Back to all articlesAI action safety

Choose Absolute and Relative Limits for AI Value Changes

Alactic Editorial10 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

An AI assistant proposes changing estimated service effort from 2 hours to 18 hours. Both values might fit the field’s permitted range, but the proposal increases the estimate ninefold. A range check alone would accept the destination without examining the journey.

To set numeric limits for AI write actions, combine three controls: an absolute range for the resulting value, limits on the size of the change, and explicit exceptions that route uncertain cases to review. Evaluate them together before an authorised write. Passing these checks means a proposal is eligible under a defined policy—not that the estimate is correct.

The fictional service-effort policy below shows how to make those controls concrete, including exact boundaries, missing baselines, repeated small changes, and recovery after a failed write. All example values and thresholds are hypothetical and require local validation.

1. Separate the destination from the size of the move

Absolute bounds answer: “May this field contain this value?” Change-size bounds answer: “May this workflow move the existing value this far?” These are different questions, and neither replaces the other.

  • Absolute value range: the permitted minimum and maximum for the proposed final estimate.
  • Absolute change cap: the maximum increase or decrease measured in hours.
  • Relative change cap: the maximum increase or decrease measured as a percentage of a valid baseline.
  • Review exceptions: conditions that prevent an ordinary write even when the arithmetic appears acceptable.

For a valid, positive baseline, calculate the absolute change as abs(proposed − baseline). Calculate the percentage change as abs(proposed − baseline) / baseline × 100. Using the magnitude of the change applies the same limits to increases and decreases. If your operation needs different rules by direction, specify both explicitly.

Require every applicable condition to pass. An “hours cap OR percentage cap” rule can admit changes that the other cap was intended to stop. An “hours cap AND percentage cap” rule preserves both protections.

2. Define a complete hypothetical policy

Fictional worked example: a service operations team lets an authorised assistant propose revisions to the estimated effort for a service job. The estimate is measured in hours, not elapsed time, staffing levels, or a customer commitment. This policy governs revisions to that field only.

The following hypothetical policy is intentionally specific enough to implement and test:

  • Permitted final value: 1 through 80 hours, inclusive.
  • Permitted precision: increments of 0.5 hours. Do not silently round other values.
  • Ordinary change limits: no more than 4 hours and no more than 25% of the positive baseline, with equality allowed.
  • Baseline: the most recent human-approved estimate, captured with its record version. Before approval of a new baseline, evaluate later proposals against the same baseline.
  • Current value: read separately to detect a no-op or a conflicting edit. It may differ from the approved baseline because of an earlier authorised revision.
  • Required context: a recorded reason for the revision and a valid approval-baseline reference.
  • Review conditions: missing, zero, invalid, or disputed baselines; a change above either cap; or missing supporting context.
  • Rejected proposals: nonnumeric or nonfinite values, wrong units, unsupported precision, and final values outside the absolute range.

In this hypothetical policy, zero is not a valid estimate. A missing estimate means “not yet established,” not zero. Creating an initial estimate therefore follows a separate reviewed path rather than masquerading as a revision.

Permission is another independent gate. A numeric proposal cannot confer authority to edit a record. OWASP’s authorization guidance recommends least privilege, default denial, and permission checks on every request. Apply that distinction even when all numeric tests pass.

3. Use this filled decision table

The table applies the hypothetical policy above. Except where noted, the current stored estimate equals the baseline, the supporting reason is present, and the actor has permission to request the change. “Eligible” means eligible for an authorised write, not already written or proven correct.

Case | Baseline | Proposed | Change | Change % | Decision
A    | 16 h     | 20 h     | 4 h    | 25%      | Eligible: both caps met exactly
B    | 16 h     | 12 h     | 4 h    | 25%      | Eligible: decrease at both caps
C    | 2 h      | 6 h      | 4 h    | 200%     | Review: percentage cap exceeded
D    | 40 h     | 46 h     | 6 h    | 15%      | Review: hours cap exceeded
E    | 16 h     | 20.5 h   | 4.5 h  | 28.125%  | Review: both caps exceeded
F    | 1 h      | 1.5 h    | 0.5 h  | 50%      | Review: small baseline, large ratio
G    | 1 h      | 1 h      | 0 h    | 0%       | No-op: already at lower boundary
H    | 80 h     | 80 h     | 0 h    | 0%       | No-op: already at upper boundary
I    | 80 h     | 80.5 h   | 0.5 h  | 0.625%   | Reject: above final-value maximum
J    | 1 h      | 0.5 h    | 0.5 h  | 50%      | Reject: below final-value minimum
K    | Missing  | 8 h      | N/A    | N/A      | Review: initial estimate required
L    | 0 h      | 4 h      | 4 h    | Undefined| Review: invalid zero baseline
M    | 16 h     | 18.25 h  | 2.25 h | 14.0625% | Reject: unsupported precision
N    | 16 h     | 18 h     | 2 h    | 12.5%    | Review: revision reason missing
O    | 16 h     | 20 h     | 4 h    | 25%      | Stop: record version changed
P    | 16 h     | 22 h     | 6 h    | 37.5%    | Review: current value is 20 h

Cases A and B make the boundary rule unambiguous: equality passes. Cases G and H show that the endpoints themselves are valid. Cases I and J show that even a small move cannot escape the final-value range.

Case P prevents a subtler bypass. Moving from the current 20 hours to 22 hours looks modest when considered alone. Against the unchanged approved baseline of 16 hours, however, the total movement exceeds both change caps. Do not automatically reset the baseline after each assistant write.

