finds relevant material before asking a model to answer or draft. The material might be a proposal template, service catalogue, glossary, policy library, discovery notes or approved commercial clauses.

The model can still misread a passage, combine unrelated facts or produce a confident sentence the sources do not justify. Retrieval reduces one problem—drafting without relevant context—but does not eliminate review.

RAG projects fail when teams treat retrieval as an accuracy switch. Indexing a shared drive can retrieve stale documents, miss a key exclusion or present weak evidence with great confidence.

Our operator view at DevProStudio is blunt: the source system and the evaluation process matter more than the acronym. If your approved content is unclear, duplicated or out of date, a newer model will produce a more polished version of the same mess.

The SkyDraft evidence playbook

SkyDraft applies this principle to recurring business documents. Its public product description focuses on proposals, statements of work and reports. The workflow uses an uploaded document structure, brand voice, glossary and required-information checklist, drafts section by section, and asks clarifying questions where the source is unclear.

That is a better frame than “chat with all our files”. A sound evidence workflow has five parts.

1. Define the document contract

Choose one document class and state what a complete draft must contain. For an SOW, that might include objectives, inclusions, exclusions, dependencies, dates, responsibilities and approval details.

The contract is a checklist the workflow and reviewer can test. If a required fact is absent, ask for it or mark it missing—do not fill the gap with plausible text.

2. Curate the source pack

Do not point the system at every historical document. Separate approved templates and clauses from examples, archived versions and working notes. Assign an owner and review date to important material.

One clean service description is more useful than 40 old proposals containing different promises.

3. Retrieve evidence at section level

A whole document is usually too broad a retrieval unit. The evidence needed for “scope” differs from the evidence needed for “commercial assumptions”. Retrieve for the section being drafted, preserve a reference to the source, and make the retrieved passages available to the reviewer.

The aim is traceability, not a wall of client-facing citations. A reviewer should know where a claim came from and whether its source remains approved.

4. Test unsupported statements

Build a small evaluation set before rollout. Include normal examples, incomplete inputs, conflicting sources and a deliberately outdated document. Then assess:

• required-field completion;
• whether retrieved sources were relevant;
• unsupported factual or commercial statements;
• correct handling of missing information; and
• reviewer corrections before approval.

Do not score only fluency. Smooth prose is cheap. A missed exclusion or invented delivery date is expensive.

5. Keep approval outside the model

The output is a draft. A named person remains accountable for scope, assumptions, pricing language, privacy and client commitments. Approval should be an explicit workflow state, not something inferred because the AI finished generating.

A common implementation pattern

This is an implementation pattern, not a claimed customer case study.

Consider a small consultancy that prepares similar SOWs every week. It chooses one approved template, removes obsolete examples and creates a glossary for service names. Discovery notes supply client-specific facts. The workflow retrieves approved content for each section and asks a question when a dependency, date or exclusion is missing.

Before adoption, the team runs a fixed set of sample briefs. One has complete inputs, one omits the support window, one contains contradictory dates and one includes an old clause. The reviewer records unsupported claims and corrections. Only after those failure cases behave sensibly does the team use the workflow on live drafts.

That is less glamorous than “AI writes the proposal”, and far more useful.

What does a grounded drafting workflow cost?

No current public SkyDraft price was verifiable during this research, so a dollar comparison would be guesswork. Use a total-cost framework instead.

Count the work required to clean and maintain source material; configure retrieval and required fields; process model requests; review drafts; and correct failures. Then compare that with the present cost of searching old files, rewriting repeated sections and fixing errors late.

For a lightweight workflow, a general AI tool plus a disciplined source pack may be enough. A document-specific system becomes more attractive when the same document class recurs, multiple staff draft it, and missing facts create commercial risk.

The key metric is not subscription price per user. It is cost per approved document at an acceptable error rate.

Australian privacy and security controls

Source-grounded drafting may process names, contact details, project risks, commercial terms and other personal or confidential information. Treat the retrieval store as a business system, not a clever prompt library.

Under APP 11 guidance, an APP entity must take active measures to secure personal information it holds and consider whether it is permitted to retain it. APP 8 guidance is relevant when personal information may be disclosed to an overseas recipient.

In practice:

• minimise personal information before upload;
• restrict source libraries and draft generation by role;
• record where data is stored, processed and retained;
• confirm whether overseas disclosure is involved;
• separate approved sources from general shared-drive content;
• log source and version identifiers for review; and
• require human approval before external use.

These controls will not rescue poor content, but they prevent a document assistant from becoming an uncontrolled copy of sensitive business records.

RAG is alive when the evidence matters

RAG should disappear into the workflow. Staff should not need to think about embeddings or vector databases. They should see the right source, a clear gap when information is missing and a draft that is easier to verify.

For SkyDraft, that is the useful product test: not whether it can generate a polished page, but whether it helps a team turn approved material into a reviewable document without losing control of facts and commitments.

If your team is still copying last year's proposal and asking a chatbot to tidy it up, the next step is not a larger prompt. It is a smaller, better-governed evidence system.

FAQ

Is RAG obsolete in 2026?

No. Larger context windows and better models change how retrieval systems are designed, but businesses still need current, approved source material. RAG remains useful when evidence must be selected, traced and reviewed rather than buried in a giant prompt.

What is grounded AI document drafting?

Grounded drafting generates text from defined source material such as approved templates, clauses, notes and glossaries. A good workflow exposes missing information, preserves source references and requires human approval for business commitments.

How should an SME evaluate a RAG workflow?

Test it with a fixed set of complete, incomplete, conflicting and outdated inputs. Measure source relevance, required-field completion, unsupported statements and reviewer corrections. Fluency alone is not a quality measure.

Does source grounding guarantee an accurate document?

No. Retrieval can supply relevant evidence, but the model may still misinterpret it or generate unsupported language. Source quality, evaluation, access controls and accountable human review remain essential.

Build a document workflow you can defend

DevProStudio builds practical AI systems for Australian SMEs, including source-grounded document workflows and custom AI apps. If you want to replace copy-and-paste drafting with a controlled evidence process, talk to DevProStudio.