Design a Source-Grounded Generative AI Drafting Workflow

Source-grounded drafting is not a cleverer prompt or a fact check bolted onto polished prose. It is a controlled chain in which approved sources, qualified evidence, generated wording, review decisions and final approval remain separate and inspectable. Consider a pilot brief that says, “The pilot proved the workflow will save analysts time.” If the evidence records only that a small, self-selected group spent less time preparing first drafts during measured pilot tasks, the sentence has crossed the evidence boundary. Its confident tone and nearby citation do not repair that leap. A sound workflow exposes the limitation before drafting, makes unsupported gaps visible and records who knowingly accepts the finished document.

What must be settled before the model drafts?

Generation should begin only after the team has assembled and approved a minimum drafting packet. That packet defines the document's purpose, audience, deadline, accountable owner, consequence of error, required format and the decision or action it must support. These details prevent a useful-sounding draft from quietly becoming a different document: an exploratory note turning into a recommendation, for example, or an internal observation becoming an external commitment. The owner should also state who may accept the output as a business document.

The packet must then fix the operating boundary: the approved tool, permitted information, registered sources, evidence-card format, output template, review map and record rule. It should say what the model may produce, which inferences are prohibited and how a gap must appear. Confidential, personal, privileged, restricted or contractually protected material belongs only in tools and workflows that the relevant organisational owners have approved for that information; the drafting team must not improvise the permission decision.

This discipline is consistent with NIST's voluntary AI Risk Management Framework, which calls for documented knowledge limits, intended application scope, output oversight and differentiated human-AI responsibilities. Applied at document level, that means the model's role is drafting within a stated boundary, while people retain ownership of evidence, exceptions and release. Approval of a source set is still only a control boundary. It does not certify that every source is accurate, complete, current, suitable or legally usable.

How do approved sources become usable evidence?

Approved documents become drafting evidence through a source register and evidence cards, not by being poured wholesale into a model. Give each source a stable identifier and record its owner or publisher, version or date, relevant scope, approval status, access conditions, known limitations and refresh trigger. Approval should consider authority, relevance, currency, suitability, permissions and limitations. The appropriate information-governance, privacy, security, legal, records or domain owner determines the checks that apply inside the organisation.

Each evidence card should carry the source identifier, exact passage or measurement, precise locator, supported point, qualifications, allowed use and prohibited inference. This structure lets a reviewer inspect evidence without first untangling the model's interpretation. NIST's Generative AI Profile recommends documenting reliance on upstream sources and reviewing data for accuracy, representativeness, relevance and suitability across lifecycle stages. Its guidance is voluntary, but the underlying control is practical: availability is not approval, and approval is not unlimited permission to infer.

Take evidence card E-04 from a controlled drafting pilot. Its source is the pilot evaluation, section 3.2, and its supported point is that participants reduced first-draft preparation time within measured tasks. The card retains three limitations: the sample was small and self-selected, there was no comparison group, and review time was recorded separately. It permits a description of the observed result but prohibits predicting a percentage saving for analysts or claiming organisation-wide productivity. That prohibited inference is as important as the usable finding.

A claim-to-evidence map makes this boundary visible downstream. Current NIST research describes an experimental pipeline that combines an authoritative document corpus, cited outputs, citation evaluation and structured results. That work is not a finished production standard, but it illustrates a useful separation: source material, downstream claims and evaluation results remain distinguishable. A simple register and card system can apply the same traceability principle without pretending that the mapping itself proves the source or claim correct.

How can the model draft without quietly filling gaps?

Keep the evidence layer physically or logically separate from the reader-facing draft, and give the model a narrow generation contract. Supply evidence-card identifiers and an output template, but instruct the model to draft only the supported decision context, observations, unresolved risks, options and recommendations. Generated sentences should point back to cards without overwriting their passages or qualifications. Reviewers can then compare the model's wording with the underlying evidence rather than treating fluent prose as a faithful summary by default.

The contract must define what happens when support is absent: insert a visible marker such as [EVIDENCE NEEDED], leave the point unresolved and do not complete it from general model knowledge. It should also prohibit invented citations, unregistered market statistics, vendor claims and unstated interpretations of organisational policy. A gap is useful operational information. It tells the owner to obtain approved evidence, narrow the requested claim or remove it, instead of allowing an attractive sentence to conceal uncertainty.

Citations improve traceability, but a source label does not establish that the adjacent wording is entailed, accurate, complete or suitable. The NIST Generative AI Profile recommends reviewing and verifying sources and citations in generated outputs. Verification therefore returns to the cited passage and its qualifications. The drafting model should never serve as the only fact-checker for its own consequential claims, because repeating or defending its earlier wording is not an independent examination of the approved evidence.

Who reviews each part of the draft?

