In-house AI team vs an agency.

Why this question is harder than it looks

The instinct, especially for operators who've built engineering teams before, is "we should hire AI people". It feels like building capability. It feels like the responsible long-term move. And in some cases it is. But for most mid-market businesses, the decision to hire an in-house AI team is being made before anyone has answered the harder question: what are we actually trying to build?

Hiring is the wrong starting move for the same reason hiring before you have product-market fit is the wrong starting move. You're committing to a fixed cost structure before you know what work needs doing. The mismatch shows up six to twelve months later when you have a team and a roadmap that don't fit each other.

The real cost of an in-house AI team in Australia

Let's price it properly. These are 2026 Australian market rates for full-time AI/ML engineers, observed in the Sydney, Melbourne and Brisbane markets across product companies, financial services and consulting.

Salaries

• Senior AI/ML engineer: $180K to $260K base, plus 12% super. Total comp typically $210K to $300K including super and equity-equivalent. Top of band for someone who has shipped agents into production at scale.
• Mid-level AI/ML engineer: $130K to $180K base, plus super. Total $150K to $210K.
• Junior: $90K to $130K. Useful but unable to lead or solo a production system.
• AI/ML engineering lead or staff engineer: $260K to $360K total comp. Adds people leadership, architectural decisions, accountability.

A minimum viable team that can actually ship is one senior plus one mid-level, with a part-time staff engineer or fractional CTO providing direction. Annual fully-loaded cost: $400K to $550K, before any infrastructure or tooling.

Recruitment time and cost

Senior AI talent in Australia is in short supply, and the salary band keeps moving up. Realistic time-to-hire for one senior AI engineer in 2026 Australian market is three to six months from the day you start looking. Mid-level slightly faster. Recruiter fees if you use an agency: typically 18% to 22% of first-year base, so $35K to $50K per senior hire. Internal recruiting time is hidden but real: ten to fifteen hours per week of leadership attention for the duration of the search.

Ramp time before they ship

Even a senior hire costs you for two to three months before they're producing meaningful output. They need to learn your systems, your data, your business, your customers, your stakeholders. During this period you're paying full salary for partial output. Cost to budget: roughly $50K to $80K per senior hire in non-productive ramp.

Infrastructure and tooling

For a small team, expect:

• Cloud compute and inference: $5K to $15K per month at the low end, $20K to $50K per month for production workloads with real traffic.
• Model API spend: variable, but a small production agent system commonly runs $3K to $20K per month in 2026 prices.
• Observability, evaluation, deployment tooling: $2K to $5K per month for the small-team stack (LangSmith, Weights & Biases, plus standard DevOps tooling).
• Vector database, retrieval infrastructure: $1K to $5K per month at modest scale.

Annual run cost for a small in-house team's infrastructure: $130K to $400K depending on workload, before model spend.

The opportunity cost

The cost that doesn't show up on a budget but determines whether the team is worth it: what are they actually building, and is it the right thing? Six months of a $500K annual team building the wrong system is a $250K loss plus the cost of throwing the work away. This happens routinely, and it's almost always upstream of the team itself: leadership had a vision but not a workflow audit, the team built what was asked, the asked-for thing wasn't the highest-leverage problem.

The cost shape of an agency

An agency engagement costs more on a fully-loaded hourly basis but less on a calendar-year basis, and a lot less on a time-to-first-shipped-system basis.

Engagement structures

• Fixed-scope sprint: $30K to $150K depending on scope. Defined deliverable, defined timeline, defined acceptance criteria. Useful when you know exactly what you want built.
• Monthly retainer: typically $15K to $80K per month for capacity equivalent to one to three senior engineers, plus an embedded technical lead. Useful when you have an evolving roadmap and want continuity.
• Fractional engagement: $8K to $20K per month for senior advisory and oversight, with delivery handled by your team or a separate vendor. Useful when you have engineering capacity but no AI seniority.
• One-off audit or strategy work: $15K to $50K. Useful before you commit to a build path.

What you're actually buying

The thing an agency provides that an in-house team can't, in the early phase, is compressed time. A good agency starts production work in week one because they have the infrastructure templates, the evaluation patterns, and the deployment muscle already in place. The same work in-house, even with great hires, takes three to six months to spin up because the team is building the foundations from scratch.