Keep a structured result for every case: proposed value, baseline, calculated differences, failed rules, policy version, and disposition. A reviewer should not need to reconstruct the calculation from a conversational explanation.

4. Handle missing and unstable baselines explicitly

A percentage check is only meaningful if the denominator represents the intended reference value. Never substitute a convenient default for a missing baseline simply to complete the formula.

A zero baseline makes percentage change undefined. An invalid negative estimate may yield a number from a formula but still fail the field’s meaning. A positive baseline can also be untrustworthy if it belongs to an older scope of work or a different unit.

Use distinct exception reasons rather than one generic “validation failed” state:

  • Missing baseline: ask a reviewer to establish an initial estimate.
  • Invalid baseline: correct the source record through an authorised process before evaluating a revision.
  • Disputed baseline: show the competing values and pause the write.
  • Changed scope: request review of both the scope and the new baseline rather than treating a large revision as routine.
  • Stale record version: reread the record and evaluate the proposal again.

Small positive baselines create another trade-off. In case F, the smallest permitted increment already exceeds the percentage cap. That may be intentional: low estimates receive human attention. If it creates excessive review work, consider a separately validated low-baseline rule. Do not insert a hidden denominator floor that quietly weakens the percentage check.

5. Make review an explicit decision, not an override button

Review should explain which kind of exception is being requested. A proposal outside a change cap is different from a proposal that cannot be represented in the field.

Under the hypothetical policy, an authorised reviewer may approve a specific change that exceeds the hours or percentage cap. That approval does not waive the absolute range or precision rules. A legitimate need for a value outside those constraints requires a separate policy or field-design change, followed by fresh validation.

A useful review packet contains:

  • The current value, approved baseline, proposed value, and units.
  • The absolute and percentage changes, including any undefined result.
  • The exact rules that failed and the revision reason.
  • The relevant scope change or supporting record.
  • The record version, policy version, and requested action.

Bind approval to that proposal and record version. If the value changes during review, do not reuse the approval for a different edit. The reviewer can approve the proposal, reject it, request more context, or submit a different value for fresh evaluation.

After a reviewed change is successfully written, the policy may designate it as the new human-approved baseline. Record that transition explicitly. A rejected proposal, an assistant suggestion, or a failed write must not reset the baseline.

6. Enforce the rules where the write happens

Instructions to an assistant are useful for shaping proposals, but they are not the enforcement boundary. Validate again in the trusted write path using the current record, applicable policy, and request permissions.

A recommended sequence is:

  1. Confirm the actor may change this field on this specific record.
  2. Parse the proposal into the expected numeric type and units; reject nonfinite or ambiguous input.
  3. Load the current value, approved baseline, record version, and policy version.
  4. Check absolute bounds and precision without silently clipping or rounding.
  5. If the proposed value already equals the current value, return a no-change result without a mutation.
  6. Evaluate baseline validity, change caps, required context, and any proposal-specific approval.
  7. Write only if the record version still matches the evaluated version; otherwise stop and reevaluate.
  8. Record the confirmed outcome separately from the proposal and approval.

Clipping a proposal to the nearest permitted value is not a safe substitute for rejection. If an assistant proposes 90 hours, storing 80 hours invents a different estimate that nobody necessarily intended.

If people also enter values through a form, give errors that explain the failed condition and the next step. W3C’s input-validation guidance calls for understandable, accessible errors and notes that client-side checks do not replace server-side validation. A message such as “This change is 37.5% of the approved baseline; review is required above 25%” is more actionable than “Invalid value.” These numbers remain hypothetical.

7. Recover without compounding the mistake

A timeout does not establish whether a write succeeded. Before retrying, check the recorded operation outcome and current record. A repeated absolute-value assignment may leave the same estimate, but it can still duplicate downstream activity unless the workflow deliberately handles repeated requests.

Prefer a proposal that says “set the estimate to 20 hours” over an unqualified instruction to “add 4 hours.” Bind it to the evaluated record version and a unique operation reference. This recommended design makes ambiguous retries easier to reconcile; it is not a claim that a particular product provides those controls.

If a wrong estimate was written, pause further assistant changes to that record. Preserve the proposal, approval, and write history. Have an authorised person evaluate a correction against the current record rather than blindly restoring an older value that may now be stale.

Also distinguish a policy failure from a reasoning failure. A write that exceeded a configured cap indicates an enforcement problem. A plausible but incorrect estimate that passed every check indicates that numeric limits were insufficient to establish business correctness. Tightening thresholds alone may not address the underlying issue.

8. Validate thresholds with a small, measurable rollout

Start by copying the decision table into a test checklist and adding representative local revisions with known review outcomes. Test exact boundaries, missing values, conflicting edits, repeated proposals, and increases as well as decreases.

Before permitting writes, run proposals through the policy without changing records. Track the share eligible for writing, the share routed to review, reviewer acceptance of exceptions, and proposals reviewers consider wrong despite passing the numeric gates. Segment results by baseline size: one overall review rate can hide friction around small estimates.

Then choose a limited, authorised rollout with a named owner and a review date. Your next concrete deliverable is a policy record containing the units, inclusive bounds, precision, both change caps, baseline-reset rule, exception authority, and test results. The observable goal is consistent routing and traceable writes—not a promise that every permitted estimate is right.

AI action safetyNumeric controlsApproval workflowsService operations

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications