How AI Contributes to Business Resilience and Agility
AI can help an organization notice a disruption, sort records for review, and prepare options for the people who keep a service running. Business resilience concerns maintaining and recovering essential work. Agility concerns how quickly an organization can make and carry out a suitable adjustment. Both depend on verified data, staff capacity and a usable fallback.
Business resilience and agility
Start with the service and the people who rely on it. For a wholesaler, the resilience question might be whether priority deliveries continue during a dispatch outage. The agility question might be how long staff take to agree on revised delivery slots. Track delivery coverage and time to approval separately.
When requests carry forward, use the service backlog calculator to see how the queue changes across intervals.
How data and AI work together
Create a register for the records the decision actually uses. Record who owns each source, when it was observed, when it was last updated, who can see it, and how corrections enter the workflow. Explicitly mark absent records and disagreements. A recent import can still contain stale observations.
Decide which record can authorize an action. A carrier’s current signed receipt may support a change to inventory. A draft email or an AI summary may point staff to a record that still needs verification. If the reference is missing, route the decision to the person responsible for that data.
What AI can assist with during a disruption
| Disruption | Available data | Possible assistance | Human decision | Fallback | Measure | Limitation |
|---|---|---|---|---|---|---|
| Late inbound shipment | Dated carrier notices and receiving log | Group inconsistent status reports for triage | Verify receipt and prioritize deliveries | Review the log and call the carrier | Time to verified status | An estimate can be mistaken for a receipt. |
| Staffing gap | Roster and known service demand | Draft options for coverage | Authorize a feasible roster | Use the duty roster and escalation list | Urgent shifts covered | Availability may differ from the roster. |
| Demand surge | Order history and current requests | Summarize demand patterns or compare forecast assumptions | Set a review threshold and choose the plan | Use verified order counts and a fixed decision rule | Error at the decision horizon | A shift in demand can invalidate old patterns. |
| Service outage | Approved status and customer contacts | Draft a status notice for review | Check facts and release the notice | Use an approved status template | Time to an accurate notice and corrections | A draft can omit exceptions. |
Worked example: keeping priority orders moving
Illustrative exercise. A regional distributor has 60 priority orders to handle during a four-hour window. Available stock can cover all 60. Staff can verify 48 orders and dispatch can handle 45 in that window. The figures are stipulated.
The service ceiling is the smallest constraint: min(60 requests, 60 orders in stock, 48 reviews, 45 dispatch slots) = 45 orders. At least 15 requests remain unmet in this window under these assumptions. The same ceiling applies to the current process, a structured human-only queue and AI-assisted triage with the same inputs.
A fair exercise gives the three groups the same incoming records, staff and slots. Measure time to the first verified urgent order, priority errors and staff minutes spent checking suggestions. The exercise contains no observations for those measures. Use the serial service bottleneck calculator to vary the stated constraints or the manual and AI review workload method to compare work you have actually measured.
How to evaluate AI services and advisers
Ask the supplier to define one decision, the minimum service, the baseline and the evidence that would justify a change. Require a data inventory with permission boundaries and a trial plan that includes missing inputs and service outages. Identify who owns each decision during the trial and after handover.
Specify deliverables before purchase: a documented workflow, representative test cases, review costs, a recovery runbook, exportable records and an exit plan. Name conditions for stopping the trial, removing access and returning to the existing process. Ask who fixes errors after launch and how those fixes will be verified.
Use the process cost break-even calculator to compare declared setup and recurring costs at the same workload.
Generative AI, pretrained models and conventional analytics
Generative systems produce outputs such as text or images. A pretrained model starts with capabilities learned before local deployment and can be adapted or used directly. Pretraining describes the starting point, so a generative model and a classifier can both be pretrained. Check the local vocabulary, data access, input changes and version behavior before relying on either.
A spreadsheet, rules engine or conventional forecasting model may be sufficient when the task is well specified. A rules-based duplicate check can match known order IDs. A language model may help with inconsistent free-text records, with extra review for unsupported matches.
A dated change in enterprise guidance
In July 2024, NIST published a companion profile for generative AI that addresses risks such as unreliable generated information and data privacy. A practical procurement implication is to ask how a vendor evaluates those risks in the proposed workflow and what staff can verify when the system changes. The publication records a change in available guidance. It provides no measure of enterprise adoption or business continuity gains.
Continuity use cases by business function
| Function | Disruption | Possible assistance | Required input | Human control | Continuity measure | Fallback |
|---|---|---|---|---|---|---|
| Operations | Dispatch or scheduling interruption | Group urgent jobs and draft a recovery queue | Verified orders, capacity and priorities | Supervisor assigns jobs and approves exceptions | Essential jobs completed in the window | Use a printed queue and phone confirmations. |
| Accounting | Late or inconsistent settlement records | Flag candidate mismatches and compare cash-flow scenarios | Dated ledger, authorized statements and explicit scenario inputs | Controller approves reconciliation and entries | Unresolved items and time to reconcile | Match records manually and record exceptions. |
| Sales | Order system outage or supply shortage | Summarize existing commitments and demand changes | Confirmed orders and service commitments | Account owner verifies promises and approves changes | Commitments correctly updated | Use the approved customer list and call log. |
| Marketing | Conflicting public outage messages | Draft consistent service notices from approved facts | Verified status, approved copy and channel list | Communications lead checks and releases the notice | Correct notices and time to publish corrections | Publish an approved manual status template. |
| Rules-based automation (RPA) | Manual queue grows while a system is unavailable | Run fixed reconciliation or routing rules, with optional AI review of exceptions | Stable fields, explicit rules and exception criteria | Process owner approves exceptions and can pause the job | Unhandled exceptions, errors and restored throughput | Stop the job and reconcile the queue manually. |
Implement and evaluate one continuity decision
- Define the service and its minimum useful level.
- Measure the existing human and process baseline.
- Specify the decision AI would assist and who approves it.
- Document data, dependencies, failure costs and permissions.
- Run representative tests with matching staff, information and constraints.
- Rehearse manual operation, rollback and exit procedures.
- Review service outcomes, errors and effort before deciding whether to retain the system.
Dependencies and errors to watch
Provider outages, stale inputs, feedback delays and plausible but inaccurate summaries can shift work onto the same people handling the disruption. Test the fallback with the actual staff and equipment available on a difficult shift. Include time spent verifying and correcting output in the comparison.
Questions before adopting AI for continuity
What should we pilot first? Pick a reversible decision with a known owner and enough reliable records to compare outcomes.
What if the review workload rises? Count verification and correction time. A rule or manual workflow may be easier to sustain.
What should management see? Show minimum service achieved, unresolved cases, errors, recovery time and workload as separate results.
Sources and what they support
- NIST — AI Risk Management Framework 1.0: AI Risks and Trustworthiness. Published January 2023. Supports evaluating reliability and robustness under conditions representative of intended use. Retrieved .
- NIST — Generative Artificial Intelligence Profile (AI 600-1). Published 26 July 2024. Documents the 2024 NIST generative AI risk profile, including risks involving unreliable output, privacy and information integrity. Retrieved .
- NIST — Contingency Planning Guide for Federal Information Systems (SP 800-34 Rev. 1). Published 2010. Provides a reference for setting continuity priorities and planning recovery around business impacts. Retrieved .