# Where should we start with AI?

AI-assisted editorial draft • Bill Allred review pending • Draft and sources checked 19 October 2026

*Illustrations are hypothetical, not Altiport client work.*

## Executive answer

Start with a business constraint you can describe, an outcome somebody owns, and a bounded change you can test. Include opportunities to serve more customers, offer a new service, or improve decisions alongside efficiency. The first investment should produce useful business evidence and teach you something about a larger opportunity.

A useful starting sentence is: “We cannot currently deliver this customer outcome because this part of the operating model limits us.” If the sentence only names software, keep working.

## Find the constraint behind the complaint

“We spend too much time preparing proposals” is an observation. Investigate what that prevents: pursuing smaller opportunities, responding while the buyer is still interested, or preparing technically sound alternatives. These lead to different designs and different measures.

Interview the person receiving the work as well as the person doing it. Review actual examples: completed work, abandoned requests, exceptions, and rework. Ask where time goes, what information is missing, and which decisions require judgment. A busy activity may have little effect on the overall result. Improving it could simply move a queue downstream.

Write three kinds of opportunity:

- **Improve an existing service:** make answers more complete or reduce avoidable handoffs.
- **Expand useful capacity:** handle demand the business currently declines or leaves waiting.
- **Create a capability:** provide a previously uneconomic analysis, preparation step, or customer interaction.

Then compare each with a non-AI alternative. Better forms, clearer policy, conventional automation, and eliminating an unnecessary step all deserve consideration. Anthropic’s engineering guidance recommends starting with the simplest suitable approach and adding complexity when it improves results. That supports testing the design rather than assuming every opportunity needs an agent. [Source: Anthropic](https://www.anthropic.com/engineering/building-effective-agents)

## Use gates before scores

An opportunity should pass four gates before entering a funded experiment:

1. **Ownership:** a business owner can change the workflow and accept its results.
2. **Evidence:** representative inputs and a credible baseline can be obtained appropriately.
3. **Containment:** consequential mistakes can be detected, escalated, or prevented within the test’s scope.
4. **Actionability:** a positive result would lead to a real operating decision.

A missing gate means resolve the dependency, narrow the test, or defer. Do not hide a missing permission or unacceptable consequence inside an attractive average score.

For opportunities that pass, compare customer value, capacity or growth potential, feasibility, learning value, and uncertainty. Use high/medium/low with a written reason. Numerical precision adds little if the underlying judgment has no evidence.

## Worked example: choosing a first move

**Hypothetical illustration.** An industrial distributor receives 120 quotation requests each week. Staff complete 80 and defer 40. Management considers three projects: drafting internal meeting summaries, preparing quote packets, and automatically approving discounts.

Quote preparation appears promising because it connects to unserved demand. But the team discovers that inventory validation, rather than writing, causes much of the delay. A useful experiment therefore retrieves approved product information, flags missing quantities, and assembles a review packet. Staff retain price and delivery commitments.

Suppose the test reduces preparation from 30 to 20 minutes for 60 eligible requests. That creates 600 minutes, or 10 hours, of weekly capacity. At 20 minutes per additional packet, the theoretical ceiling is 30 extra packets. It is not yet 30 additional completed quotations: inventory checks, reviewer availability, customer demand, and rework may reduce the gain.

The decision is to test quote packets while measuring completed, accurate quotations and reviewer workload. Discount approval waits for stronger controls. Meeting summaries may remain useful, but they answer a different business priority.

## Reusable opportunity worksheet

Copy these fields for each candidate. Record an evidence owner and “unknown” wherever the team is guessing.

| Field | Your entry | Hypothetical example |
|---|---|---|
| Customer outcome | What improves for whom? | Buyers receive complete quotes |
| Binding constraint | What prevents that today? | Inventory validation and preparation queue |
| New capability | What becomes possible? | More small requests can be considered |
| Proposed change | What would people do differently? | Review an assembled evidence packet |
| Baseline | Volume, quality, elapsed time | 80 completed requests weekly |
| Dependencies | Data, permissions, systems, people | Approved catalog and inventory access |
| Failure boundary | What cannot happen? | No unapproved price or delivery promise |
| Evidence needed | What would change the decision? | Net throughput after review and rework |
| Learning value | What transfers elsewhere? | Retrieval quality and reviewer adoption |
| Owner and next decision | Who decides what? | Sales operations: expand, revise, or stop |

## Limits and the next decision

This is a proposed prioritization method, not a validated Altiport scoring model. High-impact or safety-sensitive applications need domain-specific review. NIST’s AI Risk Management Framework is voluntary guidance for considering trustworthiness throughout design, use, and evaluation; completing a worksheet does not establish compliance or safety. [Source: NIST](https://www.nist.gov/itl/ai-risk-management-framework)

Finish with a one-page experiment decision: scope, owner, evidence, failure boundary, and the operational choice the evidence will support. Then [redesign the selected workflow](https://altiport.ai/resources/redesign-workflow-before-choosing-tool/) before selecting technology.