Annual cost of a competent retainer that ships a comparable volume of production AI systems to a small in-house team: $400K to $800K. Higher per-hour than salaries, but no ramp time, no recruitment cost, no infrastructure capex, no exposure to a senior hire leaving.

The decision framework

Four questions, asked honestly, decide which path fits.

1. Do you know what to build?

If yes, with confidence, hiring is more attractive. You're paying for execution against a known target.

If no, an agency is dramatically better. The cost of building the wrong thing in-house is enormous; an agency that does discovery as part of every engagement absorbs that risk.

2. How much AI work is on your roadmap?

If your roadmap has continuous AI/ML work for the next three years (call it five plus production systems, ongoing), in-house starts to make economic sense once you're past the unknown phase. The fixed cost amortises across many shipped systems.

If your roadmap has two or three production systems plus ongoing iteration, an agency retainer is cheaper and easier to start and stop.

3. Is the data or domain extremely sensitive?

For some workloads, in-house is preferred regardless of TCO. Defence, healthcare with PII, financial systems with regulator scrutiny, IP-sensitive R&D. The right agency can work in these contexts under strong NDA and security controls, but if your security model genuinely requires "no external party touches this", in-house is the answer.

4. What's your time-to-value tolerance?

If you need to ship a working system in eight to twelve weeks, agency is the only option. The earliest a new senior in-house hire ships meaningful production output is month four to six.

The hybrid pattern that wins

For most mid-market operators, the right shape isn't pure in-house or pure agency. It's a senior in-house technical lead, brought on as the first hire, paired with an agency engagement that builds the first systems while the team scales. This pattern works because:

• The senior in-house hire owns the AI agenda, sets direction, and accumulates institutional knowledge that survives any agency change.
• The agency ships the first production systems on a known timeline, while the lead is still ramping and recruiting.
• By month nine to twelve, you have running systems, an in-house lead with full context, and the option to bring more execution in-house or keep the agency relationship for capacity.
• When the agency engagement ends, the systems and the knowledge stay, because the in-house lead has been embedded in the work the entire time.

This is structurally cheaper than building from zero in-house, and structurally safer than running pure agency forever (no captive expertise, full vendor lock-in).

What goes wrong with each path

In-house failure modes

• Hiring before knowing the work. Team built; roadmap unclear; six months wasted on infrastructure with no business outcome. The most common failure.
• Senior hire churns. One person leaves and the team's capability drops 60% overnight. Common at year one to two.
• Slow start. Six to nine months from kickoff to first production system, while the business waits.
• Building everything bespoke. Reinventing patterns the rest of the industry has solved. Common when the team is junior or isolated.

Agency failure modes

• Demo shop. Agency builds polished demos that don't survive contact with production. Avoidable by picking an agency with a track record of shipped systems and asking to talk to existing clients.
• Lock-in. Critical systems built without proper handover documentation, leaving you dependent. Avoidable by writing handover into the engagement scope.
• No internal champion. Engagement runs without an internal lead, so when the agency leaves, no one understands the system. The hybrid model fixes this.

What this means for the R&D Tax Incentive

One asymmetry worth noting: agency work is generally easier to claim under the Australian R&D Tax Incentive than in-house work, in our experience. Sprint structure, hypothesis logging, and outcome documentation are baked into a properly run engagement. In-house teams often skip the documentation and lose the offset. We've seen the 43.5% offset turn an agency engagement into a net cheaper option than the in-house equivalent on paper, once R&D-eligible work is properly captured. Worth modelling for any business committing to material AI spend.

The honest answer

If you have a clear, ongoing AI roadmap, you've already validated which workflows are worth building, your data is too sensitive to share, and you can wait six months for first output, build in-house. Otherwise, start with an agency engagement that ships your first one or two production systems, hire your senior lead in parallel, and keep the agency on retainer until your team is shipping at the pace the business needs. The fastest path to durable AI capability isn't the path that feels most strategically virtuous. It's the one that gets a working system in front of customers soonest, with the lowest risk of paying for the wrong build.