“How much does an AI agent cost?” is the wrong question. It’s like asking how much a car costs. A used Hyundai i30 and a Ferrari are both cars, both legally drivable, both priced in dollars. Quoting a single number is useless.
The right question is: what work do I want done, who’s accountable for it, and what would I pay a human to do the same job? Once you can answer that, the agent’s price tag stops being a mystery and starts being a procurement decision.
This post is the answer I give CFOs, CIOs, and founders when they sit down for the first scoping call. Five buying patterns, their honest 2026 cost ranges, the line items most vendors hide, and a three-question test to make sure you’re comparing apples with apples. Numbers come from running an AI Factory that shipped 55+ production agents, plus the public pricing of every major platform as at May 2026.
The five patterns, and what each one costs
There are essentially five ways to “hire” an AI agent in 2026. They differ by who builds it, who runs it, and who’s accountable when it gets the answer wrong. Pick the wrong pattern for your problem and you’ll either overpay by 10x or underpay by 10x and ship nothing.
Pattern 1: per-resolution SaaS agents (cheapest to start, most expensive at scale)
Examples: Intercom Fin, Salesforce Agentforce, Zendesk AI Agents, ServiceNow AI Agents, Decagon, Sierra.
You bolt a vendor’s pre-built agent onto your support, sales, or service workflow. They handle the model, the prompt, the tools, the eval. You pay per ticket resolved, per conversation, or per “agent action”.
Honest ranges (May 2026):
- Intercom Fin: USD 0.99 per resolution.
- Salesforce Agentforce: USD 2.00 per conversation, plus a Data Cloud licence floor.
- Zendesk AI Agents: USD 1.00–1.50 per automated resolution.
- Sierra, Decagon, Cresta: priced per conversation or per outcome, typically negotiated, often USD 1–4 per resolution at enterprise volume.
Where it makes sense. High-volume, low-variance work where a 60–80% deflection rate is the goal and a human safety net handles the rest. Customer support is the canonical fit.
Where the sticker price lies. At 100 resolutions a month it looks free. At 50,000 it can cost more than the support team it was meant to deflect. Always model the cost at projected volume, not pilot volume.
Total annual envelope: USD 5K–500K+ depending on volume. Build cost: zero. Run cost: 100% of the bill.
Pattern 2: seat-based copilots (cheap, broad, shallow)
Examples: Microsoft 365 Copilot, Google Gemini for Workspace, ChatGPT Business / Enterprise, Claude Teams / Enterprise, GitHub Copilot.
A per-user licence that puts a general-purpose AI inside the tools your team already uses. Not really an “agent” in the strict sense (most of these are still copilots, see the operator primer), but the line is blurring fast as agentic features ship.
Honest ranges:
- Microsoft 365 Copilot: USD 30 / user / month (annual).
- Google Gemini for Workspace: USD 20–30 / user / month.
- ChatGPT Business: USD 25 / user / month.
- Claude Teams: USD 30 / user / month; Enterprise on negotiation.
- GitHub Copilot Business / Enterprise: USD 19–39 / user / month.
Where it makes sense. Productivity uplift across knowledge work. Drafting, summarising, code completion, research. The fastest place to get any AI value into a business with low risk and zero integration cost.
Where the cost gets ugly. Buying seats for everyone “because we should” with no measurement plan. A 500-person org on M365 Copilot is USD 180K a year before you’ve shown a single hour of saved work. Pilot the seat licence on the function with the highest documented hours-saved opportunity, then scale.
Total annual envelope: USD 5K–500K+ scaling linearly with headcount.
Pattern 3: no-code / low-code agent builds
Examples: n8n, Zapier with AI actions, Make, Microsoft Power Automate with Copilot Studio, OpenClaw, Lindy, Relevance AI.
You (or an integrator) wire up an agent on a visual builder. Trigger fires, the agent calls a model, calls some tools, writes a result somewhere. This is the sweet spot for SMEs and for the first 5–10 internal agents in any larger business.
