AI consultant vs software developer for New Zealand businesses.

Which question does each one answer?

An AI consultant answers “what should we do, and is it worth it?” A software developer answers “how do we build this reliably?” Most failed AI projects got one of those questions wrong, so the useful comparison is about the decisions each is good at, not job titles.

The line blurs at the edges. Some consultants prototype, and some developers have shipped several AI systems. Judge people on what they show you, not on the title.

What does an AI-aware adviser catch that others miss?

The value of advice is in specific, local catches. Three examples from 2026:

• Xero’s developer terms. Since 2 March 2026 Xero has charged developers through usage-based plans, and its terms ban using API data to train AI models. A build plan that ignores either is wrong from day one.
• Model retirements. OpenAI shut down its Assistants API on 26 August 2026, and Anthropic retires Claude Sonnet 4.5 from its API on 30 November 2026. A plan without a migration budget understates cost.
• Privacy expectations. The Privacy Commissioner expects a privacy impact assessment before AI tools are used, and human review before acting on AI outputs. Both shape the design.

A strong developer may catch all three. The point is that someone must, before money is spent.

How do you test both on the same brief?

Give each candidate, consultant or developer, the same paid brief from your business: “cut the time it takes to turn job notes into quotes”. Use real, anonymised job notes. Our quote preparation workflow shows what good looks like.

1. A one-page plan. Ask for the success measure, the steps AI would take, the steps that stay manual and the risks. Look for a number to move and an honest “leave this part alone”.
2. A 20-case test. Ask them to run 20 real job notes through a quick prototype or a business AI workspace and report how many quotes needed no correction. Look for measured results, not adjectives.
3. A running-cost estimate. Ask for a monthly cost in NZD, with assumptions, and what happens when the model is retired. Look for review time in the estimate, not just model fees.

Then ask both the same closing question: “Do you earn commission or margin on any tool you recommended?” A good adviser answers plainly.

What does each cost, and when does advice pay?

Use expected value with your own numbers. Let B be the build cost, p the chance the build targets the wrong workflow, tool or scope, f the share of that waste good advice would avoid, and C the cost of advice. Advice pays when C < p × f × B.

Illustrative example with stated assumptions, not quotes: a build estimated at B = NZ$60,000, a one-in-three chance of building the wrong thing (p = 0.33) and advice that avoids half of that waste (f = 0.5). The threshold is 0.33 × 0.5 × 60,000 = NZ$9,900. Advice costing less than that is worth buying; advice costing more is not, on cost alone.

Three decision rules follow:

• Small builds: when B is close to C, skip the plan and build a narrow prototype instead; it is the cheapest advice available. See AI proof of concept.
• A product already does it: you may need neither. Configure the product and train staff; see custom AI versus off-the-shelf tools.
• Large or risky builds: p and B are both high, so advice clears the bar easily.

Running costs belong in both quotes: model fees, monitoring, the evaluation set and migrations. Our AI project brief template asks for them up front.

Which should you choose?

Choose an AI consultant if

• You cannot yet name the workflow and the number you want to move.
• You are choosing between a product, a configured product and a custom build.
• Personal information, regulated work or customers are involved and you need the risks mapped.
• Leadership needs a costed recommendation before approving spend.

Choose a software developer if

• The workflow, the tool and the success measure are written down and agreed.
• The work is mostly integration with systems such as Xero, HubSpot or your own database.
• They can show how they will measure accuracy before launch.
• Someone in your business will own the system after it ships.

If you cannot find both skills in one person or firm, pair them, and make the consultant’s plan the developer’s acceptance test. Levy Street does both, through AI consulting and custom AI development; judge us on the same three tasks.

Where do both go wrong?

• Decks without numbers. A consultant’s report with no running costs or measured accuracy is opinion.
• Builds without an evaluation set. A developer who cannot say how often the system is right will not know when it gets worse. See evaluating AI generated code for the review habits that help.
• Agents where a workflow would do. Both can over-engineer. Many quoting and triage jobs need a fixed workflow with one AI step, not an autonomous agent.
• Nobody owns the outcome. Whoever you hire, name an owner inside the business who signs off the success measure.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.