AI agency vs internal hire for New Zealand businesses.

How do an AI agency and an internal hire compare?

Levy Street is an agency, so read this with that in mind; we have tried to write it so it would say the same if we were not. The honest summary is that each option fails in a predictable way, and the right one depends on the shape of your workload.

The skills gap is real. In Datacom’s 2026 survey of New Zealand organisations, lack of internal capability or skills was the most-cited barrier to scaling AI, at 20%. That is a vendor survey of larger organisations, but it matches what both options are trying to solve.

What does each cost over three years?

No salary survey or agency rate is reliable enough to publish here, so use your own numbers in these formulas.

• Internal hire, per year: I = S × (1 + o) + T, where S is salary, o is the on-cost rate (employer KiwiSaver, ACC levies, leave cover, equipment) and T is tools, training and model spend. Add a one-off recruitment cost R in year one.
• Agency, per year: A = W × d, where W is the days of AI work you actually commission and d is the agency’s day rate. Fixed-price projects work the same way once you divide the price by the effort.

Break-even rule: the hire is cheaper when W > I ÷ d. Worked example with ratios rather than invented rates: if an agency day costs one-hundredth of the hire’s fully loaded yearly cost, the hire breaks even at 100 days of real AI work a year. If it costs one-hundred-and-fiftieth, break-even is 150 days. Below that line, the agency is cheaper even at a higher day rate.

Three adjustments usually move the answer:

• Delay. If recruitment takes L months and a working system is worth V a month, waiting costs L × V. Three months of delay on a system worth V is 3V.
• Ramp-up. A new hire needs time to learn your systems and data before they are productive. Count those months at full cost.
• Coverage. One person rarely covers data engineering, evaluation, security and front-end work equally well. Budget agency days for the gaps even if you hire.

Ongoing costs exist either way. Models change: Anthropic retires Claude Sonnet 4.5 from its API on 30 November 2026, and any system built on it needs migrating. Someone has to do that work, inside or out. Our lesson on creating an AI budget lays out the running costs to plan for.

How do you test both on the same real task?

Give an agency and your shortlisted candidates the same small, paid piece of work from your own business, with personal details removed. Paying candidates for a work sample is fairer and gives better evidence than an interview alone.

1. Design an evaluation. Provide 30 anonymised supplier invoices and your Xero chart of accounts. Ask how they would measure automated coding accuracy and what accuracy would be needed before going live.
2. Estimate running costs. Ask for a monthly cost in NZD to code 2,000 invoices, with every assumption stated: model, tokens per invoice, platform fees and review time.
3. Write a handover note. Ask for a one-page runbook for your office manager: what to check each morning, and what to do when it fails.

Look for people who measure before promising, raise privacy and offshore processing without being asked, recommend leaving some steps manual, and write plainly. Our AI vendor evaluation checklist and AI procurement questions add the contract checks.

Which should you choose?

Choose an agency if

• This is your first production AI system and you are not yet sure what works.
• The work comes in bursts: a build, then months of light support.
• You need several specialist skills for a short period.
• You need a working system within a quarter, faster than you could recruit.

Choose an internal hire if

• You have a pipeline of AI work that will last several years.
• AI is part of a product you sell, not just internal efficiency.
• A senior manager can direct the role and give it data access and authority.
• Operators need someone beside them every day to adjust the system.

The common hybrid: an agency builds the first system with your internal owner working alongside, hands it over properly, and stays on call. Several agencies work this way, including Levy Street’s forward deployed AI engineers, and it only succeeds if you appoint the owner; our lesson on training an internal AI champion covers that role.

Where do both go wrong?

• The agency that leaves nothing behind. No documentation, no evaluation set, systems in the agency’s accounts. Require the handover items in our FAQ, and see documenting an AI workflow for handover.
• The lone AI hire. One person with a big title, no data access and no decision rights produces demos, not systems.
• No owner, no number. Neither option works without a business owner and an operating measure, such as turnaround time or error rate.
• Privacy in the contract. Under section 11 of the Privacy Act 2020, an agency that handles personal information only on your behalf is treated as part of your organisation, and you stay responsible. If it uses the data for its own purposes, the Privacy Commissioner says sharing may become a disclosure. Write that limit into the contract.
• Restructuring by stealth. If the project may affect jobs, the Employment Relations Act 2000 requires information and a chance to comment before you decide. That applies whoever builds it.

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