Assign named reviewers to distinct questions rather than asking for generic human oversight. The source owner confirms the approved set and the fidelity of its evidence cards. An evidence reviewer compares consequential clauses with those cards and their underlying passages. An editor improves structure, clarity, tone and audience fit without silently changing evidentiary meaning. A specialist reviews only when a policy, sensitivity or domain trigger appears, while the approver accepts the recommendation, its conditions and its residual uncertainty.

• Source decision: Is this the approved, current source set for the stated purpose?
• Evidence decision: Does each consequential clause stay within the passage and its qualifications?
• Editorial decision: Is the document clear and usable without strengthening or weakening its meaning?
• Specialist decision: Does a triggered policy, permission, security or domain issue require correction or escalation?
• Release decision: Is the accountable owner prepared to accept the recommendation and remaining uncertainty?

A small South African team may give several roles to one person, but it should preserve the questions and record which decision that person is making. NIST's AI RMF supports clear roles, responsibilities, communication lines and differentiated human-AI oversight. For long or consequential drafts, bounded claim or clause review is also a sensible practice. LongEval found lower annotator disagreement with finer-grained judgements across two long-form summarisation datasets, although transferring that result to business drafting is a practical inference, not a universally measured outcome.

Automated claim-to-source checks can help prioritise review, flag missing locators or assemble an evidence map. NIST's current grounding work presents such checks as experimental evaluation probes that store structured evidence-linked results. It does not establish an automated release authority. People must still decide whether the passage supports the wording, whether specialist risk has been addressed and whether the document may be issued. Automation can organise the queue; it cannot accept accountability on the organisation's behalf.

A grounded draft is one whose consequential claims can be traced, challenged, corrected and knowingly approved.

Which edits need an explicit record?

Record an edit explicitly when it changes factual meaning, interpretation, recommendation, commitment, obligation, risk treatment, source basis or approval status. Ordinary spelling, layout and stylistic changes can remain in normal version history unless organisational policy requires more. This proportional test keeps the record focused on decisions another reviewer may need to reconstruct. It also prevents a seemingly editorial rewrite from changing what the evidence supports after the evidence pass has already been completed.

A lightweight entry needs the wording before and after the change, the reason, affected evidence, reviewer, approver and decision time. The NIST Generative AI Profile describes provenance metadata that can include creators, dates, modifications and sources, and recommends retaining content-change records with associated metadata. It does not impose a universal retention period or a keep-everything rule. The organisation's relevant records, privacy, legal and information-governance owners must decide which prompts, inputs, outputs and drafts require preservation.

The pilot claim shows the test in practice. Replace “The pilot proved the workflow will save analysts time” with “In the measured pilot tasks, participants spent less time preparing first drafts; the small, self-selected sample does not establish the effect for analyst work.” The record should say that the edit removed unsupported causal and population-wide inferences, restored evidence card E-04's limitations and was accepted by the evidence reviewer and approver. That is a meaning change, not cosmetic copyediting.

How should controls differ by document type?

Use the same evidence-to-approval pattern for briefs, reports and routine messages, but scale its weight to consequence, uncertainty and sensitivity. A decision brief needs a clear owner, supported options, recommendation and residual uncertainty. An analytical report needs method, exclusions, limitations and source versions. A routine message can rely on approved facts and fixed language with a sender check. Length is not the deciding factor: a short email creating a contractual or financial commitment may warrant closer scrutiny than a long, low-risk internal summary.

The table is a minimum pattern, not a universal control schedule. Teams can reduce ceremony for recurring, low-consequence messages whose facts and fixed wording are already approved. They should increase scrutiny when a document changes a commitment, introduces uncertain evidence, carries sensitive information or requires specialist judgement. A template should make escalation easier, not invite staff to tick boxes around a real issue. If the risk falls outside the template, pause and route it to the named owner.

How does the workflow stay reliable as work changes?

Keep the workflow reliable by assigning refresh ownership, sampling completed packets and changing upstream controls when the same correction recurs. Revisit a registered source when its facts, version, relevance, permissions or organisational status changes. The NIST Generative AI Profile recommends defining periodic-review responsibilities for content provenance and documenting human oversight roles, which supports explicit ownership. The interval itself should reflect the source and document risk rather than an arbitrary universal calendar.

Sample completed packets to test whether consequential claims still trace to approved passages, limitations survive editing, triggered reviews occur and material edits receive a record. Examine exceptions as well as successful releases. If the model repeatedly converts a pilot observation into an organisation-wide forecast, do not merely add another reviewer at the end. Tighten the evidence card's prohibited inference, adjust the output template or narrow the generation contract so that the error becomes less likely to travel downstream.

When a new source sits outside the approved boundary, or a permission, sensitivity, policy or specialist-review trigger appears, pause the affected draft. Route the issue to the named owner, record the boundary decision and resume only after the source set or exception is explicitly resolved. Begin with one recurring document type and make this path visible. Consult the appropriate organisational owners for regulated statements, uncertain permissions, protected information, contractual commitments, specialist judgements and retention obligations; the drafting model must not determine those requirements.