Testing Manual AI Fallbacks: George Dagliyan Research Guide

What should happen when the AI document assistant goes down?

When an internal AI document assistant is unavailable, the immediate goal is not to imitate the system perfectly; it is to preserve continuity, accuracy, and accountability. For George Dagliyan, whose work centers on AI adoption, technology acceptance, and adoption inhibitors, this is a practical test of whether the organization can keep moving when a favored tool disappears. The core answer is to shift to a manual, preassigned workflow, protect authoritative records, and require a clear signoff when the system returns.

This article’s editorial analysis treats the outage rehearsal as a continuity drill, not a claim that manual work is always fast or risk-free. A manual fallback can reduce dependency on one tool, but it can also introduce delay, transcription errors, and confusion if roles are unclear. The value of the rehearsal is in exposing those weak points before a real outage does.

What assumptions should be written down before the rehearsal begins?

Start by stating the scenario in plain English: the document assistant cannot generate, retrieve, summarize, or route documents for a defined period, but email, file storage, and core business systems remain available. Label this as a hypothetical outage, not a prediction. Specify which document classes are in scope, such as policies, procedures, customer-facing drafts, internal memos, or meeting notes, because the fallback process may differ by sensitivity and business impact.

Also record the authority assumptions. Identify who may approve manual edits, who may release a document, and who may restore the normal workflow after recovery. NIST’s Cybersecurity Framework is a voluntary framework designed to help organizations better understand and improve cybersecurity risk management, so this rehearsal can be aligned to internal governance without pretending that NIST prescribes this exact checklist. NIST’s AI Risk Management Framework likewise supports voluntary AI risk management; this article uses that spirit as a practical planning lens.

How should the alternate manual process work step by step?

A good fallback process should be simple enough to run under pressure. The manual route can begin with intake: a staff member logs the request in a shared tracker, captures the requester, deadline, document type, and urgency, then assigns a human owner. Next comes drafting from the authoritative source, followed by a second-person review for accuracy, consistency, and required fields. Finally, the owner releases the document and records the action in the tracker.

The important distinction is between assistance and authority. The AI assistant may normally help with formatting, retrieval, or summarization, but in an outage the human process must rely on source documents, approved templates, and named approvers. If an organization uses controlled templates, the rehearsal should specify where those templates live, who may update them, and how version numbering is applied so that staff do not improvise different copies.

Who should do what during the outage?

The rehearsal should assign roles before the exercise begins. One person should be the intake coordinator, one the drafter, one the verifier, and one the release owner. A separate recorder should track timestamps, issues, and exceptions. This division matters because outages often fail not from lack of effort, but from unclear handoffs and duplicated work.

To keep the test realistic, each role should have a short duty statement. The intake coordinator confirms what was requested and whether the request is urgent. The drafter works only from authoritative records. The verifier checks names, dates, file references, and any claims copied into the document. The release owner decides whether the document is ready to distribute. That structure is also a useful adoption lesson: technologies are easier to sustain when the organization can describe who does what without relying on the tool to fill in the gaps.

What counts as an authoritative record, and where should it come from?

Authoritative records are the approved sources that define what the organization should trust when the AI assistant is offline. These may include official policies, controlled templates, final signed versions, system of record entries, and approved source files. The rehearsal should force the team to identify which record wins when two copies differ, because conflicts are common during outages and recovery.

The worksheet below is the article’s proposed tool, not an official standard. For each document type, list the primary source, the backup source, the owner, the location, and the version rule. Example: policy updates may come only from the controlled policy repository; meeting summaries may come from the meeting notes folder plus attendee confirmation; customer-facing letters may require a final reviewed copy in the approval archive. The point is to make trust explicit before a crisis, not after one.

What worksheet can make the drill measurable without pretending to be a standard?

Use a Continuity Rehearsal Worksheet with five columns: task, manual path, authoritative source, reviewer, and recovery note. For each essential document workflow, record the normal AI-assisted path and the manual fallback path side by side. Then note the source that must be used if the two paths produce different text. This worksheet is a practical artifact created for this article; it is not an NIST requirement or a universal control set.

Add a simple decision rule for the exercise: if a request can be completed manually within the agreed service window and the source record is available, proceed manually; if the source is missing or the required approver is unavailable, pause and escalate instead of improvising. This rule is deliberately conservative. It does not guarantee safety, but it helps prevent silent errors from entering the record under time pressure.

How should the team test staff readiness before relying on the fallback?

Readiness is not proven by a policy memo. It is shown when staff can perform the fallback without searching for instructions every minute. A short tabletop exercise can test whether people know where the templates live, how to locate the source record, who approves release, and how to log exceptions. Tabletop here means a discussion-based rehearsal, not a live production disruption.

A stronger test is a timed dry run using a non-production document set. Give each role a packet of hypothetical inputs and ask the team to produce one controlled document manually. The evaluator should look for missed handoffs, duplicated edits, and uncertainty over source precedence. Because this is a rehearsal design, no measured results are assumed here; the value comes from observing where the process slows or breaks before the real outage does.

How should reconciliation happen after the AI system comes back online?

Recovery is where many organizations create avoidable confusion. Once the document assistant returns, the team should not immediately resume normal use without checking what was produced during the outage. The reconciliation step compares manual outputs against what the AI-assisted workflow would have stored, identifies duplicate files, and marks the authoritative version. Any unresolved discrepancy should be routed back to the release owner, not left to habit or convenience.

The signoff step is equally important. Owner signoff means the designated business or document owner confirms that the manual fallback was executed, exceptions were recorded, and the restored workflow is safe to resume for that document class. This is also the place to capture lessons learned: what slowed the process, which source was hard to find, and which approval step created a bottleneck. In Dagliyan’s research themes, this is where adoption facilitators and adoption inhibitors become visible in practice, because the organization sees which controls support continuity and which ones block it.

Frequently Asked Questions

Is a manual fallback meant to replace the AI document assistant permanently?

No. The fallback is a continuity measure for temporary outage conditions, not a claim that human processing is always better or scalable. A manual path may preserve operations when the system is down, but it can be slower and more error-prone. The rehearsal should help the organization understand where that path works, where it does not, and when escalation is preferable to improvisation.

What is the difference between an authoritative record and a convenient copy?

An authoritative record is the version the organization has decided to trust for decisions, approvals, or release. A convenient copy may be easier to find, but it may be outdated, incomplete, or edited without control. During a fallback rehearsal, the team should name the authoritative source first so staff do not rely on the nearest file when versions conflict.

How long should the outage rehearsal last?

There is no universal duration. The right length depends on the number of document types in scope, the number of roles being tested, and the time needed to exercise recovery and signoff. Many organizations can start with a short tabletop discussion and then add a timed dry run for one document class. The point is to test the workflow, not to stage a dramatic disruption.

What should be logged during the fallback?

At minimum, log the request, the time received, the manual owner, the source used, the reviewer, the release decision, and any exception or delay. Also note where the original AI-assisted path would have handled the task differently. Those logs support reconciliation after recovery and make it easier to spot recurring issues such as unclear ownership or missing source material.

How does this continuity rehearsal relate to AI adoption research?

It reflects a common adoption reality: organizations do not just adopt a tool, they adopt a way of working that depends on trust, process clarity, and human readiness. If the fallback is confusing, the technology feels more fragile. If the manual path is clear, the organization has a better chance of sustaining use under stress. That is consistent with the research themes associated with adoption facilitators, adoption inhibitors, and technology acceptance.

Sources and further reading

• NIST — Cybersecurity Framework
• NIST — AI Risk Management Framework