Honest ranges (build + run, first year):
- Platform licence: USD 0 (n8n self-hosted) to USD 5K / year (n8n Cloud Pro, Zapier teams plans, Make enterprise).
- Model API spend: USD 50 / month low, USD 1K / month for a busy production agent. Budget USD 600–12K / year per agent.
- Build labour: 2–10 days of an integrator at AUD 1,500–2,500 / day, so AUD 3K–25K per agent depending on tool depth.
- Hosting (if self-hosted): USD 20–200 / month for the orchestration layer.
Where it makes sense. Internal automation across known systems (Xero, HubSpot, Slack, Jira, Notion). Agents that handle defined work with reasonable volume but don’t need bespoke infrastructure. The vast majority of the agents I ship at SME and mid-market scale fit here.
Total annual envelope: AUD 10K–60K per agent in year one, dropping in year two as build cost retires.
If you want a worked example with a full architecture, the build playbook for AI voice agents is a representative case at this tier.
Pattern 4: custom-built single agent
Examples: a bespoke agent built directly on the Claude, OpenAI, or Gemini API, exposed through your own tooling. Often built around the Model Context Protocol so the same tools work across model providers.
You pay for engineering time to build the prompt, the tools, the evaluation harness, the monitoring, the human-in-the-loop UI. You then pay for the model usage and the infrastructure to run it.
Honest ranges (year one, single production agent):
- Engineering build: 4–12 engineering weeks at AUD 1,800–3,000 / day, so AUD 36K–180K. Higher if the integration surface is dirty.
- Eval and quality harness: 1–3 weeks of additional engineering, AUD 9K–45K. Skip this and you’ll pay it back as production incidents.
- Model API spend: USD 200–5,000 / month depending on volume and model choice. Budget USD 2.4K–60K / year.
- Infrastructure and observability: USD 100–1,000 / month.
- Governance and security work: 2–4 weeks of legal/risk/compliance time. Often hidden in fully-loaded labour, but real.
- Ongoing run team: a fractional engineer at 0.1–0.3 FTE for monitoring, model upgrades, prompt iteration, error triage. AUD 25K–75K / year.
Where it makes sense. Work that touches sensitive data, requires bespoke logic, or is core to your competitive position. Anything you wouldn’t trust to a vendor’s black box.
Total annual envelope: AUD 80K–300K in year one, AUD 40K–120K / year ongoing.
Pattern 5: multi-agent factory
A standing capability that ships agents continuously across the business, with shared platform components (orchestration, RAG, MCP layer, eval, observability, governance) and a backlog rather than a project plan. This is what we ran at the Oman conglomerate, and what I now build for clients through the AI Factory service.
Honest ranges (year one, building the factory plus the first 5–10 agents):
- Platform stand-up: AUD 150K–400K. Orchestration, RAG, MCP, eval, observability, the governance scaffold.
- First agents: AUD 30K–80K each, marginal cost dropping fast as platform reuse kicks in.
- Standing team: 2–6 FTE blend (platform engineering, AI engineers, governance lead, change lead). AUD 400K–1.2M / year fully loaded.
- Model API spend: USD 5K–50K / month at portfolio scale. USD 60K–600K / year.
- Tooling and licences: USD 30K–100K / year (vector DB, observability, security tooling, model providers).
Where it makes sense. Mid-market and enterprise organisations with 50+ FTE in repeatable work, structured data, and an executive sponsor willing to fund this as an operating line, not a project. Read why AI factories beat AI projects before you sign off on this tier.
Total annual envelope: AUD 800K–3M+ year one, AUD 600K–2M+ ongoing. Compounds to 5–10x ROI when run with discipline. Stalls and burns when run as a project.
The line items vendors quietly leave out
Every cost range above gets a 20–40% uplift the first time you run it for real. The hidden line items:
-
Data preparation. Your CRM has 14 ways to spell the same customer. Your invoice store has PDFs, scans, and email attachments. Cleaning the data the agent will read costs more than building the agent. Budget 30–50% of build cost for data work on any agent that touches operational data.
