Planning
When NOT to use AI in a business workflow
Use rules for exact decisions, APIs for handoffs and people for accountable approvals. A practical guide to where AI adds cost or avoidable risk.
By Automate HQ · Updated · 4 min read
At a glance
Do not add AI when a written rule already determines the answer, when reliable source data is missing, or when errors cannot be detected before an important action. Use AI for a bounded interpretation task only when a measured pilot improves the complete workflow, including review and maintenance.
- 01Define the decision
- 02Try a rule
- 03Measure review work
- 04Limit AI authority
Exact calculations belong in rules
Invoice totals, due-date calculations and permission checks should follow explicit rules. Asking a language model to decide whether a required field is empty makes a predictable operation harder to reproduce. The same applies to copying an approved value between systems.
Suppose an invoice needs approval whenever its total exceeds an internal threshold. Store the threshold in configuration, normalize the currency under an agreed policy and compare values in code. AI might help read the invoice, but the approval rule should remain independently inspectable.
Choose the mechanism for each step
A workflow rarely needs one technology for every stage. Separate reading an ambiguous message from updating a record. The first may benefit from language interpretation; the second needs identity checks, permissions and duplicate protection.
The table describes AutomateHQ's design recommendations, not measured performance claims. Start with the simplest method that can express the requirement and test it against actual examples.
| Task | Start with | Where AI may help |
|---|---|---|
| Copy a form submission to a CRM | API integration and validation | Usually unnecessary |
| Calculate amounts or deadlines | Code and explicit business rules | Extract proposed inputs for checking |
| Route a known dropdown category | Lookup table | Usually unnecessary |
| Interpret varied customer messages | Reviewed classification pilot | Suggest category with fallback |
| Answer a published opening-hours question | Approved answer or menu | Understand varied wording |
| Change bank details or approve payment | Verified process and named approver | Summarize evidence without authority |
Unclear policy produces unclear automation
If two experienced employees disagree about a refund because the policy is incomplete, a model cannot resolve the underlying business decision. It can produce a plausible response that hides the disagreement. Write the policy and exception path before automating the answer.
The same problem appears with stale product lists and conflicting spreadsheets. Choose the source of truth and an owner for updates. A fluent answer based on an outdated price is still wrong. When the source does not contain an answer, the workflow should ask for clarification or hand the request to a person.
Avoid AI when checking costs more than doing
Measure the full task, not the time until the first draft appears. Include reading the suggestion, locating its evidence, correcting it and resolving exceptions. For low-volume work, setup and maintenance may consume more time than the automation returns.
Consider an explicitly illustrative calculation: 100 requests taking two minutes each use 200 minutes. If an AI-assisted process still needs 90 seconds of review per request and 40 minutes of weekly maintenance, it uses 190 minutes before implementation cost. The ten-minute difference would be a weak basis for a project. Replace every input with your own observed numbers.
Do not give a reader permission to become an approver
A model that reads invoices or customer messages encounters content written by outsiders. That content can include requests to ignore rules or reveal information. Keep available actions constrained by code and access controls; do not let text inside a document expand the model's permissions.
For a first deployment, let AI create a draft, category or proposed field value. Validate allowed values and show the source to the reviewer. Keep consequential actions, such as changing supplier payment details, behind a separate verified process. A confident explanation from the model is not independent evidence.
What evidence would justify adding AI?
Use a held-out sample representing ordinary work and awkward exceptions. Compare the current process, a rules-only baseline and the proposed AI-assisted process. Measure completed tasks, correction time, missed exceptions and total operating cost. Record the exact configuration so the test can be repeated.
Define a stop condition before the pilot. For example, pause automatic replies whenever the approved source is missing or a request asks for an action outside the workflow's scope. Expand authority only when the team can show that the checks work and someone owns the failures. The related invoice test guide provides a concrete evaluation protocol.