-
Integration debt. “Connect to our ERP” is a sentence. The actual work is figuring out which of the four sandboxes is current, which fields the previous integrator deprecated, and who has the SAP licence to grant API access. Budget 1–3 extra weeks of engineering per integration that touches a legacy system.
-
Evaluation infrastructure. Golden sets, regression tests, drift monitoring, faithfulness metrics for any RAG pipeline. The single biggest predictor of whether an agent survives its first year. The work to build this is invisible until you’re trying to swap a model and have no way to know if quality dropped. Budget 15–25% of build labour.
-
Human-in-the-loop UI. Most agents need a screen where a human reviews, approves, or overrides decisions. That UI is real software with real cost. AUD 5K–30K depending on complexity.
-
Change management and adoption. McKinsey’s long-standing finding is that organisations investing in cultural change see 5.3x higher transformation success rates. The agent that nobody uses is the most expensive thing you can buy. Budget 10–20% of total programme cost for adoption work, run in parallel with the build.
-
Model and platform drift. The model layer changes every six weeks. Your prompt that scored 92% on the eval set last quarter scores 84% next quarter because the provider quietly deprecated a feature. You will pay for re-tuning. Budget 0.1–0.3 FTE of ongoing engineering per agent for the first 18 months.
Skip any of those line items and the headline price looks great. Then you find them in the budget, in the wrong quarter, with the wrong owner, and the programme stalls. See why 95% of AI pilots fail for the longer version of this argument.
How an agent actually compares to a human FTE
The fair comparison: would I rather pay an agent or a junior FTE to do this work? Be honest about both sides.
A junior FTE in Australia, fully loaded (salary, super, payroll tax, equipment, software, manager time, recruitment cost, training, leave loading): AUD 95K–150K / year for an entry-level role; AUD 150K–250K / year for a mid-level role.
A reasonably busy custom-built agent (Pattern 4): AUD 80K–150K total cost of ownership in year one, AUD 40K–80K ongoing.
So an agent that genuinely replaces 1.0 FTE of repeatable work pays back inside 12–18 months. An agent that replaces 0.3 FTE pays back in 3–4 years. An agent that doesn’t actually replace anyone (just makes the existing humans 10% faster) is a productivity tool, not a labour substitution. That’s a different ROI calc, and a much harder one to defend in front of a CFO.
For the full version of this calculation, including how I built the 300% ROI number and what actually pays back versus what doesn’t, read agentic AI ROI: the honest numbers.
The three-question test before you commit a budget
Before you sign anything, answer these three out loud. If you can’t, you’re not ready to buy.
1. What specific process am I automating, and what does it cost me today? A dollar number. Hours per week × fully-loaded hourly rate × 52. If you can’t say it, you don’t have a business case, you have a vibe. Map the process first (process inventory is the moat nobody maps).
2. Who owns this agent in production, and what’s their pager number? Not the consultant. Not the vendor. An internal name, an internal job, an internal accountability. If nobody can answer this, the agent will quietly stop running 90 days after the launch party.
3. What’s the human escalation path when the agent gets it wrong? And it will get it wrong. The cost of the agent is partially the cost of the humans who handle the 5–15% of cases the agent escalates. Bake that into the model, not the post-launch retro.
The one-paragraph answer
Pre-built SaaS agents start at a few thousand dollars a year and scale per usage. Seat-based copilots cost USD 20–40 per user per month. Custom-built single agents land around AUD 80K–300K total cost in year one. Multi-agent factories cost AUD 800K–3M+ in year one and pay back at 3–10x when run as standing capability rather than a project. Add 20–40% to whatever the headline number is for data prep, integration debt, eval, change management, and model drift. Compare every option to the fully-loaded cost of the human FTE doing the same work today. Then pick the smallest pattern that solves the actual problem.
That’s it. The question stops being “how much does an AI agent cost” and becomes “what’s my smallest credible step toward a measurable hours-saved number”. Which is exactly the right question.
If you want a one-hour scoping call to size your specific agent against these patterns, book a 30-minute intro and we’ll work the numbers on your actual processes. Or if you want the deeper read on what gets built versus bought, see the AI Factory case study.